reagent.training package
Subpackages
Submodules
reagent.training.c51_trainer module
- class reagent.training.c51_trainer.C51Trainer(q_network, q_network_target, actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), rl: reagent.core.parameters.RLParameters = Field(name='rl',type=<class 'reagent.core.parameters.RLParameters'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.RLParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), double_q_learning: bool = True, minibatch_size: int = 1024, minibatches_per_step: int = 1, num_atoms: int = 51, qmin: float = -100, qmax: float = 200, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD))
Bases:
reagent.training.rl_trainer_pytorch.RLTrainerMixin
,reagent.training.reagent_lightning_module.ReAgentLightningModule
Implementation of 51 Categorical DQN (C51)
See https://arxiv.org/abs/1707.06887 for details
- argmax_with_mask(q_values, possible_actions_mask)
- boost_rewards(rewards: torch.Tensor, actions: torch.Tensor) torch.Tensor
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.DiscreteDqnInput, batch_idx: int)
Implement training step as generator here
reagent.training.cem_trainer module
- The Trainer for Cross-Entropy Method. The idea is that an ensemble of
world models are fitted to predict transitions and reward functions.
A cross entropy method-based planner will then plan the best next action based on simulation data generated by the fitted world models.
The idea is inspired by: https://arxiv.org/abs/1805.12114
- class reagent.training.cem_trainer.CEMTrainer(cem_planner_network: reagent.models.cem_planner.CEMPlannerNetwork, world_model_trainers: List[reagent.training.world_model.mdnrnn_trainer.MDNRNNTrainer], parameters: reagent.core.parameters.CEMTrainerParameters)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- train_step_gen(training_batch: reagent.core.types.MemoryNetworkInput, batch_idx: int)
Implement training step as generator here
- training: bool
- reagent.training.cem_trainer.print_mdnrnn_losses(minibatch, model_index, losses) None
reagent.training.discrete_crr_trainer module
- class reagent.training.discrete_crr_trainer.DiscreteCRRTrainer(actor_network, actor_network_target, q1_network, q1_network_target, reward_network, q2_network=None, q2_network_target=None, q_network_cpe=None, q_network_cpe_target=None, metrics_to_score=None, evaluation: reagent.core.parameters.EvaluationParameters = Field(name=None,type=None,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.EvaluationParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=None), rl: reagent.core.parameters.RLParameters = Field(name='rl',type=<class 'reagent.core.parameters.RLParameters'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.RLParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), double_q_learning: bool = True, q_network_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='q_network_optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), actor_network_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='actor_network_optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), use_target_actor: bool = False, actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), delayed_policy_update: int = 1, beta: float = 1.0, entropy_coeff: float = 0.0, clip_limit: float = 10.0, max_weight: float = 20.0)
Bases:
reagent.training.dqn_trainer_base.DQNTrainerBaseLightning
Critic Regularized Regression (CRR) algorithm trainer as described in https://arxiv.org/abs/2006.15134
- compute_actor_loss(batch_idx, action, logged_action_probs, all_q_values, all_action_scores)
- compute_target_q_values(next_state, rewards, not_terminal, next_q_values)
- compute_td_loss(q_network, state, action, target_q_values)
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_detached_model_outputs(state) Tuple[torch.Tensor, None]
- property q_network
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.DiscreteDqnInput, batch_idx: int)
IMPORTANT: the input action here is preprocessed according to the training_batch type, which in this case is DiscreteDqnInput. Hence, the preprocessor in the DiscreteDqnInputMaker class in the trainer_preprocessor.py is used, which converts acion taken to a one-hot representation.
- validation_step(batch, batch_idx)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
reagent.training.dqn_trainer module
- class reagent.training.dqn_trainer.BCQConfig(drop_threshold: float = 0.1)
Bases:
object
- drop_threshold: float = 0.1
- class reagent.training.dqn_trainer.DQNTrainer(q_network, q_network_target, reward_network, q_network_cpe=None, q_network_cpe_target=None, metrics_to_score=None, evaluation: reagent.core.parameters.EvaluationParameters = Field(name=None,type=None,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.EvaluationParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=None), imitator=None, actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), rl: reagent.core.parameters.RLParameters = Field(name='rl',type=<class 'reagent.core.parameters.RLParameters'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.RLParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), double_q_learning: bool = True, bcq: Optional[reagent.training.dqn_trainer.BCQConfig] = None, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD))
Bases:
reagent.training.dqn_trainer_base.DQNTrainerBaseLightning
- compute_discount_tensor(batch: reagent.core.types.DiscreteDqnInput, boosted_rewards: torch.Tensor)
- compute_td_loss(batch: reagent.core.types.DiscreteDqnInput, boosted_rewards: torch.Tensor, discount_tensor: torch.Tensor)
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_detached_model_outputs(state) Tuple[torch.Tensor, Optional[torch.Tensor]]
Gets the q values from the model and target networks
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.DiscreteDqnInput, batch_idx: int)
Implement training step as generator here
- validation_step(batch, batch_idx)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
reagent.training.dqn_trainer_base module
- class reagent.training.dqn_trainer_base.DQNTrainerBaseLightning(rl_parameters: reagent.core.parameters.RLParameters, metrics_to_score=None, actions: Optional[List[str]] = None, evaluation_parameters: Optional[reagent.core.parameters.EvaluationParameters] = None)
Bases:
reagent.training.dqn_trainer_base.DQNTrainerMixin
,reagent.training.rl_trainer_pytorch.RLTrainerMixin
,reagent.training.reagent_lightning_module.ReAgentLightningModule
- boost_rewards(rewards: torch.Tensor, actions: torch.Tensor) torch.Tensor
- gather_eval_data(validation_step_outputs)
- property num_actions: int
- rl_parameters: reagent.core.parameters.RLParameters
- validation_epoch_end(valid_step_outputs)
Called at the end of the validation epoch with the outputs of all validation steps.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
outputs – List of outputs you defined in
validation_step()
, or if there are multiple dataloaders, a list containing a list of outputs for each dataloader.- Returns
None
Note
If you didn’t define a
validation_step()
, this won’t be called.Examples
With a single dataloader:
def validation_epoch_end(self, val_step_outputs): for out in val_step_outputs: ...
With multiple dataloaders, outputs will be a list of lists. The outer list contains one entry per dataloader, while the inner list contains the individual outputs of each validation step for that dataloader.
def validation_epoch_end(self, outputs): for dataloader_output_result in outputs: dataloader_outs = dataloader_output_result.dataloader_i_outputs self.log("final_metric", final_value)
- validation_step(batch, batch_idx)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
- class reagent.training.dqn_trainer_base.DQNTrainerMixin
Bases:
object
- ACTION_NOT_POSSIBLE_VAL = -1000000000.0
- get_max_q_values(q_values, possible_actions_mask)
- get_max_q_values_with_target(q_values, q_values_target, possible_actions_mask)
Used in Q-learning update.
- Parameters
q_values – PyTorch tensor with shape (batch_size, action_dim). Each row contains the list of Q-values for each possible action in this state.
q_values_target – PyTorch tensor with shape (batch_size, action_dim). Each row contains the list of Q-values from the target network for each possible action in this state.
possible_actions_mask – PyTorch tensor with shape (batch_size, action_dim). possible_actions[i][j] = 1 iff the agent can take action j from state i.
- Returns a tensor of maximum Q-values for every state in the batch
and also the index of the corresponding action (which is used in evaluation_data_page.py, in create_from_tensors_dqn()).
reagent.training.imitator_training module
- reagent.training.imitator_training.get_valid_actions_from_imitator(imitator, input, drop_threshold)
Create mask for non-viable actions under the imitator.
reagent.training.multi_stage_trainer module
- class reagent.training.multi_stage_trainer.MultiStageTrainer(trainers: List[reagent.training.reagent_lightning_module.ReAgentLightningModule], epochs: List[int], assign_reporter_function=None, flush_reporter_function=None, automatic_optimization=True)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- property multi_stage_total_epochs
- on_fit_end()
Called at the very end of fit.
If on DDP it is called on every process
- on_fit_start()
Called at the very beginning of fit.
If on DDP it is called on every process
- on_test_end()
Called at the end of testing.
- on_test_start()
Called at the beginning of testing.
- optimizer_step(epoch: int, batch_idx: int, optimizer, optimizer_idx: int, optimizer_closure, on_tpu: int = False, using_native_amp: int = False, using_lbfgs: int = False)
Override this method to adjust the default way the
Trainer
calls each optimizer. By default, Lightning callsstep()
andzero_grad()
as shown in the example once per optimizer. This method (andzero_grad()
) won’t be called during the accumulation phase whenTrainer(accumulate_grad_batches != 1)
.- Parameters
epoch – Current epoch
batch_idx – Index of current batch
optimizer – A PyTorch optimizer
optimizer_idx – If you used multiple optimizers, this indexes into that list.
optimizer_closure – Closure for all optimizers. This closure must be executed as it includes the calls to
training_step()
,optimizer.zero_grad()
, andbackward()
.on_tpu –
True
if TPU backward is requiredusing_native_amp –
True
if using native ampusing_lbfgs – True if the matching optimizer is
torch.optim.LBFGS
Examples:
# DEFAULT def optimizer_step(self, epoch, batch_idx, optimizer, optimizer_idx, optimizer_closure, on_tpu, using_native_amp, using_lbfgs): optimizer.step(closure=optimizer_closure) # Alternating schedule for optimizer steps (i.e.: GANs) def optimizer_step(self, epoch, batch_idx, optimizer, optimizer_idx, optimizer_closure, on_tpu, using_native_amp, using_lbfgs): # update generator opt every step if optimizer_idx == 0: optimizer.step(closure=optimizer_closure) # update discriminator opt every 2 steps if optimizer_idx == 1: if (batch_idx + 1) % 2 == 0 : optimizer.step(closure=optimizer_closure) else: # call the closure by itself to run `training_step` + `backward` without an optimizer step optimizer_closure() # ... # add as many optimizers as you want
Here’s another example showing how to use this for more advanced things such as learning rate warm-up:
# learning rate warm-up def optimizer_step( self, epoch, batch_idx, optimizer, optimizer_idx, optimizer_closure, on_tpu, using_native_amp, using_lbfgs, ): # warm up lr if self.trainer.global_step < 500: lr_scale = min(1.0, float(self.trainer.global_step + 1) / 500.0) for pg in optimizer.param_groups: pg["lr"] = lr_scale * self.learning_rate # update params optimizer.step(closure=optimizer_closure)
- set_reporter(reporter)
- test_epoch_end(outputs)
Called at the end of a test epoch with the output of all test steps.
# the pseudocode for these calls test_outs = [] for test_batch in test_data: out = test_step(test_batch) test_outs.append(out) test_epoch_end(test_outs)
- Parameters
outputs – List of outputs you defined in
test_step_end()
, or if there are multiple dataloaders, a list containing a list of outputs for each dataloader- Returns
None
Note
If you didn’t define a
test_step()
, this won’t be called.Examples
With a single dataloader:
def test_epoch_end(self, outputs): # do something with the outputs of all test batches all_test_preds = test_step_outputs.predictions some_result = calc_all_results(all_test_preds) self.log(some_result)
With multiple dataloaders, outputs will be a list of lists. The outer list contains one entry per dataloader, while the inner list contains the individual outputs of each test step for that dataloader.
def test_epoch_end(self, outputs): final_value = 0 for dataloader_outputs in outputs: for test_step_out in dataloader_outputs: # do something final_value += test_step_out self.log("final_metric", final_value)
- test_step(*args, **kwargs)
Operates on a single batch of data from the test set. In this step you’d normally generate examples or calculate anything of interest such as accuracy.
# the pseudocode for these calls test_outs = [] for test_batch in test_data: out = test_step(test_batch) test_outs.append(out) test_epoch_end(test_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch.
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple test dataloaders used).
- Returns
Any of.
Any object or value
None
- Testing will skip to the next batch
# if you have one test dataloader: def test_step(self, batch, batch_idx): ... # if you have multiple test dataloaders: def test_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single test dataset def test_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) test_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'test_loss': loss, 'test_acc': test_acc})
If you pass in multiple test dataloaders,
test_step()
will have an additional argument.# CASE 2: multiple test dataloaders def test_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to test you don’t need to implement this method.
Note
When the
test_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of the test epoch, the model goes back to training mode and gradients are enabled.
- training: bool
- training_epoch_end(outputs)
Called at the end of the training epoch with the outputs of all training steps. Use this in case you need to do something with all the outputs returned by
training_step()
.# the pseudocode for these calls train_outs = [] for train_batch in train_data: out = training_step(train_batch) train_outs.append(out) training_epoch_end(train_outs)
- Parameters
outputs – List of outputs you defined in
training_step()
. If there are multiple optimizers, it is a list containing a list of outputs for each optimizer. If usingtruncated_bptt_steps > 1
, each element is a list of outputs corresponding to the outputs of each processed split batch.- Returns
None
Note
If this method is not overridden, this won’t be called.
def training_epoch_end(self, training_step_outputs): # do something with all training_step outputs for out in training_step_outputs: ...
- training_step(batch, batch_idx: int, optimizer_idx: int = 0)
Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger.
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (
int
) – Integer displaying index of this batchoptimizer_idx (
int
) – When using multiple optimizers, this argument will also be present.hiddens (
Any
) – Passed in if :paramref:`~pytorch_lightning.core.lightning.LightningModule.truncated_bptt_steps` > 0.
- Returns
Any of.
Tensor
- The loss tensordict
- A dictionary. Can include any keys, but must include the key'loss'
None
- Training will skip to the next batch. This is only for automatic optimization.This is not supported for multi-GPU, TPU, IPU, or DeepSpeed.
In this step you’d normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something model specific.
Example:
def training_step(self, batch, batch_idx): x, y, z = batch out = self.encoder(x) loss = self.loss(out, x) return loss
If you define multiple optimizers, this step will be called with an additional
optimizer_idx
parameter.# Multiple optimizers (e.g.: GANs) def training_step(self, batch, batch_idx, optimizer_idx): if optimizer_idx == 0: # do training_step with encoder ... if optimizer_idx == 1: # do training_step with decoder ...
If you add truncated back propagation through time you will also get an additional argument with the hidden states of the previous step.
# Truncated back-propagation through time def training_step(self, batch, batch_idx, hiddens): # hiddens are the hidden states from the previous truncated backprop step out, hiddens = self.lstm(data, hiddens) loss = ... return {"loss": loss, "hiddens": hiddens}
Note
The loss value shown in the progress bar is smoothed (averaged) over the last values, so it differs from the actual loss returned in train/validation step.
- validation_epoch_end(outputs)
Called at the end of the validation epoch with the outputs of all validation steps.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
outputs – List of outputs you defined in
validation_step()
, or if there are multiple dataloaders, a list containing a list of outputs for each dataloader.- Returns
None
Note
If you didn’t define a
validation_step()
, this won’t be called.Examples
With a single dataloader:
def validation_epoch_end(self, val_step_outputs): for out in val_step_outputs: ...
With multiple dataloaders, outputs will be a list of lists. The outer list contains one entry per dataloader, while the inner list contains the individual outputs of each validation step for that dataloader.
def validation_epoch_end(self, outputs): for dataloader_output_result in outputs: dataloader_outs = dataloader_output_result.dataloader_i_outputs self.log("final_metric", final_value)
- validation_step(*args, **kwargs)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
reagent.training.parameters module
- class reagent.training.parameters.C51TrainerParameters(actions: List[str] = <factory>, rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, minibatch_size: int = 1024, minibatches_per_step: int = 1, num_atoms: int = 51, qmin: float = -100, qmax: float = 200, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>)
Bases:
object
- actions: List[str]
- asdict()
- double_q_learning: bool = True
- minibatch_size: int = 1024
- minibatches_per_step: int = 1
- num_atoms: int = 51
- optimizer: reagent.optimizer.union.Optimizer__Union
- qmax: float = 200
- qmin: float = -100
- class reagent.training.parameters.CRRTrainerParameters(rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, q_network_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, actor_network_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, use_target_actor: bool = False, actions: List[str] = <factory>, delayed_policy_update: int = 1, beta: float = 1.0, entropy_coeff: float = 0.0, clip_limit: float = 10.0, max_weight: float = 20.0)
Bases:
object
- actions: List[str]
- actor_network_optimizer: reagent.optimizer.union.Optimizer__Union
- asdict()
- beta: float = 1.0
- clip_limit: float = 10.0
- delayed_policy_update: int = 1
- double_q_learning: bool = True
- entropy_coeff: float = 0.0
- max_weight: float = 20.0
- q_network_optimizer: reagent.optimizer.union.Optimizer__Union
- use_target_actor: bool = False
- class reagent.training.parameters.DQNTrainerParameters(actions: List[str] = <factory>, rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, bcq: Optional[reagent.training.dqn_trainer.BCQConfig] = None, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>)
Bases:
object
- actions: List[str]
- asdict()
- bcq: Optional[reagent.training.dqn_trainer.BCQConfig] = None
- double_q_learning: bool = True
- minibatch_size: int = 1024
- minibatches_per_step: int = 1
- optimizer: reagent.optimizer.union.Optimizer__Union
- class reagent.training.parameters.LinUCBTrainerParameters(num_actions: int = - 1, use_interaction_features: bool = True)
Bases:
object
- asdict()
- num_actions: int = -1
- use_interaction_features: bool = True
- class reagent.training.parameters.PPOTrainerParameters(gamma: float = 0.9, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, optimizer_value_net: reagent.optimizer.union.Optimizer__Union = <factory>, actions: List[str] = <factory>, reward_clip: float = 1000000.0, normalize: bool = True, subtract_mean: bool = True, offset_clamp_min: bool = False, update_freq: int = 1, update_epochs: int = 1, ppo_batch_size: int = 1, ppo_epsilon: float = 0.2, entropy_weight: float = 0.0)
Bases:
object
- actions: List[str]
- asdict()
- entropy_weight: float = 0.0
- gamma: float = 0.9
- normalize: bool = True
- offset_clamp_min: bool = False
- optimizer: reagent.optimizer.union.Optimizer__Union
- optimizer_value_net: reagent.optimizer.union.Optimizer__Union
- ppo_batch_size: int = 1
- ppo_epsilon: float = 0.2
- reward_clip: float = 1000000.0
- subtract_mean: bool = True
- update_epochs: int = 1
- update_freq: int = 1
- class reagent.training.parameters.ParametricDQNTrainerParameters(rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>)
Bases:
object
- asdict()
- double_q_learning: bool = True
- minibatches_per_step: int = 1
- optimizer: reagent.optimizer.union.Optimizer__Union
- class reagent.training.parameters.QRDQNTrainerParameters(actions: List[str] = <factory>, rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, num_atoms: int = 51, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, cpe_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>)
Bases:
object
- actions: List[str]
- asdict()
- cpe_optimizer: reagent.optimizer.union.Optimizer__Union
- double_q_learning: bool = True
- minibatch_size: int = 1024
- minibatches_per_step: int = 1
- num_atoms: int = 51
- optimizer: reagent.optimizer.union.Optimizer__Union
- class reagent.training.parameters.ReinforceTrainerParameters(gamma: float = 0.0, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, optimizer_value_net: reagent.optimizer.union.Optimizer__Union = <factory>, actions: List[str] = <factory>, off_policy: bool = False, reward_clip: float = 1000000.0, clip_param: float = 1000000.0, normalize: bool = True, subtract_mean: bool = True, offset_clamp_min: bool = False)
Bases:
object
- actions: List[str]
- asdict()
- clip_param: float = 1000000.0
- gamma: float = 0.0
- normalize: bool = True
- off_policy: bool = False
- offset_clamp_min: bool = False
- optimizer: reagent.optimizer.union.Optimizer__Union
- optimizer_value_net: reagent.optimizer.union.Optimizer__Union
- reward_clip: float = 1000000.0
- subtract_mean: bool = True
- class reagent.training.parameters.RewardNetworkTrainerParameters(optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, loss_type: reagent.training.reward_network_trainer.LossFunction = <LossFunction.MSE: 'MSE_Loss'>, reward_ignore_threshold: Optional[float] = None, weighted_by_inverse_propensity: bool = False)
Bases:
object
- asdict()
- loss_type: reagent.training.reward_network_trainer.LossFunction = 'MSE_Loss'
- optimizer: reagent.optimizer.union.Optimizer__Union
- reward_ignore_threshold: Optional[float] = None
- weighted_by_inverse_propensity: bool = False
- class reagent.training.parameters.SACTrainerParameters(rl: reagent.core.parameters.RLParameters = <factory>, q_network_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, value_network_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, actor_network_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, alpha_optimizer: Optional[reagent.optimizer.union.Optimizer__Union] = <factory>, minibatch_size: int = 1024, entropy_temperature: float = 0.01, logged_action_uniform_prior: bool = True, target_entropy: float = -1.0, action_embedding_kld_weight: Optional[float] = None, apply_kld_on_mean: bool = False, action_embedding_mean: Optional[List[float]] = None, action_embedding_variance: Optional[List[float]] = None, crr_config: Optional[reagent.training.sac_trainer.CRRWeightFn] = None, backprop_through_log_prob: bool = True)
Bases:
object
- action_embedding_kld_weight: Optional[float] = None
- action_embedding_mean: Optional[List[float]] = None
- action_embedding_variance: Optional[List[float]] = None
- actor_network_optimizer: reagent.optimizer.union.Optimizer__Union
- alpha_optimizer: Optional[reagent.optimizer.union.Optimizer__Union]
- apply_kld_on_mean: bool = False
- asdict()
- backprop_through_log_prob: bool = True
- crr_config: Optional[reagent.training.sac_trainer.CRRWeightFn] = None
- entropy_temperature: float = 0.01
- logged_action_uniform_prior: bool = True
- minibatch_size: int = 1024
- q_network_optimizer: reagent.optimizer.union.Optimizer__Union
- target_entropy: float = -1.0
- value_network_optimizer: reagent.optimizer.union.Optimizer__Union
- class reagent.training.parameters.Seq2SlateTrainerParameters(params: reagent.core.parameters.Seq2SlateParameters = <factory>, policy_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, baseline_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, policy_gradient_interval: int = 1, print_interval: int = 100, calc_cpe: bool = False, reward_network: Optional[torch.nn.modules.module.Module] = None)
Bases:
reagent.core.base_dataclass.BaseDataClass
- asdict()
- baseline_optimizer: reagent.optimizer.union.Optimizer__Union
- calc_cpe: bool = False
- policy_gradient_interval: int = 1
- policy_optimizer: reagent.optimizer.union.Optimizer__Union
- print_interval: int = 100
- reward_network: Optional[torch.nn.modules.module.Module] = None
- class reagent.training.parameters.SlateQTrainerParameters(rl: reagent.core.parameters.RLParameters = <factory>, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, slate_opt_parameters: Optional[reagent.core.parameters.SlateOptParameters] = None, discount_time_scale: Optional[float] = None, single_selection: bool = True, next_slate_value_norm_method: reagent.training.slate_q_trainer.NextSlateValueNormMethod = <NextSlateValueNormMethod.NORM_BY_CURRENT_SLATE_SIZE: 'norm_by_current_slate_size'>, minibatch_size: int = 1024, evaluation: reagent.core.parameters.EvaluationParameters = <factory>)
Bases:
object
- asdict()
- discount_time_scale: Optional[float] = None
- evaluation: reagent.core.parameters.EvaluationParameters
- minibatch_size: int = 1024
- next_slate_value_norm_method: reagent.training.slate_q_trainer.NextSlateValueNormMethod = 'norm_by_current_slate_size'
- optimizer: reagent.optimizer.union.Optimizer__Union
- single_selection: bool = True
- slate_opt_parameters: Optional[reagent.core.parameters.SlateOptParameters] = None
- class reagent.training.parameters.TD3TrainerParameters(rl: reagent.core.parameters.RLParameters = <factory>, q_network_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, actor_network_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, minibatch_size: int = 64, noise_variance: float = 0.2, noise_clip: float = 0.5, delayed_policy_update: int = 2, minibatches_per_step: int = 1)
Bases:
object
- actor_network_optimizer: reagent.optimizer.union.Optimizer__Union
- asdict()
- delayed_policy_update: int = 2
- minibatch_size: int = 64
- minibatches_per_step: int = 1
- noise_clip: float = 0.5
- noise_variance: float = 0.2
- q_network_optimizer: reagent.optimizer.union.Optimizer__Union
reagent.training.parametric_dqn_trainer module
- class reagent.training.parametric_dqn_trainer.ParametricDQNTrainer(q_network, q_network_target, reward_network, rl: reagent.core.parameters.RLParameters = Field(name='rl', type=<class 'reagent.core.parameters.RLParameters'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<class 'reagent.core.parameters.RLParameters'>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD), double_q_learning: bool = True, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer', type=<class 'reagent.optimizer.union.Optimizer__Union'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD))
Bases:
reagent.training.dqn_trainer_base.DQNTrainerMixin
,reagent.training.rl_trainer_pytorch.RLTrainerMixin
,reagent.training.reagent_lightning_module.ReAgentLightningModule
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_detached_model_outputs(state, action) Tuple[torch.Tensor, torch.Tensor]
Gets the q values from the model and target networks
- train_step_gen(training_batch: reagent.core.types.ParametricDqnInput, batch_idx: int)
Implement training step as generator here
reagent.training.ppo_trainer module
- class reagent.training.ppo_trainer.PPOTrainer(policy: reagent.gym.policies.policy.Policy, gamma: float = 0.9, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), optimizer_value_net: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer_value_net',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), reward_clip: float = 1000000.0, normalize: bool = True, subtract_mean: bool = True, offset_clamp_min: bool = False, update_freq: int = 1, update_epochs: int = 1, ppo_batch_size: int = 1, ppo_epsilon: float = 0.2, entropy_weight: float = 0.0, value_net: Optional[reagent.models.base.ModelBase] = None)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
Proximal Policy Optimization (PPO). See https://arxiv.org/pdf/1707.06347.pdf This is the “clip” version of PPO. It does not include: - KL divergence - Bootstrapping with a critic model (our approach only works if full trajectories up to terminal state are fed in) Optionally, a value network can be trained and used as a baseline for rewards.
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_optimizers()
- training: bool
- training_step(training_batch: Union[reagent.core.types.PolicyGradientInput, Dict[str, torch.Tensor]], batch_idx: int)
Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger.
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (
int
) – Integer displaying index of this batchoptimizer_idx (
int
) – When using multiple optimizers, this argument will also be present.hiddens (
Any
) – Passed in if :paramref:`~pytorch_lightning.core.lightning.LightningModule.truncated_bptt_steps` > 0.
- Returns
Any of.
Tensor
- The loss tensordict
- A dictionary. Can include any keys, but must include the key'loss'
None
- Training will skip to the next batch. This is only for automatic optimization.This is not supported for multi-GPU, TPU, IPU, or DeepSpeed.
In this step you’d normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something model specific.
Example:
def training_step(self, batch, batch_idx): x, y, z = batch out = self.encoder(x) loss = self.loss(out, x) return loss
If you define multiple optimizers, this step will be called with an additional
optimizer_idx
parameter.# Multiple optimizers (e.g.: GANs) def training_step(self, batch, batch_idx, optimizer_idx): if optimizer_idx == 0: # do training_step with encoder ... if optimizer_idx == 1: # do training_step with decoder ...
If you add truncated back propagation through time you will also get an additional argument with the hidden states of the previous step.
# Truncated back-propagation through time def training_step(self, batch, batch_idx, hiddens): # hiddens are the hidden states from the previous truncated backprop step out, hiddens = self.lstm(data, hiddens) loss = ... return {"loss": loss, "hiddens": hiddens}
Note
The loss value shown in the progress bar is smoothed (averaged) over the last values, so it differs from the actual loss returned in train/validation step.
- update_model()
reagent.training.qrdqn_trainer module
- class reagent.training.qrdqn_trainer.QRDQNTrainer(q_network, q_network_target, metrics_to_score=None, reward_network=None, q_network_cpe=None, q_network_cpe_target=None, actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), rl: reagent.core.parameters.RLParameters = Field(name='rl',type=<class 'reagent.core.parameters.RLParameters'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.RLParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), double_q_learning: bool = True, num_atoms: int = 51, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), cpe_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='cpe_optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), evaluation: reagent.core.parameters.EvaluationParameters = Field(name=None,type=None,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.EvaluationParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=None))
Bases:
reagent.training.dqn_trainer_base.DQNTrainerBaseLightning
Implementation of QR-DQN (Quantile Regression Deep Q-Network)
See https://arxiv.org/abs/1710.10044 for details
- allow_zero_length_dataloader_with_multiple_devices: bool
- argmax_with_mask(q_values, possible_actions_mask)
- boost_rewards(rewards: torch.Tensor, actions: torch.Tensor) torch.Tensor
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_detached_model_outputs(state: reagent.core.types.FeatureData) Tuple[torch.Tensor, torch.Tensor]
Gets the q values from the model and target networks
- huber(x)
- precision: int
- prepare_data_per_node: bool
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.DiscreteDqnInput, batch_idx: int)
Implement training step as generator here
- training: bool
- use_amp: bool
reagent.training.reagent_lightning_module module
- class reagent.training.reagent_lightning_module.ReAgentLightningModule(automatic_optimization=True)
Bases:
pytorch_lightning.core.lightning.LightningModule
- increase_next_stopping_epochs(num_epochs: int)
- on_epoch_end()
Called when either of train/val/test epoch ends.
- on_test_batch_end(*args, **kwargs)
Called in the test loop after the batch.
- Parameters
outputs – The outputs of test_step_end(test_step(x))
batch – The batched data as it is returned by the test DataLoader.
batch_idx – the index of the batch
dataloader_idx – the index of the dataloader
- on_train_batch_end(*args, **kwargs)
Called in the training loop after the batch.
- Parameters
outputs – The outputs of training_step_end(training_step(x))
batch – The batched data as it is returned by the training DataLoader.
batch_idx – the index of the batch
unused – Deprecated argument. Will be removed in v1.7.
- on_validation_batch_end(*args, **kwargs)
Called in the validation loop after the batch.
- Parameters
outputs – The outputs of validation_step_end(validation_step(x))
batch – The batched data as it is returned by the validation DataLoader.
batch_idx – the index of the batch
dataloader_idx – the index of the dataloader
- optimizers(use_pl_optimizer: bool = True)
Returns the optimizer(s) that are being used during training. Useful for manual optimization.
- Parameters
use_pl_optimizer – If
True
, will wrap the optimizer(s) in aLightningOptimizer
for automatic handling of precision and profiling.- Returns
A single optimizer, or a list of optimizers in case multiple ones are present.
- property reporter
- set_clean_stop(clean_stop: bool)
- set_reporter(reporter)
- soft_update_result() torch.Tensor
A dummy loss to trigger soft-update
- property summary_writer
Accessor to TensorBoard’s SummaryWriter
- train(*args)
Sets the module in training mode.
This has any effect only on certain modules. See documentations of particular modules for details of their behaviors in training/evaluation mode, if they are affected, e.g.
Dropout
,BatchNorm
, etc.- Parameters
mode (bool) – whether to set training mode (
True
) or evaluation mode (False
). Default:True
.- Returns
self
- Return type
Module
- train_step_gen(training_batch, batch_idx: int)
Implement training step as generator here
- training: bool
- training_step(batch, batch_idx: int, optimizer_idx: int = 0)
Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger.
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (
int
) – Integer displaying index of this batchoptimizer_idx (
int
) – When using multiple optimizers, this argument will also be present.hiddens (
Any
) – Passed in if :paramref:`~pytorch_lightning.core.lightning.LightningModule.truncated_bptt_steps` > 0.
- Returns
Any of.
Tensor
- The loss tensordict
- A dictionary. Can include any keys, but must include the key'loss'
None
- Training will skip to the next batch. This is only for automatic optimization.This is not supported for multi-GPU, TPU, IPU, or DeepSpeed.
In this step you’d normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something model specific.
Example:
def training_step(self, batch, batch_idx): x, y, z = batch out = self.encoder(x) loss = self.loss(out, x) return loss
If you define multiple optimizers, this step will be called with an additional
optimizer_idx
parameter.# Multiple optimizers (e.g.: GANs) def training_step(self, batch, batch_idx, optimizer_idx): if optimizer_idx == 0: # do training_step with encoder ... if optimizer_idx == 1: # do training_step with decoder ...
If you add truncated back propagation through time you will also get an additional argument with the hidden states of the previous step.
# Truncated back-propagation through time def training_step(self, batch, batch_idx, hiddens): # hiddens are the hidden states from the previous truncated backprop step out, hiddens = self.lstm(data, hiddens) loss = ... return {"loss": loss, "hiddens": hiddens}
Note
The loss value shown in the progress bar is smoothed (averaged) over the last values, so it differs from the actual loss returned in train/validation step.
- class reagent.training.reagent_lightning_module.StoppingEpochCallback(num_epochs)
Bases:
pytorch_lightning.callbacks.base.Callback
We use this callback to control the number of training epochs in incremental training. Epoch & step counts are not reset in the checkpoint. If we were to set max_epochs on the trainer, we would have to keep track of the previous max_epochs and add to it manually. This keeps the infomation in one place.
Note that we need to set _cleanly_stopped back to True before saving the checkpoint. This is done in ModelManager.save_trainer().
- on_pretrain_routine_end(trainer, pl_module)
Called when the pretrain routine ends.
- reagent.training.reagent_lightning_module.has_test_step_override(trainer_module: reagent.training.reagent_lightning_module.ReAgentLightningModule)
Detect if a subclass of LightningModule has test_step overridden
reagent.training.reinforce_trainer module
- class reagent.training.reinforce_trainer.ReinforceTrainer(policy: reagent.gym.policies.policy.Policy, gamma: float = 0.0, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), optimizer_value_net: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer_value_net',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), off_policy: bool = False, reward_clip: float = 1000000.0, clip_param: float = 1000000.0, normalize: bool = True, subtract_mean: bool = True, offset_clamp_min: bool = False, value_net: Optional[reagent.models.base.ModelBase] = None)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
- allow_zero_length_dataloader_with_multiple_devices: bool
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- precision: int
- prepare_data_per_node: bool
- train_step_gen(training_batch: reagent.core.types.PolicyGradientInput, batch_idx: int)
Implement training step as generator here
- training: bool
- use_amp: bool
reagent.training.reward_network_trainer module
- class reagent.training.reward_network_trainer.LossFunction(value)
Bases:
enum.Enum
An enumeration.
- BCELoss = 'BCE_Loss'
- L1Loss = 'L1_Loss'
- MSE = 'MSE_Loss'
- SmoothL1Loss = 'SmoothL1_Loss'
- class reagent.training.reward_network_trainer.RewardNetTrainer(reward_net: reagent.models.base.ModelBase, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer', type=<class 'reagent.optimizer.union.Optimizer__Union'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD), loss_type: reagent.training.reward_network_trainer.LossFunction = LossFunction.MSE, reward_ignore_threshold: Optional[float] = None, weighted_by_inverse_propensity: bool = False)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
- allow_zero_length_dataloader_with_multiple_devices: bool
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- precision: int
- prepare_data_per_node: bool
- train_step_gen(training_batch: reagent.core.types.PreprocessedRankingInput, batch_idx: int)
Implement training step as generator here
- training: bool
- use_amp: bool
- validation_epoch_end(outputs)
Called at the end of the validation epoch with the outputs of all validation steps.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
outputs – List of outputs you defined in
validation_step()
, or if there are multiple dataloaders, a list containing a list of outputs for each dataloader.- Returns
None
Note
If you didn’t define a
validation_step()
, this won’t be called.Examples
With a single dataloader:
def validation_epoch_end(self, val_step_outputs): for out in val_step_outputs: ...
With multiple dataloaders, outputs will be a list of lists. The outer list contains one entry per dataloader, while the inner list contains the individual outputs of each validation step for that dataloader.
def validation_epoch_end(self, outputs): for dataloader_output_result in outputs: dataloader_outs = dataloader_output_result.dataloader_i_outputs self.log("final_metric", final_value)
- validation_step(batch: reagent.core.types.PreprocessedRankingInput, batch_idx: int)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
- warm_start_components()
reagent.training.rl_trainer_pytorch module
- class reagent.training.rl_trainer_pytorch.RLTrainerMixin
Bases:
object
- ACTION_NOT_POSSIBLE_VAL = -1000000000.0
- property gamma: float
- property maxq_learning: bool
- property multi_steps: Optional[int]
- rl_parameters: reagent.core.parameters.RLParameters
- property rl_temperature: float
- property tau: float
- property use_seq_num_diff_as_time_diff: bool
reagent.training.sac_trainer module
- class reagent.training.sac_trainer.CRRWeightFn(indicator_fn_threshold: Optional[float] = None, exponent_beta: Optional[float] = None, exponent_clamp: Optional[float] = None)
Bases:
object
- exponent_beta: Optional[float] = None
- exponent_clamp: Optional[float] = None
- get_weight_from_advantage(advantage)
- indicator_fn_threshold: Optional[float] = None
- class reagent.training.sac_trainer.SACTrainer(actor_network, q1_network, q2_network=None, value_network=None, rl: reagent.core.parameters.RLParameters = Field(name='rl',type=<class 'reagent.core.parameters.RLParameters'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.RLParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), q_network_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='q_network_optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), value_network_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='value_network_optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), actor_network_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='actor_network_optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), alpha_optimizer: Optional[reagent.optimizer.union.Optimizer__Union] = Field(name='alpha_optimizer',type=typing.Optional[reagent.optimizer.union.Optimizer__Union],default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), minibatch_size: int = 1024, entropy_temperature: float = 0.01, logged_action_uniform_prior: bool = True, target_entropy: float = -1.0, action_embedding_kld_weight: Optional[float] = None, apply_kld_on_mean: bool = False, action_embedding_mean: Optional[List[float]] = None, action_embedding_variance: Optional[List[float]] = None, crr_config: Optional[reagent.training.sac_trainer.CRRWeightFn] = None, backprop_through_log_prob: bool = True)
Bases:
reagent.training.rl_trainer_pytorch.RLTrainerMixin
,reagent.training.reagent_lightning_module.ReAgentLightningModule
Soft Actor-Critic trainer as described in https://arxiv.org/pdf/1801.01290
The actor is assumed to implement reparameterization trick.
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.PolicyNetworkInput, batch_idx: int)
IMPORTANT: the input action here is assumed to match the range of the output of the actor.
reagent.training.slate_q_trainer module
- class reagent.training.slate_q_trainer.NextSlateValueNormMethod(value)
Bases:
enum.Enum
The Q value of the current slate item is the sum of the item’s short-term reward and the normalized sum of all item Q-values on the next slate. We can normalize the sum by either the current slate size (NORM_BY_CURRENT_SLATE_SIZE) or the next slate size (NORM_BY_NEXT_SLATE_SIZE). This enum distinguishes between these two different ways of normalizing the next slate value.
- NORM_BY_CURRENT_SLATE_SIZE = 'norm_by_current_slate_size'
- NORM_BY_NEXT_SLATE_SIZE = 'norm_by_next_slate_size'
- class reagent.training.slate_q_trainer.SlateQTrainer(q_network, q_network_target, slate_size, rl: reagent.core.parameters.RLParameters = Field(name='rl', type=<class 'reagent.core.parameters.RLParameters'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<function SlateQTrainer.<lambda>>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD), optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer', type=<class 'reagent.optimizer.union.Optimizer__Union'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD), slate_opt_parameters: Optional[reagent.core.parameters.SlateOptParameters] = None, discount_time_scale: Optional[float] = None, single_selection: bool = True, next_slate_value_norm_method: reagent.training.slate_q_trainer.NextSlateValueNormMethod = NextSlateValueNormMethod.NORM_BY_CURRENT_SLATE_SIZE, minibatch_size: int = 1024, evaluation: reagent.core.parameters.EvaluationParameters = Field(name='evaluation', type=<class 'reagent.core.parameters.EvaluationParameters'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<function SlateQTrainer.<lambda>>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD))
Bases:
reagent.training.rl_trainer_pytorch.RLTrainerMixin
,reagent.training.reagent_lightning_module.ReAgentLightningModule
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.SlateQInput, batch_idx: int)
Implement training step as generator here
reagent.training.td3_trainer module
- class reagent.training.td3_trainer.TD3Trainer(actor_network, q1_network, q2_network=None, rl: reagent.core.parameters.RLParameters = Field(name='rl', type=<class 'reagent.core.parameters.RLParameters'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<class 'reagent.core.parameters.RLParameters'>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD), q_network_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='q_network_optimizer', type=<class 'reagent.optimizer.union.Optimizer__Union'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD), actor_network_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='actor_network_optimizer', type=<class 'reagent.optimizer.union.Optimizer__Union'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD), minibatch_size: int = 64, noise_variance: float = 0.2, noise_clip: float = 0.5, delayed_policy_update: int = 2, minibatches_per_step: int = 1)
Bases:
reagent.training.rl_trainer_pytorch.RLTrainerMixin
,reagent.training.reagent_lightning_module.ReAgentLightningModule
Twin Delayed Deep Deterministic Policy Gradient algorithm trainer as described in https://arxiv.org/pdf/1802.09477
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.PolicyNetworkInput, batch_idx: int)
IMPORTANT: the input action here is assumed to be preprocessed to match the range of the output of the actor.
reagent.training.utils module
- reagent.training.utils.discounted_returns(rewards: torch.Tensor, gamma: float = 0) torch.Tensor
Perform rollout to compute reward to go and do a baseline subtraction.
- reagent.training.utils.gen_permutations(seq_len: int, num_action: int) torch.Tensor
generate all seq_len permutations for a given action set the return shape is (SEQ_LEN, PERM_NUM, ACTION_DIM)
- reagent.training.utils.rescale_actions(actions: torch.Tensor, new_min: torch.Tensor, new_max: torch.Tensor, prev_min: torch.Tensor, prev_max: torch.Tensor) torch.Tensor
Scale from [prev_min, prev_max] to [new_min, new_max]
- reagent.training.utils.whiten(x: torch.Tensor, subtract_mean: bool) torch.Tensor
Module contents
- class reagent.training.BanditRewardNetTrainer(reward_net: reagent.models.base.ModelBase, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name=None, type=None, default=<dataclasses._MISSING_TYPE object>, default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=None), loss_type: reagent.training.reward_network_trainer.LossFunction = LossFunction.MSE, reward_ignore_threshold: Optional[float] = None, weighted_by_inverse_propensity: bool = False)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
- allow_zero_length_dataloader_with_multiple_devices: bool
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- precision: int
- prepare_data_per_node: bool
- train_step_gen(training_batch: reagent.core.types.BanditRewardModelInput, batch_idx: int)
Implement training step as generator here
- training: bool
- use_amp: bool
- validation_epoch_end(outputs)
Called at the end of the validation epoch with the outputs of all validation steps.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
outputs – List of outputs you defined in
validation_step()
, or if there are multiple dataloaders, a list containing a list of outputs for each dataloader.- Returns
None
Note
If you didn’t define a
validation_step()
, this won’t be called.Examples
With a single dataloader:
def validation_epoch_end(self, val_step_outputs): for out in val_step_outputs: ...
With multiple dataloaders, outputs will be a list of lists. The outer list contains one entry per dataloader, while the inner list contains the individual outputs of each validation step for that dataloader.
def validation_epoch_end(self, outputs): for dataloader_output_result in outputs: dataloader_outs = dataloader_output_result.dataloader_i_outputs self.log("final_metric", final_value)
- validation_step(batch: reagent.core.types.BanditRewardModelInput, batch_idx: int)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
- class reagent.training.C51Trainer(q_network, q_network_target, actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), rl: reagent.core.parameters.RLParameters = Field(name='rl',type=<class 'reagent.core.parameters.RLParameters'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.RLParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), double_q_learning: bool = True, minibatch_size: int = 1024, minibatches_per_step: int = 1, num_atoms: int = 51, qmin: float = -100, qmax: float = 200, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD))
Bases:
reagent.training.rl_trainer_pytorch.RLTrainerMixin
,reagent.training.reagent_lightning_module.ReAgentLightningModule
Implementation of 51 Categorical DQN (C51)
See https://arxiv.org/abs/1707.06887 for details
- allow_zero_length_dataloader_with_multiple_devices: bool
- argmax_with_mask(q_values, possible_actions_mask)
- boost_rewards(rewards: torch.Tensor, actions: torch.Tensor) torch.Tensor
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- precision: int
- prepare_data_per_node: bool
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.DiscreteDqnInput, batch_idx: int)
Implement training step as generator here
- training: bool
- use_amp: bool
- class reagent.training.C51TrainerParameters(actions: List[str] = <factory>, rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, minibatch_size: int = 1024, minibatches_per_step: int = 1, num_atoms: int = 51, qmin: float = -100, qmax: float = 200, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>)
Bases:
object
- actions: List[str]
- asdict()
- double_q_learning: bool = True
- minibatch_size: int = 1024
- minibatches_per_step: int = 1
- num_atoms: int = 51
- optimizer: reagent.optimizer.union.Optimizer__Union
- qmax: float = 200
- qmin: float = -100
- class reagent.training.CEMTrainer(cem_planner_network: reagent.models.cem_planner.CEMPlannerNetwork, world_model_trainers: List[reagent.training.world_model.mdnrnn_trainer.MDNRNNTrainer], parameters: reagent.core.parameters.CEMTrainerParameters)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
- allow_zero_length_dataloader_with_multiple_devices: bool
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- precision: int
- prepare_data_per_node: bool
- train_step_gen(training_batch: reagent.core.types.MemoryNetworkInput, batch_idx: int)
Implement training step as generator here
- training: bool
- use_amp: bool
- class reagent.training.CRRTrainerParameters(rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, q_network_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, actor_network_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, use_target_actor: bool = False, actions: List[str] = <factory>, delayed_policy_update: int = 1, beta: float = 1.0, entropy_coeff: float = 0.0, clip_limit: float = 10.0, max_weight: float = 20.0)
Bases:
object
- actions: List[str]
- actor_network_optimizer: reagent.optimizer.union.Optimizer__Union
- asdict()
- beta: float = 1.0
- clip_limit: float = 10.0
- delayed_policy_update: int = 1
- double_q_learning: bool = True
- entropy_coeff: float = 0.0
- max_weight: float = 20.0
- q_network_optimizer: reagent.optimizer.union.Optimizer__Union
- use_target_actor: bool = False
- class reagent.training.DQNTrainer(q_network, q_network_target, reward_network, q_network_cpe=None, q_network_cpe_target=None, metrics_to_score=None, evaluation: reagent.core.parameters.EvaluationParameters = Field(name=None,type=None,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.EvaluationParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=None), imitator=None, actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), rl: reagent.core.parameters.RLParameters = Field(name='rl',type=<class 'reagent.core.parameters.RLParameters'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.RLParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), double_q_learning: bool = True, bcq: Optional[reagent.training.dqn_trainer.BCQConfig] = None, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD))
Bases:
reagent.training.dqn_trainer_base.DQNTrainerBaseLightning
- allow_zero_length_dataloader_with_multiple_devices: bool
- compute_discount_tensor(batch: reagent.core.types.DiscreteDqnInput, boosted_rewards: torch.Tensor)
- compute_td_loss(batch: reagent.core.types.DiscreteDqnInput, boosted_rewards: torch.Tensor, discount_tensor: torch.Tensor)
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_detached_model_outputs(state) Tuple[torch.Tensor, Optional[torch.Tensor]]
Gets the q values from the model and target networks
- precision: int
- prepare_data_per_node: bool
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.DiscreteDqnInput, batch_idx: int)
Implement training step as generator here
- training: bool
- use_amp: bool
- validation_step(batch, batch_idx)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
- class reagent.training.DQNTrainerParameters(actions: List[str] = <factory>, rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, bcq: Optional[reagent.training.dqn_trainer.BCQConfig] = None, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>)
Bases:
object
- actions: List[str]
- asdict()
- bcq: Optional[reagent.training.dqn_trainer.BCQConfig] = None
- double_q_learning: bool = True
- minibatch_size: int = 1024
- minibatches_per_step: int = 1
- optimizer: reagent.optimizer.union.Optimizer__Union
- class reagent.training.DiscreteCRRTrainer(actor_network, actor_network_target, q1_network, q1_network_target, reward_network, q2_network=None, q2_network_target=None, q_network_cpe=None, q_network_cpe_target=None, metrics_to_score=None, evaluation: reagent.core.parameters.EvaluationParameters = Field(name=None,type=None,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.EvaluationParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=None), rl: reagent.core.parameters.RLParameters = Field(name='rl',type=<class 'reagent.core.parameters.RLParameters'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.RLParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), double_q_learning: bool = True, q_network_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='q_network_optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), actor_network_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='actor_network_optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), use_target_actor: bool = False, actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), delayed_policy_update: int = 1, beta: float = 1.0, entropy_coeff: float = 0.0, clip_limit: float = 10.0, max_weight: float = 20.0)
Bases:
reagent.training.dqn_trainer_base.DQNTrainerBaseLightning
Critic Regularized Regression (CRR) algorithm trainer as described in https://arxiv.org/abs/2006.15134
- allow_zero_length_dataloader_with_multiple_devices: bool
- compute_actor_loss(batch_idx, action, logged_action_probs, all_q_values, all_action_scores)
- compute_target_q_values(next_state, rewards, not_terminal, next_q_values)
- compute_td_loss(q_network, state, action, target_q_values)
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_detached_model_outputs(state) Tuple[torch.Tensor, None]
- precision: int
- prepare_data_per_node: bool
- property q_network
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.DiscreteDqnInput, batch_idx: int)
IMPORTANT: the input action here is preprocessed according to the training_batch type, which in this case is DiscreteDqnInput. Hence, the preprocessor in the DiscreteDqnInputMaker class in the trainer_preprocessor.py is used, which converts acion taken to a one-hot representation.
- training: bool
- use_amp: bool
- validation_step(batch, batch_idx)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
- class reagent.training.MDNRNNTrainer(memory_network: reagent.models.world_model.MemoryNetwork, params: reagent.core.parameters.MDNRNNTrainerParameters, cum_loss_hist: int = 100)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
Trainer for MDN-RNN
- allow_zero_length_dataloader_with_multiple_devices: bool
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_loss(training_batch: reagent.core.types.MemoryNetworkInput, state_dim: Optional[int] = None)
- Compute losses:
GMMLoss(next_state, GMMPredicted) / (STATE_DIM + 2) + MSE(reward, predicted_reward) + BCE(not_terminal, logit_not_terminal)
- The STATE_DIM + 2 factor is here to counteract the fact that the GMMLoss scales
approximately linearly with STATE_DIM, dim of states. All losses are averaged both on the batch and the sequence dimensions (the two first dimensions).
- Parameters
training_batch – training_batch has these fields: - state: (SEQ_LEN, BATCH_SIZE, STATE_DIM) torch tensor - action: (SEQ_LEN, BATCH_SIZE, ACTION_DIM) torch tensor - reward: (SEQ_LEN, BATCH_SIZE) torch tensor - not-terminal: (SEQ_LEN, BATCH_SIZE) torch tensor - next_state: (SEQ_LEN, BATCH_SIZE, STATE_DIM) torch tensor
state_dim – the dimension of states. If provided, use it to normalize gmm loss
- Returns
dictionary of losses, containing the gmm, the mse, the bce and the averaged loss.
- precision: int
- prepare_data_per_node: bool
- test_step(training_batch: reagent.core.types.MemoryNetworkInput, batch_idx: int)
Operates on a single batch of data from the test set. In this step you’d normally generate examples or calculate anything of interest such as accuracy.
# the pseudocode for these calls test_outs = [] for test_batch in test_data: out = test_step(test_batch) test_outs.append(out) test_epoch_end(test_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch.
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple test dataloaders used).
- Returns
Any of.
Any object or value
None
- Testing will skip to the next batch
# if you have one test dataloader: def test_step(self, batch, batch_idx): ... # if you have multiple test dataloaders: def test_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single test dataset def test_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) test_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'test_loss': loss, 'test_acc': test_acc})
If you pass in multiple test dataloaders,
test_step()
will have an additional argument.# CASE 2: multiple test dataloaders def test_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to test you don’t need to implement this method.
Note
When the
test_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of the test epoch, the model goes back to training mode and gradients are enabled.
- train_step_gen(training_batch: reagent.core.types.MemoryNetworkInput, batch_idx: int)
Implement training step as generator here
- training: bool
- use_amp: bool
- validation_step(training_batch: reagent.core.types.MemoryNetworkInput, batch_idx: int)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
- class reagent.training.MultiStageTrainer(trainers: List[reagent.training.reagent_lightning_module.ReAgentLightningModule], epochs: List[int], assign_reporter_function=None, flush_reporter_function=None, automatic_optimization=True)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
- allow_zero_length_dataloader_with_multiple_devices: bool
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- property multi_stage_total_epochs
- on_fit_end()
Called at the very end of fit.
If on DDP it is called on every process
- on_fit_start()
Called at the very beginning of fit.
If on DDP it is called on every process
- on_test_end()
Called at the end of testing.
- on_test_start()
Called at the beginning of testing.
- optimizer_step(epoch: int, batch_idx: int, optimizer, optimizer_idx: int, optimizer_closure, on_tpu: int = False, using_native_amp: int = False, using_lbfgs: int = False)
Override this method to adjust the default way the
Trainer
calls each optimizer. By default, Lightning callsstep()
andzero_grad()
as shown in the example once per optimizer. This method (andzero_grad()
) won’t be called during the accumulation phase whenTrainer(accumulate_grad_batches != 1)
.- Parameters
epoch – Current epoch
batch_idx – Index of current batch
optimizer – A PyTorch optimizer
optimizer_idx – If you used multiple optimizers, this indexes into that list.
optimizer_closure – Closure for all optimizers. This closure must be executed as it includes the calls to
training_step()
,optimizer.zero_grad()
, andbackward()
.on_tpu –
True
if TPU backward is requiredusing_native_amp –
True
if using native ampusing_lbfgs – True if the matching optimizer is
torch.optim.LBFGS
Examples:
# DEFAULT def optimizer_step(self, epoch, batch_idx, optimizer, optimizer_idx, optimizer_closure, on_tpu, using_native_amp, using_lbfgs): optimizer.step(closure=optimizer_closure) # Alternating schedule for optimizer steps (i.e.: GANs) def optimizer_step(self, epoch, batch_idx, optimizer, optimizer_idx, optimizer_closure, on_tpu, using_native_amp, using_lbfgs): # update generator opt every step if optimizer_idx == 0: optimizer.step(closure=optimizer_closure) # update discriminator opt every 2 steps if optimizer_idx == 1: if (batch_idx + 1) % 2 == 0 : optimizer.step(closure=optimizer_closure) else: # call the closure by itself to run `training_step` + `backward` without an optimizer step optimizer_closure() # ... # add as many optimizers as you want
Here’s another example showing how to use this for more advanced things such as learning rate warm-up:
# learning rate warm-up def optimizer_step( self, epoch, batch_idx, optimizer, optimizer_idx, optimizer_closure, on_tpu, using_native_amp, using_lbfgs, ): # warm up lr if self.trainer.global_step < 500: lr_scale = min(1.0, float(self.trainer.global_step + 1) / 500.0) for pg in optimizer.param_groups: pg["lr"] = lr_scale * self.learning_rate # update params optimizer.step(closure=optimizer_closure)
- precision: int
- prepare_data_per_node: bool
- set_reporter(reporter)
- test_epoch_end(outputs)
Called at the end of a test epoch with the output of all test steps.
# the pseudocode for these calls test_outs = [] for test_batch in test_data: out = test_step(test_batch) test_outs.append(out) test_epoch_end(test_outs)
- Parameters
outputs – List of outputs you defined in
test_step_end()
, or if there are multiple dataloaders, a list containing a list of outputs for each dataloader- Returns
None
Note
If you didn’t define a
test_step()
, this won’t be called.Examples
With a single dataloader:
def test_epoch_end(self, outputs): # do something with the outputs of all test batches all_test_preds = test_step_outputs.predictions some_result = calc_all_results(all_test_preds) self.log(some_result)
With multiple dataloaders, outputs will be a list of lists. The outer list contains one entry per dataloader, while the inner list contains the individual outputs of each test step for that dataloader.
def test_epoch_end(self, outputs): final_value = 0 for dataloader_outputs in outputs: for test_step_out in dataloader_outputs: # do something final_value += test_step_out self.log("final_metric", final_value)
- test_step(*args, **kwargs)
Operates on a single batch of data from the test set. In this step you’d normally generate examples or calculate anything of interest such as accuracy.
# the pseudocode for these calls test_outs = [] for test_batch in test_data: out = test_step(test_batch) test_outs.append(out) test_epoch_end(test_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch.
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple test dataloaders used).
- Returns
Any of.
Any object or value
None
- Testing will skip to the next batch
# if you have one test dataloader: def test_step(self, batch, batch_idx): ... # if you have multiple test dataloaders: def test_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single test dataset def test_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) test_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'test_loss': loss, 'test_acc': test_acc})
If you pass in multiple test dataloaders,
test_step()
will have an additional argument.# CASE 2: multiple test dataloaders def test_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to test you don’t need to implement this method.
Note
When the
test_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of the test epoch, the model goes back to training mode and gradients are enabled.
- training: bool
- training_epoch_end(outputs)
Called at the end of the training epoch with the outputs of all training steps. Use this in case you need to do something with all the outputs returned by
training_step()
.# the pseudocode for these calls train_outs = [] for train_batch in train_data: out = training_step(train_batch) train_outs.append(out) training_epoch_end(train_outs)
- Parameters
outputs – List of outputs you defined in
training_step()
. If there are multiple optimizers, it is a list containing a list of outputs for each optimizer. If usingtruncated_bptt_steps > 1
, each element is a list of outputs corresponding to the outputs of each processed split batch.- Returns
None
Note
If this method is not overridden, this won’t be called.
def training_epoch_end(self, training_step_outputs): # do something with all training_step outputs for out in training_step_outputs: ...
- training_step(batch, batch_idx: int, optimizer_idx: int = 0)
Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger.
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (
int
) – Integer displaying index of this batchoptimizer_idx (
int
) – When using multiple optimizers, this argument will also be present.hiddens (
Any
) – Passed in if :paramref:`~pytorch_lightning.core.lightning.LightningModule.truncated_bptt_steps` > 0.
- Returns
Any of.
Tensor
- The loss tensordict
- A dictionary. Can include any keys, but must include the key'loss'
None
- Training will skip to the next batch. This is only for automatic optimization.This is not supported for multi-GPU, TPU, IPU, or DeepSpeed.
In this step you’d normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something model specific.
Example:
def training_step(self, batch, batch_idx): x, y, z = batch out = self.encoder(x) loss = self.loss(out, x) return loss
If you define multiple optimizers, this step will be called with an additional
optimizer_idx
parameter.# Multiple optimizers (e.g.: GANs) def training_step(self, batch, batch_idx, optimizer_idx): if optimizer_idx == 0: # do training_step with encoder ... if optimizer_idx == 1: # do training_step with decoder ...
If you add truncated back propagation through time you will also get an additional argument with the hidden states of the previous step.
# Truncated back-propagation through time def training_step(self, batch, batch_idx, hiddens): # hiddens are the hidden states from the previous truncated backprop step out, hiddens = self.lstm(data, hiddens) loss = ... return {"loss": loss, "hiddens": hiddens}
Note
The loss value shown in the progress bar is smoothed (averaged) over the last values, so it differs from the actual loss returned in train/validation step.
- use_amp: bool
- validation_epoch_end(outputs)
Called at the end of the validation epoch with the outputs of all validation steps.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
outputs – List of outputs you defined in
validation_step()
, or if there are multiple dataloaders, a list containing a list of outputs for each dataloader.- Returns
None
Note
If you didn’t define a
validation_step()
, this won’t be called.Examples
With a single dataloader:
def validation_epoch_end(self, val_step_outputs): for out in val_step_outputs: ...
With multiple dataloaders, outputs will be a list of lists. The outer list contains one entry per dataloader, while the inner list contains the individual outputs of each validation step for that dataloader.
def validation_epoch_end(self, outputs): for dataloader_output_result in outputs: dataloader_outs = dataloader_output_result.dataloader_i_outputs self.log("final_metric", final_value)
- validation_step(*args, **kwargs)
Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.
# the pseudocode for these calls val_outs = [] for val_batch in val_data: out = validation_step(val_batch) val_outs.append(out) validation_epoch_end(val_outs)
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (int) – The index of this batch
dataloader_idx (int) – The index of the dataloader that produced this batch (only if multiple val dataloaders used)
- Returns
Any object or value
None
- Validation will skip to the next batch
# pseudocode of order val_outs = [] for val_batch in val_data: out = validation_step(val_batch) if defined("validation_step_end"): out = validation_step_end(out) val_outs.append(out) val_outs = validation_epoch_end(val_outs)
# if you have one val dataloader: def validation_step(self, batch, batch_idx): ... # if you have multiple val dataloaders: def validation_step(self, batch, batch_idx, dataloader_idx): ...
Examples:
# CASE 1: A single validation dataset def validation_step(self, batch, batch_idx): x, y = batch # implement your own out = self(x) loss = self.loss(out, y) # log 6 example images # or generated text... or whatever sample_imgs = x[:6] grid = torchvision.utils.make_grid(sample_imgs) self.logger.experiment.add_image('example_images', grid, 0) # calculate acc labels_hat = torch.argmax(out, dim=1) val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0) # log the outputs! self.log_dict({'val_loss': loss, 'val_acc': val_acc})
If you pass in multiple val dataloaders,
validation_step()
will have an additional argument.# CASE 2: multiple validation dataloaders def validation_step(self, batch, batch_idx, dataloader_idx): # dataloader_idx tells you which dataset this is. ...
Note
If you don’t need to validate you don’t need to implement this method.
Note
When the
validation_step()
is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.
- class reagent.training.PPOTrainer(policy: reagent.gym.policies.policy.Policy, gamma: float = 0.9, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), optimizer_value_net: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer_value_net',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), reward_clip: float = 1000000.0, normalize: bool = True, subtract_mean: bool = True, offset_clamp_min: bool = False, update_freq: int = 1, update_epochs: int = 1, ppo_batch_size: int = 1, ppo_epsilon: float = 0.2, entropy_weight: float = 0.0, value_net: Optional[reagent.models.base.ModelBase] = None)
Bases:
reagent.training.reagent_lightning_module.ReAgentLightningModule
Proximal Policy Optimization (PPO). See https://arxiv.org/pdf/1707.06347.pdf This is the “clip” version of PPO. It does not include: - KL divergence - Bootstrapping with a critic model (our approach only works if full trajectories up to terminal state are fed in) Optionally, a value network can be trained and used as a baseline for rewards.
- allow_zero_length_dataloader_with_multiple_devices: bool
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_optimizers()
- precision: int
- prepare_data_per_node: bool
- training: bool
- training_step(training_batch: Union[reagent.core.types.PolicyGradientInput, Dict[str, torch.Tensor]], batch_idx: int)
Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger.
- Parameters
batch (
Tensor
| (Tensor
, …) | [Tensor
, …]) – The output of yourDataLoader
. A tensor, tuple or list.batch_idx (
int
) – Integer displaying index of this batchoptimizer_idx (
int
) – When using multiple optimizers, this argument will also be present.hiddens (
Any
) – Passed in if :paramref:`~pytorch_lightning.core.lightning.LightningModule.truncated_bptt_steps` > 0.
- Returns
Any of.
Tensor
- The loss tensordict
- A dictionary. Can include any keys, but must include the key'loss'
None
- Training will skip to the next batch. This is only for automatic optimization.This is not supported for multi-GPU, TPU, IPU, or DeepSpeed.
In this step you’d normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something model specific.
Example:
def training_step(self, batch, batch_idx): x, y, z = batch out = self.encoder(x) loss = self.loss(out, x) return loss
If you define multiple optimizers, this step will be called with an additional
optimizer_idx
parameter.# Multiple optimizers (e.g.: GANs) def training_step(self, batch, batch_idx, optimizer_idx): if optimizer_idx == 0: # do training_step with encoder ... if optimizer_idx == 1: # do training_step with decoder ...
If you add truncated back propagation through time you will also get an additional argument with the hidden states of the previous step.
# Truncated back-propagation through time def training_step(self, batch, batch_idx, hiddens): # hiddens are the hidden states from the previous truncated backprop step out, hiddens = self.lstm(data, hiddens) loss = ... return {"loss": loss, "hiddens": hiddens}
Note
The loss value shown in the progress bar is smoothed (averaged) over the last values, so it differs from the actual loss returned in train/validation step.
- update_model()
- use_amp: bool
- class reagent.training.PPOTrainerParameters(gamma: float = 0.9, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, optimizer_value_net: reagent.optimizer.union.Optimizer__Union = <factory>, actions: List[str] = <factory>, reward_clip: float = 1000000.0, normalize: bool = True, subtract_mean: bool = True, offset_clamp_min: bool = False, update_freq: int = 1, update_epochs: int = 1, ppo_batch_size: int = 1, ppo_epsilon: float = 0.2, entropy_weight: float = 0.0)
Bases:
object
- actions: List[str]
- asdict()
- entropy_weight: float = 0.0
- gamma: float = 0.9
- normalize: bool = True
- offset_clamp_min: bool = False
- optimizer: reagent.optimizer.union.Optimizer__Union
- optimizer_value_net: reagent.optimizer.union.Optimizer__Union
- ppo_batch_size: int = 1
- ppo_epsilon: float = 0.2
- reward_clip: float = 1000000.0
- subtract_mean: bool = True
- update_epochs: int = 1
- update_freq: int = 1
- class reagent.training.ParametricDQNTrainer(q_network, q_network_target, reward_network, rl: reagent.core.parameters.RLParameters = Field(name='rl', type=<class 'reagent.core.parameters.RLParameters'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<class 'reagent.core.parameters.RLParameters'>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD), double_q_learning: bool = True, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer', type=<class 'reagent.optimizer.union.Optimizer__Union'>, default=<dataclasses._MISSING_TYPE object>, default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>, init=True, repr=True, hash=None, compare=True, metadata=mappingproxy({}), _field_type=_FIELD))
Bases:
reagent.training.dqn_trainer_base.DQNTrainerMixin
,reagent.training.rl_trainer_pytorch.RLTrainerMixin
,reagent.training.reagent_lightning_module.ReAgentLightningModule
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_detached_model_outputs(state, action) Tuple[torch.Tensor, torch.Tensor]
Gets the q values from the model and target networks
- train_step_gen(training_batch: reagent.core.types.ParametricDqnInput, batch_idx: int)
Implement training step as generator here
- class reagent.training.ParametricDQNTrainerParameters(rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>)
Bases:
object
- asdict()
- double_q_learning: bool = True
- minibatches_per_step: int = 1
- optimizer: reagent.optimizer.union.Optimizer__Union
- class reagent.training.QRDQNTrainer(q_network, q_network_target, metrics_to_score=None, reward_network=None, q_network_cpe=None, q_network_cpe_target=None, actions: List[str] = Field(name='actions',type=typing.List[str],default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'list'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), rl: reagent.core.parameters.RLParameters = Field(name='rl',type=<class 'reagent.core.parameters.RLParameters'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.RLParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), double_q_learning: bool = True, num_atoms: int = 51, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), cpe_optimizer: reagent.optimizer.union.Optimizer__Union = Field(name='cpe_optimizer',type=<class 'reagent.optimizer.union.Optimizer__Union'>,default=<dataclasses._MISSING_TYPE object>,default_factory=<bound method Optimizer__Union.default of <class 'reagent.optimizer.union.Optimizer__Union'>>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), evaluation: reagent.core.parameters.EvaluationParameters = Field(name=None,type=None,default=<dataclasses._MISSING_TYPE object>,default_factory=<class 'reagent.core.parameters.EvaluationParameters'>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=None))
Bases:
reagent.training.dqn_trainer_base.DQNTrainerBaseLightning
Implementation of QR-DQN (Quantile Regression Deep Q-Network)
See https://arxiv.org/abs/1710.10044 for details
- allow_zero_length_dataloader_with_multiple_devices: bool
- argmax_with_mask(q_values, possible_actions_mask)
- boost_rewards(rewards: torch.Tensor, actions: torch.Tensor) torch.Tensor
- configure_optimizers()
Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple.
- Returns
Any of these 6 options.
Single optimizer.
List or Tuple of optimizers.
Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple
lr_scheduler_config
).Dictionary, with an
"optimizer"
key, and (optionally) a"lr_scheduler"
key whose value is a single LR scheduler orlr_scheduler_config
.Tuple of dictionaries as described above, with an optional
"frequency"
key.None - Fit will run without any optimizer.
The
lr_scheduler_config
is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.lr_scheduler_config = { # REQUIRED: The scheduler instance "scheduler": lr_scheduler, # The unit of the scheduler's step size, could also be 'step'. # 'epoch' updates the scheduler on epoch end whereas 'step' # updates it after a optimizer update. "interval": "epoch", # How many epochs/steps should pass between calls to # `scheduler.step()`. 1 corresponds to updating the learning # rate after every epoch/step. "frequency": 1, # Metric to to monitor for schedulers like `ReduceLROnPlateau` "monitor": "val_loss", # If set to `True`, will enforce that the value specified 'monitor' # is available when the scheduler is updated, thus stopping # training if not found. If set to `False`, it will only produce a warning "strict": True, # If using the `LearningRateMonitor` callback to monitor the # learning rate progress, this keyword can be used to specify # a custom logged name "name": None, }
When there are schedulers in which the
.step()
method is conditioned on a value, such as thetorch.optim.lr_scheduler.ReduceLROnPlateau
scheduler, Lightning requires that thelr_scheduler_config
contains the keyword"monitor"
set to the metric name that the scheduler should be conditioned on.Metrics can be made available to monitor by simply logging it using
self.log('metric_to_track', metric_val)
in yourLightningModule
.Note
The
frequency
value specified in a dict along with theoptimizer
key is an int corresponding to the number of sequential batches optimized with the specific optimizer. It should be given to none or to all of the optimizers. There is a difference between passing multiple optimizers in a list, and passing multiple optimizers in dictionaries with a frequency of 1:In the former case, all optimizers will operate on the given batch in each optimization step.
In the latter, only one optimizer will operate on the given batch at every step.
This is different from the
frequency
value specified in thelr_scheduler_config
mentioned above.def configure_optimizers(self): optimizer_one = torch.optim.SGD(self.model.parameters(), lr=0.01) optimizer_two = torch.optim.SGD(self.model.parameters(), lr=0.01) return [ {"optimizer": optimizer_one, "frequency": 5}, {"optimizer": optimizer_two, "frequency": 10}, ]
In this example, the first optimizer will be used for the first 5 steps, the second optimizer for the next 10 steps and that cycle will continue. If an LR scheduler is specified for an optimizer using the
lr_scheduler
key in the above dict, the scheduler will only be updated when its optimizer is being used.Examples:
# most cases. no learning rate scheduler def configure_optimizers(self): return Adam(self.parameters(), lr=1e-3) # multiple optimizer case (e.g.: GAN) def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) return gen_opt, dis_opt # example with learning rate schedulers def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) dis_sch = CosineAnnealing(dis_opt, T_max=10) return [gen_opt, dis_opt], [dis_sch] # example with step-based learning rate schedulers # each optimizer has its own scheduler def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) gen_sch = { 'scheduler': ExponentialLR(gen_opt, 0.99), 'interval': 'step' # called after each training step } dis_sch = CosineAnnealing(dis_opt, T_max=10) # called every epoch return [gen_opt, dis_opt], [gen_sch, dis_sch] # example with optimizer frequencies # see training procedure in `Improved Training of Wasserstein GANs`, Algorithm 1 # https://arxiv.org/abs/1704.00028 def configure_optimizers(self): gen_opt = Adam(self.model_gen.parameters(), lr=0.01) dis_opt = Adam(self.model_dis.parameters(), lr=0.02) n_critic = 5 return ( {'optimizer': dis_opt, 'frequency': n_critic}, {'optimizer': gen_opt, 'frequency': 1} )
Note
Some things to know:
Lightning calls
.backward()
and.step()
on each optimizer and learning rate scheduler as needed.If you use 16-bit precision (
precision=16
), Lightning will automatically handle the optimizers.If you use multiple optimizers,
training_step()
will have an additionaloptimizer_idx
parameter.If you use
torch.optim.LBFGS
, Lightning handles the closure function automatically for you.If you use multiple optimizers, gradients will be calculated only for the parameters of current optimizer at each training step.
If you need to control how often those optimizers step or override the default
.step()
schedule, override theoptimizer_step()
hook.
- get_detached_model_outputs(state: reagent.core.types.FeatureData) Tuple[torch.Tensor, torch.Tensor]
Gets the q values from the model and target networks
- huber(x)
- precision: int
- prepare_data_per_node: bool
- rl_parameters: reagent.core.parameters.RLParameters
- train_step_gen(training_batch: reagent.core.types.DiscreteDqnInput, batch_idx: int)
Implement training step as generator here
- training: bool
- use_amp: bool
- class reagent.training.QRDQNTrainerParameters(actions: List[str] = <factory>, rl: reagent.core.parameters.RLParameters = <factory>, double_q_learning: bool = True, num_atoms: int = 51, minibatch_size: int = 1024, minibatches_per_step: int = 1, optimizer: reagent.optimizer.union.Optimizer__Union = <factory>, cpe_optimizer: reagent.optimizer.union.Optimizer__Union = <factory>)
Bases:
object
- actions: List[str]
- asdict()
- cpe_optimizer: reagent.optimizer.union.Optimizer__Union
- double_q_learning: bool = True
- minibatch_size: int = 1024
- minibatches_per_step: int = 1
- num_atoms: int = 51
- optimizer: reagent.optimizer.union.Optimizer__Union
- class reagent.training.ReAgentLightningModule(automatic_optimization=True)
Bases:
pytorch_lightning.core.lightning.LightningModule
- allow_zero_length_dataloader_with_multiple_devices: bool
- increase_next_stopping_epochs(num_epochs: int)
- on_epoch_end()
Called when either of train/val/test epoch ends.
- on_test_batch_end(*args, **kwargs)
Called in the test loop after the batch.
- Parameters
outputs – The outputs of test_step_end(test_step(x))
batch – The batched data as it is returned by the test DataLoader.
batch_idx – the index of the batch
dataloader_idx – the index of the dataloader
- on_train_batch_end(*args, **kwargs)
Called in the training loop after the batch.
- Parameters
outputs – The outputs of training_step_end(training_step(x))
batch – The batched data as it is returned by the training DataLoader.
batch_idx – the index of the batch
unused – Deprecated argument. Will be removed in v1.7.
- on_validation_batch_end(*args, **kwargs)
Called in the validation loop after the batch.
- Parameters
outputs – The outputs of validation_step_end(validation_step(x))
batch – The batched data as it is returned by the validation DataLoader.
batch_idx – the index of the batch
dataloader_idx – the index of the dataloader
- optimizers(use_pl_optimizer: bool = True)
Returns the optimizer(s) that are being used during training. Useful for manual optimization.
- Parameters
use_pl_optimizer – If
True
, will wrap the optimizer(s) in aLightningOptimizer
for automatic handling of precision and profiling.- Returns
A single optimizer, or a list of optimizers in case multiple ones are present.
- precision: int
- prepare_data_per_node: bool
- property reporter
- set_clean_stop(clean_stop: bool)