!build

@greptileai - please review

CI MESSAGE: [46735031]: BUILD STARTED

Greptile Summary

This PR adds ColorJitter and Grayscale operators (OO API) and to_grayscale / rgb_to_grayscale functional API entries to the experimental torchvision compatibility layer. As part of the change, layout-dimension extraction logic previously duplicated in resize.py and centercrop.py is consolidated into two new helpers in operator.py (get_HWC_from_layout_dynamic and get_HWC_from_layout_pipeline), and both existing operators are updated to use the new 4-tuple unpacking.

Key points:

• ColorJitter delegates to fn.color_twist and supports both CPU and GPU devices. Parameters are validated in __init__ via VerificationBCS / VerificationHue before the DALI pipeline is constructed.
• Grayscale uses fn.color_space_conversion (RGB→GRAY), fn.cat (1→3 channels), or fn.hsv with saturation=0 (3→3 desaturate), covering all four input/output channel combinations.
• Validation gap in VerificationHue: a negative scalar hue (e.g. hue=-0.1) silently passes validation but causes __init__ to produce an inverted range (0.1, -0.1). This in turn causes fn.random.uniform(range=(0.1, -0.1)) to fail at runtime. The fix is a simple non-negativity guard on the scalar branch, mirroring torchvision's own check.
• Dead else branch in _create_param: returns a bare float instead of a list, which would break _get_BCSH if reached (it is currently unreachable).
• The test suite is otherwise solid, covering all device/channel combinations and invalid-parameter cases — just missing the negative-scalar-hue case.

Confidence Score: 4/5

• Safe to merge once the negative-scalar hue validation gap in VerificationHue is addressed — all other changes are clean refactors or well-tested new features.
• The PR is well-structured and previous review concerns (integer inputs, HW layout IndexError) have been resolved. One targeted P1 fix remains: negative scalar hue bypasses validation and creates an inverted range at runtime. That path is easy to reproduce and easy to fix, making this a 4 rather than a 5.
• dali/python/nvidia/dali/experimental/torchvision/v2/color.py — specifically VerificationHue.verify and the _create_param else-branch.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["User input\n(PIL.Image / torch.Tensor)"] --> B["adjust_input decorator\ntransform_input()"]
    B --> C{"Input type"}
    C -->|"PIL.Image"| D["ndd.Tensor\nlayout=HWC"]
    C -->|"torch.Tensor 3D"| E["ndd.Tensor\nlayout=CHW"]
    C -->|"torch.Tensor >3D"| F["ndd.Batch\nlayout=CHW"]

    D & E & F --> G{"Operator"}

    G -->|"ColorJitter"| H["VerificationBCS + VerificationHue\n(in super().__init__)"]
    H --> I["_create_param(brightness/contrast/saturation)\n→ list[float, float]"]
    I --> J["hue scalar → (-hue, hue) tuple"]
    J --> K["_kernel: _get_BCSH\nsamples random factors"]
    K --> L["fn.color_twist()\nbright/contrast/sat/hue"]

    G -->|"Grayscale"| M["VerificationGSOutputChannels\n(in super().__init__)"]
    M --> N["preprocess_data:\nget_HWC_from_layout_pipeline\n→ (h, w, c, tensor)"]
    N --> O{"num_output_channels × c"}
    O -->|"1×3: RGB→Gray"| P["fn.color_space_conversion\nRGB→GRAY"]
    O -->|"1×1: no-op"| Q["pass"]
    O -->|"3×1: replicate"| R["fn.cat × 3 on C axis"]
    O -->|"3×3: desaturate"| S["fn.hsv saturation=0"]

    L & P & Q & R & S --> T["output ndd.Tensor/Batch"]
    T --> U["adjust_output\n→ PIL.Image or torch.Tensor"]

_{Reviews (4): Last reviewed commit: "Review fixes" | Re-trigger Greptile}

CI MESSAGE: [46735031]: BUILD PASSED

!build

CI MESSAGE: [46765427]: BUILD STARTED

CI MESSAGE: [46765427]: BUILD FAILED

@greptileai please re-review

!build

CI MESSAGE: [46875703]: BUILD STARTED

CI MESSAGE: [46875703]: BUILD PASSED