MolecularDiffusion.callbacks.train_helper

Attributes

logger

Classes

EMA
Queue
SP_regularizer	Self-paced regularizer for curriculum learning
gradient_clipping
gradient_clipping_0

Module Contents

class MolecularDiffusion.callbacks.train_helper.EMA(beta)

update_average(old, new)

update_model_average(ma_model, current_model)

beta

class MolecularDiffusion.callbacks.train_helper.Queue(max_len=50)

add(item)

mean()

std()

items = []

max_len = 50

class MolecularDiffusion.callbacks.train_helper.SP_regularizer(regularizer: str, lambda_: float = 10, lambda_2: float = 100, lambda_update_value: float = 50, lambda_update_step: int = 2500, polynomial_p: float = 1.5, warm_up_steps: int = 100)

Self-paced regularizer for curriculum learning :param regularizer: Regularizer to use. Options are:

• hard

• linear

• logaritmic

• logistic

Parameters:

• lambda (float) – Initial lambda value

• lambda_2 (float) – Initial lambda value for the second regularizer

• lambda_update_value (float) – Value to update lambda

• lambda_update_step (int) – Number of steps to update lambda

• polynomial_p (float) – Value of p for polynomial regularizer

• warm_up_steps (int) – Number of steps to use the regularizer

hard(losses: torch.Tensor)

hard_relax(losses: torch.Tensor)

linear(losses: torch.Tensor)

logaritmic(losses: torch.Tensor)

logistic(losses: torch.Tensor)

polynomial(losses: torch.Tensor)

update_lambda()

lambda_ = 10

lambda_2 = 100

lambda_update_step = 2500

lambda_update_value = 50

n_calls = 1

p = 1.5

regularizer

warm_up_steps = 100

class MolecularDiffusion.callbacks.train_helper.gradient_clipping(m=1, max_len=200)

FACTOR = 100

m = 1

max_grad_norm = None

max_grad_norms = []

max_len = 200

class MolecularDiffusion.callbacks.train_helper.gradient_clipping_0(m=1, max_len=200)

m = 1

max_grad_norm = None

max_grad_norms = []

max_len = 200

MolecularDiffusion.callbacks.train_helper.logger