Skip to main content
Version: 0.3

PyTorchCompressionEngine​

CLASS - PyTorchCompressionEngine() :

Object that contains the Entry points to CLIKA Compression Engine.

Methods​

deploy​

deploy(

) ‑> str

Deploy the CLIKA State to the Format specified when the CLIKA state was created.

Parameters:

  • clika_state_path (Union[str, Path]) - Path to the CLIKA state directory or file
  • output_dir_path (Union[str, Path]) - Where to save the generated Model file
  • file_suffix (Optional[str]) - Any string to prepend to the filename
  • input_shapes (Optional[Union[list, tuple, List[list], List[tuple], Tuple[tuple], Tuple[list]]]) - Sequence or Sequence of Sequences that indicate the Input shape of the resulting Model. In the case of a single input model, you could pass tuple or a list such as: (1, 3, 224, 224) In-case of a multiple-input model, you could pass [(1, 3, 224, 224), (1, 3, 120)] corresponding to the input shapes that your model takes. A 'None' value represents dynamic shape.
  • graph_author (Optional[str]) - Optional string to specify who created the Graph for the Model on the resulting File
  • graph_description (Optional[str]) - Optional string to specify Description for the Model on the resulting File
  • kwargs (Optional[dict]) - Not used at the moment.

Returns:

  • str: Path to the deployed Model

optimize​

optimize(

) ‑> str

Entry point for running Compression.

Parameters:

  • output_path (str/Path) - Path to save the files being generated by the library
  • settings (Settings) - Settings that contains all the necessary information to run the model
  • model (torch.nn.Module) - torch nn Module
  • init_training_dataset_fn (Callable[[], Any]) - Pointer to a function that initializes the Dataloaders
  • init_evaluation_dataset_fn (Optional[Callable[[], Any]]) - Pointer to a function that initializes the Dataloaders
  • optimizer (torch.optim.Optimizer) - Optimizer object. In PyTorch it's an object from: torch.optim... When Resuming, 'optimizer' can be set to None unless it is a custom Optimizer.
  • training_losses (Dict[str, Callable]) - Dictionary mapping between a Name to a Loss Object used for Training. Can be a function that accepts (preds, target) or a Metric that implements an interface like torchmetrics.
  • training_metrics (Dict[str, Callable]) - Dictionary mapping between a Name to a Metric Object used for Training. Can be a function that accepts (preds, target) or a Metric that implements an interface like torchmetrics.
  • evaluation_losses (Dict[str, Callable]) - Dictionary mapping between a Name to a Loss Object used for Evaluation. Can be a function that accepts (preds, target) or a Metric that implements an interface like torchmetrics.
  • evaluation_metrics (Dict[str, Callable]) - Dictionary mapping between a Name to a Metric Object used for Evaluation. Can be a function that accepts (preds, target) or a Metric that implements an interface like torchmetrics.
  • callbacks (List[BaseCallback]) - Callbacks that will be called during compression to update the status of it.
  • kwargs (Optional[dict]) - Not used at the moment

Returns:

  • str: Path to the latest CLIKA state

resume​

resume(

) ‑> str

Resumes a Compression Job from Given path. This can be used to fine-tune the model using more Data or just resume a run that had to be stopped. This can be used to finetune on more Data or just resume a run that had to be stopped.

Parameters:

  • clika_state_path (Union[str, Path]) - Path to the CLIKA state directory or file
  • init_training_dataset_fn (Callable[[], Any]) - Pointer to a function that initializes the Dataloaders
  • init_evaluation_dataset_fn (Optional[Callable[[], Any]]) - Pointer to a function that initializes the Dataloaders
  • optimizer (torch.optim.Optimizer) - Optimizer object. In PyTorch it's an object from: torch.optim... When Resuming, 'optimizer' can be set to None unless it is a custom Optimizer.
  • training_losses (Dict[str, Callable]) - Dictionary mapping between a Name to a Loss Object used for Training. Can be a function that accepts (preds, target) or a Metric that implements an interface like torchmetrics.
  • training_metrics (Dict[str, Callable]) - Dictionary mapping between a Name to a Metric Object used for Training. Can be a function that accepts (preds, target) or a Metric that implements an interface like torchmetrics.
  • evaluation_losses (Dict[str, Callable]) - Dictionary mapping between a Name to a Loss Object used for Evaluation. Can be a function that accepts (preds, target) or a Metric that implements an interface like torchmetrics.
  • evaluation_metrics (Dict[str, Callable]) - Dictionary mapping between a Name to a Metric Object used for Evaluation. Can be a function that accepts (preds, target) or a Metric that implements an interface like torchmetrics.
  • callbacks (List[BaseCallback]) - Callbacks that will be called during compression to update the status of it.
  • multi_gpu (bool) - whether or not to do Multi-GPU compression
  • new_settings (Settings) - object to use for the Compression Run. TrainingSettings and DeploymentSettings will be updated with this.
  • resume_start_epoch (Optional[int]) - Will override the Epoch present in the CLIKA State directory at Path.
  • kwargs (Optional[dict]) - Unsupported for now.

Returns:

  • str: Path to latest CLIKA State path