Torchvision padding

[mdabek-nvidia]

Category:

New feature

Description:

Torchvision pad operator 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

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [46880686]: BUILD STARTED

[greptile-apps]

Greptile Summary

This PR adds a torchvision-compatible Pad operator (class-based and functional) to DALI's experimental torchvision layer, backed by fn.slice / ndd.slice with explicit axis naming to correctly map left/top/right/bottom padding to W and H dimensions. It also includes a rename of two internal validator classes in color.py for naming consistency, and minor test/comment cleanups. Previous review concerns (axis-ordering, non-negative validation, padding-mode validation) have all been addressed.

Confidence Score: 5/5

Safe to merge; the only remaining finding is a P2 type-annotation mismatch that does not affect runtime correctness for documented use cases.

All prior P0/P1 concerns have been resolved. The single new finding is that ndd.Batch appears in pad()'s type hint but raises TypeError at runtime — this is P2 since no public test or documented workflow passes ndd.Batch directly to the functional API. Core padding logic, axis ordering, and validation are correct and cross-validated against torchvision.

dali/python/nvidia/dali/experimental/torchvision/v2/functional/pad.py — minor type-hint cleanup needed.

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/pad.py	New _PadBase, PadConstant/Edge/Reflect/Symmetric, _ValidatePaddingMode, and Pad classes; uses axis_names='WH' to correctly map anchor/shape to W and H dimensions; non-negative padding validation is present.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/pad.py	New functional pad() API; correctly selects axes per input layout; has a type-hint/implementation mismatch where ndd.Batch is advertised but raises TypeError.
dali/test/python/torchvision/test_tv_pads.py	New test file with thorough coverage of all four padding modes, single-axis and multi-axis padding, PIL images, and CPU/GPU variants; cross-validates against torchvision reference outputs.
dali/python/nvidia/dali/experimental/torchvision/v2/color.py	Minor rename of two internal validator classes to use consistent _Validate* naming convention; no logic changes.
dali/test/python/torchvision/test_tv_color.py	Removes a dead img, batch = make_tensor_inputs() assignment that was never used in the function body; correct cleanup.
dali/test/python/torchvision/test_tv_gaussian_blur.py	Single-line comment cleanup removing stray [web:69][web:73] artifact; no functional change.
dali/python/nvidia/dali/experimental/torchvision/init.py	Exports new Pad class and adds it to all; import ordering adjusted alphabetically.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/init.py	Exports new functional pad and adds it to all.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["User input\n(PIL.Image | torch.Tensor)"] --> B{pad or Pad?}
    B -->|"functional.pad()"| C{Input type}
    B -->|"Pad()()"| D["Pad.__init__\n_ValidatePaddingMode\n_ValidateIfNonNegative\nselect PadConstant/Edge/Reflect/Symmetric"]
    C -->|PIL.Image| E["axes=[-3,-2] HWC"]
    C -->|torch.Tensor| F["axes=[-2,-1] CHW"]
    C -->|other| G["raise TypeError"]
    E --> H["_ValidatePaddingMode + _ValidateIfNonNegative"]
    F --> H
    H --> I["_pad() @adjust_input\nndd.slice(anchor=(-top,-left),\nshape=(H+top+bottom,W+left+right),\naxes, out_of_bounds_policy)"]
    D --> J["_PadBase._kernel()\nfn.slice(axis_names='WH'\nanchor=(-pad_left,-pad_top)\nshape=(W+l+r, H+t+b))"]
    I --> K["Output: ndd.Tensor or ndd.Batch"]
    J --> K

Loading

_{Reviews (11): Last reviewed commit: "Review fixes" | Re-trigger Greptile}

[dali-automaton]

CI MESSAGE: [46880686]: BUILD PASSED

[mdabek-nvidia]

@greptileai please re-review

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [47558804]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [47558804]: BUILD PASSED

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [47618654]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [47618654]: BUILD PASSED