class Config(ConfigBase):
#: Training epochs
epochs: int = 10
#: Stop after how many epochs when the eval metric is not improving
early_stop_after: int = 0
#: Clip gradient norm if set
max_clip_norm: Optional[float] = None
#: Whether metrics on training data should be computed and reported.
report_train_metrics: bool = True
#: Target time limit for training, default (None) to no time limit.
target_time_limit_seconds: Optional[int] = None
#: Whether to do evaluation and model selection based on it.
do_eval: bool = True
#: Number of samples for logging training progress.
num_samples_to_log_progress: int = 1000
#: Number of forward & backward per batch before update gradients, the
#: actual_batch_size = batch_size x num_accumulated_batches
num_accumulated_batches: int = 1
#: Define epoch as a fixed number of batches. Subsequent epochs will continue
#: to iterate through the data, cycling through it when they reach the end.
#: If not set, use exactly one pass through the dataset as one epoch.
#: This configuration only affects the train epochs, test and eval
#: will always test their entire datasets.
num_batches_per_epoch: Optional[int] = None
#: config for optimizer, used in parameter update
optimizer: Optimizer.Config = Adam.Config()
scheduler: Optional[Scheduler.Config] = None
sparsifier: Optional[Sparsifier.Config] = None
#: Define arguments for fp16 training. A fp16_optimizer will be created
#: and wraps the original optimizer, which will scale loss during
#: backward and master weight will be maintained on original optimizer.
#: https://arxiv.org/abs/1710.03740
fp16_args: FP16Optimizer.Config = FP16OptimizerFairseq.Config()
class Config(ConfigBase):
#: Training epochs
epochs: int = 10
#: Stop after how many epochs when the eval metric is not improving
early_stop_after: int = 0
#: Clip gradient norm if set
max_clip_norm: Optional[float] = None
#: Whether metrics on training data should be computed and reported.
report_train_metrics: bool = True
#: Target time limit for training, default (None) to no time limit.
target_time_limit_seconds: Optional[int] = None
#: Whether to do evaluation and model selection based on it.
do_eval: bool = True
#: if do_eval, do we load the best model state dict after training or just
# use the latest model state
load_best_model_after_train: bool = True
#: Number of samples for logging training progress.
num_samples_to_log_progress: int = 1000
#: Number of forward & backward per batch before update gradients, the
#: actual_batch_size = batch_size x num_accumulated_batches
num_accumulated_batches: int = 1
#: Define epoch as a fixed number of batches. Subsequent epochs will continue
#: to iterate through the data, cycling through it when they reach the end.
#: If not set, use exactly one pass through the dataset as one epoch.
#: This configuration only affects the train epochs, test and eval
#: will always test their entire datasets.
num_batches_per_epoch: Optional[int] = None
#: config for optimizer, used in parameter update
optimizer: Optimizer.Config = Adam.Config()
scheduler: Optional[Scheduler.Config] = None
sparsifier: Optional[Sparsifier.Config] = None
#: Define arguments for fp16 training. A fp16_optimizer will be created
#: and wraps the original optimizer, which will scale loss during
#: backward and master weight will be maintained on original optimizer.
#: https://arxiv.org/abs/1710.03740
fp16_args: FP16Optimizer.Config = FP16OptimizerFairseq.Config()
use_tensorboard: bool = False
find_unused_parameters: bool = True
#: Set a discriminative learning rate for some of the parameters in model.
#: If None, all parameters will have the same lr.
discriminative_lr: Optional[float] = None
#: Model parameters match any patterns in the list will have discriminative_lr
#: Parameters not matching any patterns will have default lr.
# E.g. ["decoder.mlp.0", "decoder.mlp.3"]
discriminative_lr_params_pattern: Optional[List[str]] = None
#: Model parameters match any patterns in the list will have lr = 0.0
freeze_params_pattern: Optional[List[str]] = None
