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 a Normalize operator and normalize functional to the DALI torchvision compatibility layer, mirroring torchvision.transforms.v2.Normalize. Both a class-based (Normalize) and functional (normalize) API are provided, with CHW/HWC layout detection handled consistently with the existing operator patterns in the codebase. Previously flagged issues (scalar hue validation in ColorJitter, scalar input crash in the functional normalize, and GaussianBlur bugs) have all been resolved.

Confidence Score: 5/5

Safe to merge — all previously raised P1 issues have been resolved and only minor style findings remain.

All P1 findings from prior review rounds are fixed. The two remaining findings are P2 style issues (a comment typo and a slightly inaccurate docstring) that do not affect correctness or runtime behaviour.

No files require special attention.

Vulnerabilities

No security concerns identified.

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/functional/normalize.py	New functional normalize implementation; layout detection via string slicing is correct; minor comment typos ("Torchvison", "nomrlization").
dali/python/nvidia/dali/experimental/torchvision/v2/normalize.py	New Normalize operator class; validation logic is correct; layout comparison uses established uint8 byte pattern; docstring slightly inaccurate about PIL Image rejection.
dali/test/python/torchvision/test_tv_normalize.py	Good test coverage across CPU/GPU, single/multi-channel, batch shapes, mismatched lengths, and zero-std cases.
dali/python/nvidia/dali/experimental/torchvision/v2/color.py	Correctly adds a non-negative guard for scalar hue before range expansion, fixing the previously flagged silent inverted-range bug.
dali/python/nvidia/dali/experimental/torchvision/init.py	Adds Normalize to the public torchvision API surface.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/init.py	Exports normalize from the functional sub-package.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/gaussian_blur.py	Whitespace-only change (blank line added before return).

Sequence Diagram

sequenceDiagram
    participant Caller
    participant normalize as functional.normalize
    participant _normalize as functional._normalize
    participant Normalize as Normalize class
    participant ndd as ndd.normalize
    participant fn as fn.normalize

    Caller->>normalize: normalize(input, mean, std)
    normalize->>Normalize: verify_args(std, mean)
    normalize->>_normalize: _normalize(input, mean, std)
    _normalize->>_normalize: detect layout (CHW vs HWC)
    _normalize->>_normalize: reshape mean/std arrays
    _normalize->>ndd: ndd.normalize(input, mean, stddev)
    ndd-->>Caller: normalized tensor

    Caller->>Normalize: Normalize(mean, std)(input)
    Normalize->>Normalize: verify_args (via Operator.__init__)
    Normalize->>Normalize: _kernel(data_input)
    Normalize->>Normalize: detect layout via property("layout")[-1]
    Normalize->>fn: fn.normalize(data_input, mean, stddev)
    fn-->>Caller: normalized tensor

Loading

_{Reviews (14): Last reviewed commit: "Typo fix" | Re-trigger Greptile}

[mdabek-nvidia]

@greptileai please re-review

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [47913207]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [47913207]: BUILD PASSED

[mdabek-nvidia]

@greptileai please re-review