peal.training.interfaces.TrainingConfig

class peal.training.interfaces.TrainingConfig

Bases: BaseModel

Hyper-parameters and switches of one predictor training run.

Besides max_epochs, learning_rate, dropout, optimizer and the three batch sizes, the config holds the ModelTrainer switches for adversarial training (adv_training, attack_epsilon, attack_num_steps, input_noise_std), mixup and label smoothing, class balancing, diffusion-based augmentation (diffusion_augmented with sampling_time_fraction / num_discretization_steps) and the early_stopping_goal metric the best checkpoint is picked by. steps_per_epoch enables the DataloaderMixer episode length; epoch, global_train_step and global_validation_step are counters the trainer and logger increment during training.

Methods

Attributes

adv_training
attack_epsilon
attack_num_steps
class_balanced
concatenate_batches
diffusion_augmented
dropout	The dropout rate the model is trained with.
early_stopping_goal
epoch	The current epoch of the model training.
global_train_step	Logs how many steps the model was trained with already.
global_validation_step	Logs how many steps the model was validated for.
input_noise_std
label_smoothing
learning_rate	The learning rate the model is trained with.
max_epochs	The learning rate the model is trained with.
mixup_alpha
no_grad_attack
num_discretization_steps
num_noise_vec
num_workers	DataLoader worker processes for the loaders built from this config.
optimizer	The optimizer used for training the model.
regulization_level
sampling_time_fraction
steps_per_epoch	The number of iterations per episode when using DataloaderMixer.
test_batch_size	The test batch size.
train_batch_size	The train batch size.
train_on_test
use_mixup
val_batch_size	The val batch size.

max_epochs: pydantic.PositiveInt = 15

The learning rate the model is trained with.

learning_rate: float = 0.0001

The learning rate the model is trained with.

dropout: float = 0.5

The dropout rate the model is trained with.

num_workers: int = 0

DataLoader worker processes for the loaders built from this config. 0 keeps loading in the training process (the historical behaviour); see peal.data.dataloaders.resolve_num_workers for the $PEAL_NUM_WORKERS override and the reason the default is conservative.

global_train_step: int = 0

Logs how many steps the model was trained with already.

global_validation_step: int = 0

Logs how many steps the model was validated for.

epoch: int = -1

The current epoch of the model training.

optimizer: str = 'Adam'

The optimizer used for training the model.

train_batch_size: pydantic.PositiveInt = 1

The train batch size. Can either be set manually or be left empty and calculated by adaptive batch_size.

val_batch_size: pydantic.PositiveInt = 1

The val batch size. Can either be set manually or be left empty and calculated by adaptive batch_size.

test_batch_size: pydantic.PositiveInt = 1

The test batch size. Can either be set manually or be left empty and calculated by adaptive batch_size.

steps_per_epoch: None | pydantic.PositiveInt = None

The number of iterations per episode when using DataloaderMixer. If it is not set, the DataloaderMixer is not used and the value implicitly becomes dataset_size / batch_size

concatenate_batches: bool = True

adv_training: bool = False

input_noise_std: float = 0.1

num_noise_vec: int = 1

no_grad_attack: bool = False

attack_epsilon: float = 1.0

attack_num_steps: int = 5

train_on_test: bool = False

class_balanced: bool = False

use_mixup: bool = False

mixup_alpha: float = 1.0

label_smoothing: float = 0.0

early_stopping_goal: str = 'average_accuracy'

diffusion_augmented: bool = False

sampling_time_fraction: float = 0.3

num_discretization_steps: int = 20

regulization_level: float = 1.3