diffusion_models.diffusion_trainer

Module Contents

class DiffusionTrainer(model, dataset, optimizer, training_configuration, loss_function=F.l1_loss, log_configuration=LogConfiguration(), reverse_transforms=lambda x: ..., device='cuda')

A diffusion trainer framework.

This is a simplified framework for training a diffusion models.

Parameters:

• model (diffusion_models.models.base_diffusion_model.BaseDiffusionModel) – A diffusion model.

• dataset (Dataset) – A dataset to train on.

• optimizer (Optimizer) – The optimizer to use.

• training_configuration (TrainingConfiguration) – The training configuration to use.

• loss_function (Callable) – The loss function to use.

• log_configuration (LogConfiguration) – The logging configuration to use.

• reverse_transforms (Callable) – The reverse transforms to use.

• device (str) – The device to use.

model

The diffusion model to use.

optimizer

The optimizer to use.

loss_function

The loss function to use.

training_configuration

The training configuration to use.

device

The device to use.

dataloader

A torch dataloader.

scaler

A torch GradScaler object.

log_configuration

A LogConfiguration object.

checkpoint_path

The path to save checkpoints.

tensorboard_manager

A tensorboard manager instance.

reverse_transforms

A set of reverse transforms.

benchmark = True

save_checkpoint(epoch, checkpoint_name)

Save a checkpoint.

Parameters:

• epoch (int) – The current epoch.

• checkpoint_name (str) – The name of the checkpoint.

train()

Start the diffusion training.

log_to_tensorboard(metrics, global_step)

Log to tensorboard.

This method logs some useful metrics and visualizations to tensorboard.

Parameters:

• metrics (Dict[str, float]) – A dictionary mapping metric names to values.

• global_step (int) – The current global step.