Log actual augmentation transforms in debug.json

[Copilot]

_save_debug_information() was logging "dataloader_train.transform": "None" despite augmentations being applied. The transforms are actually stored in dataloader.generator.transforms, not at the top level.

Changes:

• Extract and log generator.transforms when present on dataloaders
• Add exception handling for stringify failures with exception type in error message
• Refactor to use local variable dl instead of repeated getattr(self, k) calls

Result:

{
  "dataloader_train.generator.transforms": "ComposeTransforms( transforms = [SpatialTransform(...), GaussianNoiseTransform(...), ...])",
  "dataloader_train.transform": "None"
}

Original prompt

---

This section details on the original issue you should resolve

<issue_title>_save_debug_information() prints "dataloader_train.transform": "None" everytime despite augmentation applied</issue_title>
<issue_description>As in the title, the debug transform info is always set to None.
I modified the debug function to print the actual augmentations applied. Figured it would be useful to someone else, especially when you're playing around with the augmentations its useful to log them:

Add one extra check in _save_debug_information() to include the internal data loader’s transforms:

nnUNet/nnunetv2/training/nnUNetTrainer/nnUNetTrainer.py

Lines 281 to 287 in 86606c5

	if k in ['dataloader_train', 'dataloader_val']:
	if hasattr(getattr(self, k), 'generator'):
	dct[k + '.generator'] = str(getattr(self, k).generator)
	if hasattr(getattr(self, k), 'num_processes'):
	dct[k + '.num_processes'] = str(getattr(self, k).num_processes)
	if hasattr(getattr(self, k), 'transform'):
	dct[k + '.transform'] = str(getattr(self, k).transform)

Suggested edit:

  if k in ['dataloader_train', 'dataloader_val']:
      if hasattr(getattr(self, k), 'generator'):
          dct[k + '.generator'] = str(getattr(self, k).generator)
          if hasattr(dl.generator, 'transforms'):
              try:
                  dct[k + '.generator.transforms'] = str(getattr(self, k).generator.transforms)
              except Exception as e:
                  dct[k + '.generator.transforms'] = f"Could not stringify generator.transforms: {e}"

      if hasattr(getattr(self, k), 'num_processes'):
          dct[k + '.num_processes'] = str(getattr(self, k).num_processes)
      if hasattr(getattr(self, k), 'transform'):
          dct[k + '.transform'] = str(getattr(self, k).transform)

So previously in the debug.json I would get:

"dataloader_train": "<batchgenerators.dataloading.nondet_multi_threaded_augmenter.NonDetMultiThreadedAugmenter object at 0x00000255CFF9C890>",
    "dataloader_train.generator": "<nnunetv2.training.dataloading.data_loader.nnUNetDataLoader object at 0x00000255CDD05DC0>",
    "dataloader_train.num_processes": "12",
    "dataloader_train.transform": "None",

Now I get:

"dataloader_train": "<batchgenerators.dataloading.nondet_multi_threaded_augmenter.NonDetMultiThreadedAugmenter object at 0x000002704922E060>",
    "dataloader_train.generator": "<nnunetv2.training.dataloading.data_loader.nnUNetDataLoader object at 0x000002704946E930>",
    "dataloader_train.generator.transforms": "ComposeTransforms( transforms = [Convert3DTo2DTransform(  ), SpatialTransform( patch_size = [192, 192], patch_center_dist_from_border = [0, 0], random_crop = False, p_elastic_deform = 0, elastic_deform_scale = (0, 0.2), elastic_deform_magnitude = (0, 0.2), p_rotation = 0.2, rotation = (-3.141592653589793, 3.141592653589793), p_scaling = 0.4, scaling = (0.7, 1.8), p_synchronize_scaling_across_axes = 1, p_synchronize_def_scale_across_axes = 0, bg_style_seg_sampling = False, mode_seg = 'bilinear', border_mode_seg = 'zeros', center_deformation = True, padding_mode_image = 'zeros' ), Convert2DTo3DTransform(  ), RandomTransform(p=0.3, transform=GaussianNoiseTransform( noise_variance = (0, 0.2), p_per_channel = 1, synchronize_channels = True )), RandomTransform(p=0.3, transform=GaussianBlurTransform( blur_sigma = (0.5, 1.0), benchmark = True, synchronize_channels = False, synchronize_axes = False, p_per_channel = 0.5, benchmark_use_fft = {}, benchmark_num_runs = 9 )), RandomTransform(p=0.15, transform=MultiplicativeBrightnessTransform( multiplier_range = BGContrast(contrast_range=(0.75, 1.25)), synchronize_channels = False, p_per_channel = 1 )), RandomTransform(p=0.15, transform=ContrastTransform( contrast_range = BGContrast(contrast_range=(0.75, 1.25)), preserve_range = True, synchronize_channels = False, p_per_channel = 1 )), RandomTransform(p=0.25, transform=SimulateLowResolutionTransform( scale = (0.5, 1), synchronize_channels = False, synchronize_axes = True, ignore_axes = (0,), allowed_channels = None, p_per_channel = 0.5, upmodes = {1: 'linear', 2: 'bilinear', 3: 'trilinear'} )), RandomTransform(p=0.1, transform=GammaTransform( gamma = BGContrast(contrast_range=(0.7, 1.5)), p_invert_image = 1, synchronize_channels = False, p_per_channel = 1, p_retain_stats = 1 )), RandomTransform(p=0.3, transform=GammaTransform( gamma = BGContrast(contrast_range=(0.7, 1.5)), p_invert_image = 0, synchronize_channels = False, p_per_channel = 1, p_retain_stats = 1 )), RemoveLabelTansform( segmentation_channels = None, label_value = -1, set_to = 0 ), DownsampleSegForDSTransform( ds_scales = [[np.float64(1.0), np.float64(1.0), np.float64(1.0)], [np.float64(1.0), np.float64(0.5), np.float64(0.5)], [np.float64(0.5), np.float64(0.25), np.float64(0.25)], [np.float64(0.25), np.float64(0.125), np.float64(0.125)], [np.float64(0.125), np.float64(0.0625), np.float64(0.0625)]] )] )",
    "dataloader_train.num_processes": "12",
    "dataloader_train.transform": "None",
    

</issue_description>

<agent_instructions>Create a PR that includes the suggested edit.</agent_instructions>

Comments on the Issue (you are @copilot in this section)

@TaWald Hey @anan1504 Thanks for this, it ...

• Fixes _save_debug_information() prints "dataloader_train.transform": "None" everytime despite augmentation applied #2933

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[TaWald]

Included the proposed changes to log the augmentation pipeline parameters.

[TaWald]

PR of #2933

[TaWald]

Adds details of training and validation augmentation pipeline to debug.json