[Torchvision API] crop

[mdabek-nvidia]

Category:

New feature

Description:

Torchvision API operators RandomCrop and functional crop.
Unit tests checking conformance with Torchvision.
Known limitations: tensor and PIL.Image based paddings are not supported

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 torchvision-compatible RandomCrop and functional crop operators to DALI's experimental torchvision API. Padding and pad_if_needed are implemented by computing a virtual padded-image space and letting DALI's fn.slice with out_of_bounds_policy handle the boundary fill, avoiding a separate padding pass.

• randomcrop.py: _randint clamps negative max_value to 0 before passing to fn.random.uniform, preventing an inverted-range DALI crash when the crop size exceeds the padded image. _ValidateFill now correctly rejects dict fills (previously listed as supported but not implemented).
• crop.py: Functional crop dispatches PIL rounding vs. integer-only validation based on the tensor layout suffix, mirroring torchvision's behavior, then calls ndd.slice directly.
• Tests: Cover tensor/PIL/batched inputs, multiple padding modes, asymmetric padding with pad_if_needed, and error-rejection paths. Skipped dict-fill tests mark a known limitation acknowledged in the PR description.

Confidence Score: 5/5

Safe to merge; the new operators are well-tested against torchvision references and known limitations are clearly documented.

Both operators are exercised against torchvision reference outputs with deterministic inputs. Previous reviewer concerns (negative max_value in _randint, false dict-fill support, misleading range_start expression, asymmetric padding semantics) are all addressed in this revision. The only remaining item is four TODO comments in the test file that lack tracked issue references.

test_tv_randomcrop.py — four TODO comments for dict-fill support need a tracking issue before merge.

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/randomcrop.py	New RandomCrop operator. _randint correctly clamps max_value to 0 before feeding fn.random.uniform, preventing inverted-range crashes. Symmetric pad_if_needed expansion is correct and tested. Validation classes are clean.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/crop.py	New functional crop operator that delegates size validation to RandomCrop.verify_args and calls ndd.slice with out_of_bounds_policy='pad'. PIL rounding logic mirrors torchvision's behavior. No issues found.
dali/test/python/torchvision/test_tv_randomcrop.py	Comprehensive test suite covering shapes, padding variants, PIL/tensor inputs, and edge cases. Contains four TODO comments for dict-fill support without tracked issue references, violating the no-todo-on-merge rule.
dali/test/python/torchvision/test_tv_crop.py	Tests functional crop against torchvision reference for tensor, PIL, batched-tensor, dtype preservation, and error-rejection cases. Good coverage across CPU/GPU and multiple dtypes.
dali/python/nvidia/dali/experimental/torchvision/init.py	Adds RandomCrop import and all entry. Alphabetically ordered correctly.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/init.py	Adds crop import and all entry. Alphabetically ordered correctly.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User input\nTensor / PIL Image] --> B{adjust_input\ndecorator}
    B --> C[DALI Tensor/Batch\nwith layout]
    C --> D{crop / RandomCrop}
    subgraph crop [functional.crop]
        D1{PIL layout?\nlayout ends HWC} -->|Yes| D2[_round_pil_box\nfloat to int rounding]
        D1 -->|No| D3[_validate_integer_param\nreject floats]
        D2 & D3 --> D4[RandomCrop.verify_args\nsize validation]
        D4 --> D5[ndd.slice\nout_of_bounds_policy=pad]
    end
    subgraph rcrop [RandomCrop._kernel]
        E1[Compute padded_h/w\nfrom explicit padding] --> E2{pad_if_needed?}
        E2 -->|Yes| E3[dali.math.max\nexpand both sides]
        E2 -->|No| E4[max_top/left =\npadded - crop]
        E3 --> E4
        E4 --> E5[_randint\nclamp to 0, fn.random.uniform]
        E5 --> E6[fn.slice\nanchor = offset - pad\nout_of_bounds fill]
    end
    D --> D1
    C --> E1
    E6 --> F[Output Tensor / PIL]
    D5 --> F

Loading

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

[mdabek-nvidia]

@greptileai re-review please

[mdabek-nvidia]

@greptileai, re-review

[mdabek-nvidia]

@greptileai re-review

[mdabek-nvidia]

@greptileai re-review

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [52613374]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [52613374]: BUILD FAILED

[dali-automaton]

CI MESSAGE: [52613374]: BUILD PASSED

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [52656583]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [52656583]: BUILD PASSED