committing to switch branch, test_transform pipeline still to be fixed

InnerEye/ML/SSL/datamodules_and_datasets/transforms_utils.py

@@ -30,7 +30,7 @@ def get_ssl_transforms_from_config(config: CfgNode,
 
  :param config: configuration defining which augmentations to apply as well as their intensities.
  :param return_two_views_per_sample: if True the resulting transforms will return two versions of each sample they
- are called on. If False, simply return one transformed version of the sample.
+ are called on. If False, simply return one transformed version of the sample centered and cropped.
  :param use_training_augmentations_for_validation: If True, use augmentation at validation time too.
  This is required for SSL validation loss to be meaningful. If False, only apply basic processing step
  (no augmentations)

InnerEye/ML/SSL/lightning_containers/ssl_container.py

@@ -187,7 +187,7 @@ def _create_ssl_data_modules(self, is_ssl_encoder_module: bool) -> InnerEyeVisio
  """
  Returns torch lightning data module for encoder or linear head
 
- :param is_ssl_encoder_module: whether to return the data module for SSL training or for linear heard. If true,
+ :param is_ssl_encoder_module: whether to return the data module for SSL training or for linear head. If true,
  :return transforms with two views per sample (batch like (img_v1, img_v2, label)). If False, return only one
  view per sample but also return the index of the sample in the dataset (to make sure we don't use twice the same
  batch in one training epoch (batch like (index, img_v1, label), as classifier dataloader expected to be shorter
@@ -225,8 +225,8 @@ def _get_transforms(self, augmentation_config: Optional[CfgNode],
  examples.
  :param dataset_name: name of the dataset, value has to be in SSLDatasetName, determines which transformation
  pipeline to return.
- :param is_ssl_encoder_module: if True the transformation pipeline will yield two version of the image it is
- applied on. If False, return only one transformation.
+ :param is_ssl_encoder_module: if True the transformation pipeline will yield two versions of the image it is
+ applied on and it applies the same transformations for validation. If False, return only one transformation.
  :return: training transformation pipeline and validation transformation pipeline.
  """
  if dataset_name in [SSLDatasetName.RSNAKaggleCXR.value,
@@ -252,7 +252,7 @@ def _get_transforms(self, augmentation_config: Optional[CfgNode],
  expand_channels=False,
  )
  logging.warning(f"Dataset {dataset_name} unknown. The config will be consumed by "
- f"get_ssl_transforms() to create the augmentation pipeline, make sure"
+ f"get_ssl_transforms() to create the augmentation pipeline, make sure "
  f"the transformations in your configs are compatible. ")
  else:
  raise ValueError(f"Dataset {dataset_name} unknown and no config has been passed.")

Tests/ML/augmentations/test_transform_pipeline.py

@@ -133,6 +133,7 @@ def test_create_transform_pipeline_from_config(expand_channels: bool) -> None:
  all_transforms.insert(0, ExpandChannels())
  else:
  fake_3d_array = np.stack([fake_cxr_as_array for i in range(3)])
+ # TODO this is raising an error - understands what shapes/values you need in here
  fake_cxr_image = PIL.Image.fromarray(fake_3d_array).convert("RGB")
 
  np.random.seed(3)
@@ -159,6 +160,7 @@ def test_create_transform_pipeline_from_config(expand_channels: bool) -> None:
  assert torch.isclose(expected_transformed, transformed_image).all()
 
  # Test the evaluation pipeline
+ # TODO why this is not parametrized?
  transformation_pipeline = create_transforms_from_config(cxr_augmentation_config, apply_augmentations=False,
  expand_channels=expand_channels)
  transformed_image = transformation_pipeline(image)