Apply data augmentation on MVTec

[astewzxc]

I tried to apply transforms on MVTec dataset according to Add transform_config to the main config.yaml file.
But I got an error:
RuntimeError: Given groups=1, weight of size [64, 3, 7, 7], expected input[32, 224, 224, 3] to have 3 channels, but got 224 channels instead

I'm not sure if I done right. I created a transform pipeline using albumentations serialization augmentation pipeline, saved it as 'transform.json' and put the file into anomalib fold.
Then change the 'null' at transform_config of config.yaml to 'transform.json'

I've tried patchcore and cflow, neither worked
I assume that the channels of the images changed, but not sure why this happened and how to fix it.

[djdameln]

Anomalib models expect their inputs to be of shape [Batch x Channel x Height x Weight]. It seems that one of your transforms has changed the order of the dimensions of the input tensor. Would you mind sharing your transform.json so we could have a closer look?

[astewzxc]

{"__version__": "1.2.1", "transform": {"__class_fullname__": "Compose", "p": 1.0, "transforms": [{"__class_fullname__": "RGBShift", "always_apply": false, "p": 0.5, "r_shift_limit": [-20, 20], "g_shift_limit": [-20, 20], "b_shift_limit": [-20, 20]}], "bbox_params": null, "keypoint_params": null, "additional_targets": {}}}

These are all contents of the file. I applied only RGBShift, nothing else.

[djdameln]

I see. When using a custom transform configuration, the transforms specified in the .json file replace the default transforms that are used when no custom configuration is passed. These default transforms consist of Resize (resizing to specified image size), Normalize (normalizing to ImageNet stats), and ToTensorV2 (converting the inputs to a tensor with the appropriate dimension ordering). The ToTensorV2 transform is always required, because it prepares the images to be properly handled by the first layer of the PyTorch model.

Looking at it now, I feel our design regarding custom transform configurations is probably not very intuitive and should be improved, or at least better documented. Ideally Anomalib should automatically append mandatory transforms such as ToTensorV2 when these are not present in the user provided .json file. For now however, it is necessary to always include the ToTensorV2 transform in your transforms json.

In your case, your problem would be solved by adding both the ToFloat and the ToTensorV2 transforms to your .json file. The ToFloat transform is needed to make sure that the elements of the input tensor are of the right data type. By default this is accounted for by the Normalize transform. Please try the following config to which I've added the ToFloat and ToTensorV2 transforms, and let us know if it addresses your issue:

{
  "__version__": "1.2.1",
  "transform": {
    "__class_fullname__": "Compose",
    "p": 1.0,
    "transforms": [
      {
        "__class_fullname__": "RGBShift",
        "always_apply": false,
        "p": 0.5,
        "r_shift_limit": [
          -20,
          20
        ],
        "g_shift_limit": [
          -20,
          20
        ],
        "b_shift_limit": [
          -20,
          20
        ]
      },
      {
        "__class_fullname__": "ToFloat",
        "always_apply": true,
        "p": 1.0,
        "max_value": null
      },
      {
        "__class_fullname__": "ToTensorV2",
        "always_apply": true,
        "p": 1.0,
        "transpose_mask": false
      }
    ],
    "bbox_params": null,
    "keypoint_params": null,
    "additional_targets": {}
  }
}

[astewzxc]

It works, thanks for the help!