Customize training and inference

This file contains mid-level information regarding various parameters that can be leveraged to customize the training/inference in GaNDLF. To see the default parameters for certain fields, see the default configs directory.

Training Parameters

The training parameters are defined in the configuration file. The following fields are supported:

model_config:  # Configuration for the model (see below) - required
modality:  # Modality of the input data, either 'rad' or 'histo' - required
num_epochs:  # Number of epochs to train the model
batch_size:  # Batch size for training, validation and test dataloaders
data_preprocessing:  # Data preprocessing configuration (see below)
data_augmentation:  # Data augmentation configuration (see below)
data_postprocessing:  # Data postprocessing configuration (see below)
dataloader_config:  # Dataloaders configuration (see below)
inference_parameters:
  batch_size:  # Batch size for inference
  n_images_to_generate:  # Number of images to generate during inference, unused in image-to-image models that require input images. This field can be a single value or a dictionary containing the number of images to generate for each class, for example {"1": 10, "2": 20}.
save_model_every_n_epochs:  # Save checkpoint every n epochs
compute: {} # Distributed training and mixed precision configuration (see below)

Model

Model configuration is expected to be in the following format:

model_config:
    model_name:  # Name of the model to use
    labeling_paradigm:  # Labeling paradigm for the model, either 'unlabeled', 'patient', or 'custom'
    architecture:  # Architecture of the model, customizing given model. Specifics are defined in the config of the given model.
    losses: # Loss functions to use (see below).
        - name:  # Name of the loss function
        - some_parameter: some_value
     # For models containing multiple losses (for example GANS), the losses are expected to be in the following format:
    losses:
        - discriminator: # Discriminator loss
            - name: # Name of the loss function
            - some_parameter: some_value
        - generator: # Generator loss
            - name: # Name of the loss function
            - some_parameter: some_value
    # Note that to use multiple losses, the model should be prepared via config to handle it via certain subloss name.
    - optimizers: # Optimizers to use (see below)
        - name:  # Name of the optimizer
        - some_parameter: some_value
    # For models containing multiple optimizers (for example GANS), the optimizers can be defined as losses above.
    - schedulers: # Schedulers to use (see below)
        - name:  # Name of the scheduler
        - some_parameter: some_value
    # For models containing multiple schedulers (for example GANS), the schedulers can be defined as losses above.
    - n_channels:  # Number of input channels
    - n_dimensions:  # Number of dimensions of the input data (2 or 3)
    - tensor_shape:  # Shape of the input tensor
    # This model config can support additional parameters that are specific to the model, for example:
    - save_eval_images_every_n_epochs:  # Save evaluation images every n epochs, useful to assess training progress of generative models. Implemented in i.e. DCGAN.

Optimizers

GaNDLF-Synth interfaces GaNDLF core framework for optimizers. See the optimizers directory for available optimizers. They support optimizer-specific configurable parameters, interfacing Pytorch Optimizers.

Schedulers

GaNDLF-Synth interfaces GaNDLF core framework for schedulers. See the schedulers directory for available schedulers. They support scheduler-specific configurable parameters, interfacing Pytorch Schedulers.

Losses

GaNDLF-Synth supports multiple loss functions. See the losses directory for available loss functions. They support loss-specific configurable parameters, interfacing Pytorch Loss functions.

Dataloader configs

GaNDLF-Synth supports separate dataloader parameters for training, validation, test and inference dataloaders. They support configurable parameters, interfacing Pytorch Dataloader. The following fields are supported:

dataloader_config:  # Dataloaders configuration (see below)
  shuffle:  # Whether to shuffle the data
  num_workers:  # Number of processes spawned to load the data
  pin_memory:  # Whether to pin the memory for CPU-GPU transfer
  timeout:  # Timeout for the dataloader processes
  prefetch_factor:  # Number of batches to prefetch by each worker
  persistent_workers:  # Whether to keep the worker processes alive between epochs

The fields for specific dataloaders are expected to be in the following format:

dataloader_config:
    train:
        - some_parameter: some_value
    val:
        - some_parameter: some_value
    test:
        - some_parameter: some_value
    inference:
        - some_parameter: some_value

If given dataloader is not configured explicitly, the default values are used (see above).

Data preprocessing

GaNDLF-Synth interfaces GaNDLF core framework for data preprocessing. To see available data preprocessing options, see here. Separate preprocessing parameters can be defined for each dataloader (train, val, test, inference) as follows:

data_preprocessing:
    train:
        - some_transform: some_value
    val:
        - some_transform: some_value
    test:
        - some_transform: some_value
    inference:
        - some_transform: some_value

Data Augmentation

GaNDLF-Synth interfaces GaNDLF core framework for data augmentation. To see available data augmentation options, see here Augmentations are applied only to the training dataloader.

data_augmentation:
    train:
        - some_transform: some_value

Post processing

GaNDLF-Synth interfaces GaNDLF core framework for post processing. To see available post processing options, see here. Post-processing is applied only to the inference dataloader.

data_postprocessing:
    inference:
        - some_transform: some_value

Distributed Training

For detalis on using distributed training, see the usage page.