[Bug]: Transformations behaviour possibly not expected when predicting with fresh dataloader

[blaz-r]

Describe the bug

Inside the engine code the trasnforms from the datamodule and dataloader are taken before the ones from the model: https://github.com/openvinotoolkit/anomalib/blob/2bd2842ec33c6eedb351d53cf1a1082069ff69dc/src/anomalib/engine/engine.py#L382-L398

This can cause problems if datamodule was never used inside the trainer. In that case looking at the following code, the transforms returned are just resize (since there is no trainer to take the correct model transforms from). This leads to missing the potential normalization from the model: https://github.com/openvinotoolkit/anomalib/blob/2bd2842ec33c6eedb351d53cf1a1082069ff69dc/src/anomalib/data/base/datamodule.py#L266-L277

This happens only if you are calling setup and dataloader from the datamodule outside trainer (as that is not set in this case) like this:

datamodule= MVTec(..)
datamodule.setup() # this calls self.train_transforms mentioned above
datamodule.test_dataloader() # only resize transofmr here

There is also code below to reproduce this.

A workaround I have at the moment is doing this:

dataloader.dataset.transform = None

This way the dataloader transforms are ignored and the ones from model are taken. Another possibility that I see is by manually setting the transforms of the datamodule with model.configure_transforms(image_size).

Dataset

N/A

Model

N/A

Steps to reproduce the behavior

    #### CONFIGURE THIS ####
    mvtec_path = "../datasets/MVTec"
    #####

    data = MVTec(root=mvtec_path, image_size=(42, 42), category="bottle", num_workers=0)
    data.setup()

    print(data.test_dataloader().dataset.transform)

OS information

OS information:

• OS: Win11
• Python version: 3.10.3
• Anomalib version: 1.2.0dev
• PyTorch version: 2.3.0
• CUDA/cuDNN version: x
• GPU models and configuration: x
• Any other relevant information: x

Expected behavior

I would expect that in this case, the model transforms would take priority over the ones in dataloader, but I see how this would cause trouble in case of custom transforms inside datamodule.

Screenshots

No response

Pip/GitHub

GitHub

What version/branch did you use?

1.2.0dev

Configuration YAML

/

Logs

/

Code of Conduct

• [X] I agree to follow this project's Code of Conduct

[samet-akcay]

@djdameln any thoughts?

[blaz-r]

Is there any updates on this one?

[samet-akcay]

@blaz-r, not to address this one specifically, but @djdameln is working on some changes, which might address this

[blaz-r]

Okay. I'll use the workaround for now and we'll see when those changes are added.