super-gradients/src/super_gradients/recipes/user_recipe_mnist_example.yaml

# The purpose of the example below is to demonstrate the use of registry for external objects for training.
# - We train mobilenet_v2 on a user dataset which is not defined in ALL_DATASETS using the dataloader registry.
# - We leverage predefined configs from cifar_10 training recipe in our repo.
#
# In order for the registry to work, we must trigger the registry of the user's objects by importing their module at
#   the top of the training script. Hence, we created a similar script to our classic train_from_recipe but with the imports
#   on top. Once imported, all the registry decorated objects will be resolved (i.e user_mnist_train will be resolved
#   to the dataloader of our user's)
#
# Differently from user_recipe_mnist_example, here we demonstrate how to use train_from_recipe, without the need to implement a DataLoader class for registry.
# Instead- we work straight with the user defined datasets, which is the simpler option when one does not need their own DataLoader implementation.
# We do so by Dropping the train_datalaoder, valid_dataloader fields from the recipe's config, while specifying the dataset arg in
#   train_dataloader_params, valid_dataloader_params.

defaults:
  - training_hyperparams: cifar10_resnet_train_params
  - dataset_params: cifar10_dataset_params
  - checkpoint_params: default_checkpoint_params
  - _self_

arch_params:
  num_classes: 10
  in_channels: 1

dataset_params:
  train_dataset_params:
    root: ./data/mnist
    train: True
    transforms:
      - RandomHorizontalFlip
      - ToTensor
    target_transform: null
    download: True

  train_dataloader_params:
    dataset: Mnist
    batch_size: 256
    num_workers: 8
    drop_last: False
    pin_memory: True

  val_dataset_params:
    root: ./data/mnist
    train: False
    transforms:
      - ToTensor
    target_transform: null
    download: True

  val_dataloader_params:
    batch_size: 512
    num_workers: 8
    drop_last: False
    pin_memory: True

resume: False
training_hyperparams:
  resume: ${resume}
  max_epochs: 3


ckpt_root_dir:

architecture: mobilenet_v2

experiment_name: mobilenet_v2_mnist


# THE FOLLOWING PARAMS ARE DIRECTLY USED BY HYDRA
hydra:
  run:
    # Set the output directory (i.e. where .hydra folder that logs all the input params will be generated)
    dir: ${hydra_output_dir:${ckpt_root_dir}, ${experiment_name}}