Torchvision normalize

[mdabek-nvidia]

Category:

New feature

Description:

Torchvision's normalize operators implementation

Additional information:

Affected modules and functionalities:

Key points relevant for the review:

Tests:

• Existing tests apply
• New tests added
  • Python tests
  • GTests
  • Benchmark
  • Other
• N/A

Checklist

Documentation

• Existing documentation applies
• Documentation updated
  • Docstring
  • Doxygen
  • RST
  • Jupyter
  • Other
• N/A

DALI team only

Requirements

• Implements new requirements
• Affects existing requirements
• N/A

REQ IDs: N/A

JIRA TASK: N/A

[mdabek-nvidia]

@greptileai please review

[greptile-apps]

Greptile Summary

This PR adds Normalize and Pad operators to NVIDIA DALI's experimental torchvision-compatibility layer, along with their functional equivalents (normalize and pad). Both operators mirror the corresponding torchvision.transforms.v2 APIs and are backed by fn.normalize / ndd.normalize and fn.slice / ndd.slice respectively. Several issues from prior review rounds have been addressed (scalar-hue negativity guard in _ValidateHue, argument validation ordering in functional/normalize.py, isinstance fixes in gaussian_blur.py).

Key findings:

• P1 – functional/pad.py type-contract violation: The pad function's signature declares ndd.Batch as a valid input type, but the body immediately raises TypeError for any input that is not Image.Image or torch.Tensor. Any caller relying on the type hint for batch inputs will hit a confusing runtime error.
• P2 – Dead inplace parameter in _normalize: The private helper _normalize in functional/normalize.py accepts inplace but never uses it; the outer normalize already guards against this before calling the helper.
• P2 – Misleading error in get_padding: When a sequence of wrong length is passed (e.g. length 3), the error message shows the type (got <class 'list'>) rather than the length, making the cause harder to diagnose.
• Test coverage for both operators is thorough — CPU/GPU, multiple channel counts, PIL images, all four padding modes, and edge-case validation are all exercised.

Confidence Score: 4/5

Safe to merge after resolving the ndd.Batch type-contract violation in functional/pad.py.

A P1 issue remains: the pad functional advertises ndd.Batch support in its type annotation but raises TypeError unconditionally for it. This is a real API contract violation that will silently mislead users. The remaining findings are P2 style/clarity items that don't affect correctness for the supported input types. Prior P0/P1 concerns from earlier rounds have all been addressed.

dali/python/nvidia/dali/experimental/torchvision/v2/functional/pad.py — the pad function's type annotation and implementation are inconsistent regarding ndd.Batch.

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/functional/pad.py	New functional pad wrapper; type signature advertises ndd.Batch support but implementation raises TypeError for it — a P1 contract violation.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/normalize.py	New functional normalize wrapper; validates args before reshape (previous P1 fixed); minor unused inplace parameter in _normalize helper.
dali/python/nvidia/dali/experimental/torchvision/v2/normalize.py	New Normalize operator; correctly handles both HWC and CHW layouts by inspecting the last layout character; validators and inplace guard look correct.
dali/python/nvidia/dali/experimental/torchvision/v2/pad.py	New Pad operator with constant/edge/reflect/symmetric modes; get_padding error message is slightly misleading for wrong-length sequences; otherwise logic is sound.
dali/python/nvidia/dali/experimental/torchvision/v2/color.py	Adds scalar-hue negativity guard in _ValidateHue (fixing the previous P1) and renames internal classes to private naming convention.
dali/test/python/torchvision/test_tv_normalize.py	Comprehensive normalize tests covering channel counts, CPU/GPU, sequence types, mismatch errors, zero-std rejection, and spatial-broadcast correctness.
dali/test/python/torchvision/test_tv_pads.py	Thorough pad tests across all four padding modes, single and multi-border paddings, PIL images, and both CPU/GPU paths.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["User calls normalize(inpt, mean, std)
    or Normalize(mean, std)(inpt)"] --> B{Validate args}
    B -- "Invalid std/mean type\nor std==0" --> E1[Raise TypeError / ValueError]
    B -- Valid --> C{Input type}
    C -- "torch.Tensor (CHW)" --> D1["Reshape mean/std as\nmean[:, None, None]"]
    C -- "Pipeline DataNode\n(CHW layout)" --> D2["Reshape mean/std as\nmean[:, None, None]"]
    C -- "Pipeline DataNode\n(HWC layout)" --> D3["Reshape mean/std as\nmean[None, None, :]"]
    D1 --> F["ndd.normalize(input, mean, stddev)"]
    D2 --> G["fn.normalize(input, mean, stddev)"]
    D3 --> G

    A2["User calls pad(inpt, padding, mode)
    or Pad(padding, mode)(inpt)"] --> B2{Validate\npadding_mode\n& non-negative}
    B2 -- Invalid --> E2[Raise ValueError / TypeError]
    B2 -- Valid --> C2{Input type}
    C2 -- "Image.Image" --> P1["axes=[-3,-2]"]
    C2 -- "torch.Tensor" --> P2["axes=[-2,-1]"]
    C2 -- "ndd.Batch" --> P3["TypeError raised\n(type hint mismatch)"]
    P1 & P2 --> Q["ndd.slice with\nnegative anchor for padding\nborder_type from mode"]

Loading

_{Reviews (9): Last reviewed commit: "Improving type hints" | Re-trigger Greptile}

[mdabek-nvidia]

@greptileai please re-review