huggingface/pytorch-image-models · warning

{name.capitalize()} range reversed. Swapping.

Error message

{name.capitalize()} range reversed. Swapping.

What it means

_validate_range in timm.data.naflex_transforms warns and auto-swaps when a (min, max) range is given reversed (value[0] > value[1]) — e.g. final_scale_range=(0.9, 0.3). The transform still works because the pair is reordered, but the warning flags likely config intent ambiguity.

Source

Thrown at timm/data/naflex_transforms.py:490

            width=target_hw[1],
            fill=self.fill,
            padding_mode=self.padding_mode,
        )

    def __repr__(self) -> str:
        return (f"{self.__class__.__name__}(patch_size={self.patch_size}, "
                f"max_sequence_len={self.max_sequence_len}, "
                f"divisible_by_patch={self.divisible_by_patch})")


def _validate_range(value, name, length=2):
    # Validate type and length
    if not isinstance(value, Sequence) or len(value) != length:
        raise ValueError(f"{name} should be a sequence of length {length}.")

    # Validate order
    if value[0] > value[1]:
        warnings.warn(f"{name.capitalize()} range reversed. Swapping.")
        return value[1], value[0]

    return value


class RandomResizedCropToSequence(torch.nn.Module):
    """
    Randomly crop the input image to a subregion with varying area and aspect ratio
    (relative to the original), then resize that crop to a target size. The target size
    is determined such that patchifying the resized image (with `patch_size`)
    does not exceed `max_seq_len` patches, while maintaining the aspect ratio of the crop.

    This combines aspects of torchvision's RandomResizedCrop with sequence length constraints.

    Args:
        patch_size (int or tuple[int, int]):
            Patch dimensions (patch_h, patch_w) for sequence length calculation.
        max_seq_len (int):

View on GitHub (pinned to 9a5261e31b)

Solutions

1. Rewrite ranges as (min, max): final_scale_range=(0.5, 1.0)
2. If a descending range is intentional for your schedule, note the swap or suppress the warning

Example fix

# before
transform = RandomResizedCropToSequence(..., final_scale_range=(0.9, 0.3))
# after
transform = RandomResizedCropToSequence(..., final_scale_range=(0.3, 0.9))

assert lo <= hi for lo, hi in [final_scale_range], 'range must be (min, max)'

When it happens

Trigger: Passing final_scale_range=(1.0, 0.5) or similar reversed bounds to RandomResizedCropToSequence/NaFlex transforms that use _validate_range.

Common situations: Copy-paste scale ranges from configs written as (max, min); thinking of scale as 'start scale' to 'end scale' during progressive resizing (where a descending range can be intentional).

Related errors

• Transform returned None for index {idx}. Skipping sample.
• final_scale_range values should ideally be between 0.0 and 1
• Final scale randomization ({scale_factor:.2f}) resulted in s
• range should be of kind (min, max)
• Input image must have positive dimensions, got H={height}, W

AI-assisted analysis of huggingface/pytorch-image-models@9a5261e31b (2026-08-27). Data as JSON: /api/errors/d5e995d0030ee64c. Report an issue: GitHub.