fastvideo.training.ode_causal_pipeline

fastvideo.training.ode_causal_pipeline#

Module Contents#

Classes#

ODEInitTrainingPipeline

Training pipeline for ODE-init using precomputed denoising trajectories.

Functions#

main

Data#

logger

API#

class fastvideo.training.ode_causal_pipeline.ODEInitTrainingPipeline(model_path: str, fastvideo_args: fastvideo.fastvideo_args.TrainingArgs, required_config_modules: list[str] | None = None, loaded_modules: dict[str, torch.nn.Module] | None = None)[source]#

Bases: fastvideo.training.training_pipeline.TrainingPipeline

Training pipeline for ODE-init using precomputed denoising trajectories.

Supervision: predict the next latent in the stored trajectory by

  • feeding current latent at timestep t into the transformer to predict noise

  • stepping the scheduler with the predicted noise

  • minimizing MSE to the stored next latent at timestep t_next

Initialization

Initialize the pipeline. After init, the pipeline should be ready to use. The pipeline should be stateless and not hold any batch state.

initialize_pipeline(fastvideo_args: fastvideo.fastvideo_args.FastVideoArgs)[source]#
initialize_training_pipeline(training_args: fastvideo.fastvideo_args.TrainingArgs)[source]#
initialize_validation_pipeline(training_args: fastvideo.fastvideo_args.TrainingArgs)[source]#
set_schemas()[source]#
train_one_step(training_batch)[source]#
visualize_intermediate_latents(training_batch: fastvideo.pipelines.pipeline_batch_info.TrainingBatch, training_args: fastvideo.fastvideo_args.TrainingArgs, step: int)[source]#

Add visualization data to wandb logging and save frames to disk.

fastvideo.training.ode_causal_pipeline.logger[source]#

‘init_logger(…)’

fastvideo.training.ode_causal_pipeline.main(args) → None[source]#
Contents