documentations.md

SPAGHETTI Documentations

Documentations for classes and method for SPAGHETTI. For installing details, see README.md. For a tutorial on sample usage, see the directory at ./tutorials/

SPAGHETTI.dataset

Set up the dataset to train SPAGHETTI

TrainingDataset

The dataset class for the SPAGHETTI model training. This class is inherited from torch.utils.data.Dataset

• args:
  • path_1: list of strings, the paths to the images in domain 1
  • path_2: list of strings, the paths to the images in domain 2
  • transform_1: the transformation for domain 1 images, in torchvision.transforms.v2
  • transform_2: the transformation for domain 2 images, in torchvision.transforms.v2
  • num_sample: int, optional, the number of images to sample from each domain

SPAGHETTI.train

Modules for training SPAGHETTI

train_spaghetti

The function to automatically handle all SPAGHETTI training using pytorch_lightning.

• args:
  • train_loader: the PyTorch Dataloader for the training dataset
  • val_loader: the PyTorch Dataloader for the validation dataset
  • batch_size: int, the batch size for the model, default 1
  • weights: list of floats, the weights for the loss functions in the order of GAN loss, cycle loss, identity loss, and SSIM loss. Default [1.0, 10.0, 5.0, 10.0]
  • lr: float, the learning rate for the model, default 0.0002
  • save_dir: str, the directory to save the model checkpoints and logs. Default current directory
  • epochs: int, the number of epochs to train the model, default 100
  • name: str, the name of the model for the logger, default "my_spaghetti"
  • num_nodes: int, the number of nodes to train the model, default 1
  • ngpus_per_node: int, the number of GPUs per node, default "auto" to use all the available GPUs
• returns: None
• size effects: Run the training scripts and save model checkpoints, loss logs, and sampling images to save_dir.

SPAGHETTI.inferences

Modules for performing inferences with SPAGHETTI

Spaghetti

The main class housing the architecture of SPAGHETTI for inference.

• args:
  • model_path: str, the path to the model checkpoint
• returns: The SPAGHETTI inference model, with the following methods:
  • pre_processing Method to pre-process the image for SPAGHETTI transformation. This is a static method.
    • args:
      • imgs: list[PIL.Image or torch.Tensor or numpy.ndarray], the images to perform the pre-processing
      • transform: None, "default", or callable of torchvision.transform.v2, the transformation to perform on the images. If None, no transformation is performed. If "default", the default transformation (converting to ```torch.Tensor`` with range [0,1], normalize with mean and std of 0.5, and resize to (256,256)) is performed.
    • return: list[torch.Tensor], the images after the pre-processing. Each image will have the shape of [C, H, W]
  • inference Method to translate the images using the SPAGHETTI model initialized with the model checkpoint.
    • args:
      • img: list[torch.Tensor] or DataLoader, the image(s) to perform the inference. Images have to preprocessed first using the pre_processing method. For larger dataset, it is strongly recommended to use DataLoader to allow the inference using a smaller memory.
      • names: list[str], the names of the images to be saved
      • save_path: str or None. If str, images will be saved to the path to after the transformation. If None, transalted images will only be returned but not saved
    • return: list[torch.Tensor], the images after the SPAGHETTI transformaton when save_path is None, otherwise ```None``.
    • side effects: If save_path is not None, save the images to the specified path.