Reshape per-channel ImageType.scale to broadcast over channel-first input (fixes #2461)

[LeSingh1]

Summary

ImageType.scale is documented as

scale: float or list of floats

The scaling factor for all values in the image channels.

…matching bias. But when a user actually supplied a per-channel list (e.g. [1/127.5, 1/127.5, 1/127.5] for an RGB MobileNetV2 input), the mil_backend::insert_image_preprocessing_ops pass dropped it through np.array(...) and emitted an mb.mul whose scale tensor still had shape (3,). That tensor then tried to broadcast against the channel-first input (N, 3, H, W) and failed in backend_mlprogram with:

ValueError: Incompatible dim 3 in shapes (1, 3, 224, 224) vs. (1, 1, 1, 3)

Fix

The neighbouring per-channel bias path already reshapes the constant to (3, 1, 1) for rank-3 inputs and (1, 3, 1, 1) for rank-4 inputs. Apply the same reshape to scale:

scale_arr = np.array(input_type.scale, dtype=input_nptype)
if scale_arr.ndim > 0 and input_type.color_layout not in (
    _input_types.ColorLayout.GRAYSCALE,
    _input_types.ColorLayout.GRAYSCALE_FLOAT16,
):
    if len(last_output.shape) == 3:
        scale_arr = scale_arr.reshape([3, 1, 1])
    elif len(last_output.shape) == 4:
        scale_arr = scale_arr.reshape([1, 3, 1, 1])
    else:
        raise TypeError("Unsupported rank for image input type.")
last_output = mb.mul(x=last_output, y=scale_arr, name=input_var.name + "__scaled__")

Scalar scales (scale_arr.ndim == 0) and grayscale layouts skip the reshape — no other code paths change.

Tests

Two new regression tests under TestImagePreprocessingPass:

• test_program_rgb_per_channel_scale — rank-4 RGB input with scale=[1/127.5]*3, bias=[-1, -1, -1] (the exact pattern from ImageType causes an error in TensorFlow 2 model conversion with a scale parameter provided as list #2461); asserts the inserted mul.y has shape (1, 3, 1, 1) and values match the input list.
• test_program_rgb_per_channel_scale_rank3 — rank-3 sibling; asserts the (3, 1, 1) reshape path mirroring the rank-3 bias case.

Both pass locally. The full TestImagePreprocessingPass suite (14 tests including the existing scalar-scale and per-channel-bias cases) also continues to pass:

$ pytest coremltools/converters/mil/backend/mil/passes/test_passes.py::TestImagePreprocessingPass -v
============================== 14 passed in 0.05s ==============================

Issue

Fixes #2461