Grayscale Image Source

[rumanto00]

Is possible to use grayscale image on anomalib with RGB Resnet Transfer Learning?
and where the setting on anomalib for 1 channel image?

Thanks

[Jonas1302]

I haven't tested it but you could try to use a transform_config like they did with DRAEM (see its config.yaml and transform_config.yaml).
Adding ToRGB to the inner transforms list (right before ToFloat) may do it.

[samet-akcay]

@rumanto00, I agree with @Jonas1302. You could create your custom transforms and add it to your config.yaml file as @Jonas1302 suggested.

For example, here is a code snippet how you could generate those transform configs via albumentations.

from skimage import data
import albumentations as A
from albumentations.pytorch import ToTensorV2

image = data.camera()

transforms = A.Compose([A.ToRGB(), ToTensorV2()])
transformed_image = transforms(image=image)["image"]

# (512, 512) torch.Size([3, 512, 512])
print(image.shape, transformed_image.shape)

A.save(transforms, "transforms.yaml", data_format="yaml")

This would save the transforms to transforms.yaml file, which would be something like the following:

__version__: 1.3.0
transform:
  __class_fullname__: Compose
  additional_targets: {}
  bbox_params: null
  is_check_shapes: true
  keypoint_params: null
  p: 1.0
  transforms:
  - __class_fullname__: ToRGB
    always_apply: true
    p: 1.0
  - __class_fullname__: ToTensorV2
    always_apply: true
    p: 1.0
    transpose_mask: false

NOTE:
ToRGB has just been added to albumentations. You will need to install albumentations package from source since this change is not on PyPI yet.

[papago2355]

Hello. Is this 1channel config improved your results?