Use imagenet dataset from hf in conformance tests

[AlexanderDokuchaev]

Changes

Add action for ptq conformance tests
Use HF dataset
Update metrics in result change in calibration subset

Tests

[Copilot]

Pull request overview

This PR updates the post-training (PTQ) conformance tests to use the ImageNet-1k validation split from Hugging Face datasets instead of relying on a locally prepared torchvision ImageFolder, and refactors image-classification pipelines to share calibration input transformation logic in the base class.

Changes:

• Switched ImageNet data loading in image-classification PTQ pipelines from local filesystem (ImageFolder) to a Hugging Face dataset loader targeting validation.
• Moved get_transform_calibration_fn logic from torchvision/timm pipelines into ImageClassificationBase.
• Updated PTQ test invocation to no longer pass a data_dir value (now None) and adjusted documentation accordingly.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/post_training/test_quantize_conformance.py	Stops passing data_dir for PTQ runs (now HF-based).
tests/post_training/README.md	Updates data preparation guidance to HF datasets/token usage.
tests/post_training/pipelines/image_classification_torchvision.py	Removes per-pipeline calibration transform (now inherited).
tests/post_training/pipelines/image_classification_timm.py	Removes per-pipeline calibration transform (now inherited).
tests/post_training/pipelines/image_classification_base.py	Implements HF ImageNet-1k val loader and adapts calibration/validation to dict-based batches; adds shared calibration transform.

Comments suppressed due to low confidence (1)

tests/post_training/pipelines/image_classification_base.py:163

• After switching validation to the HF ImageNet dataset, val_loader yields dict batches with {"image": ..., "label": ...} (as handled in _validate_ov), but _validate_torch_compile still iterates as for i, (images, target) in enumerate(val_loader). With dict batches, this unpacks keys instead of tensors and will break FX-backend validation. Update _validate_torch_compile to read data["image"] / data["label"] (matching the new dataset format).

    def _validate(self) -> None:
        val_dataset = hf_imagenet_1k_val(self.transform)
        val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=1, num_workers=2, shuffle=False)

        dataset_size = len(val_loader)

        # Initialize result tensors for async inference support.
        predictions = np.zeros(dataset_size)
        references = -1 * np.ones(dataset_size)

        if self.backend in FX_BACKENDS:
            predictions, references = self._validate_torch_compile(val_loader, predictions, references)
        else:
            predictions, references = self._validate_ov(val_loader, predictions, references, dataset_size)