[Torchvision API] resized_crop and RandomResizedCrop

[mdabek-nvidia]

Category:

New feature

Description:

Torchvision API operators:

• RandomResizedCrop
• Functional resized_crop functional
• Unification of InterpolationMode validation

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 review please

[greptile-apps]

Greptile Summary

This PR adds two new torchvision-API-compatible operators to the DALI experimental torchvision layer: the resized_crop functional (crop + resize in one call) and the RandomResizedCrop transform class. It also unifies PIL integer interpolation-code handling via a shared Resize.normalize_interpolation classmethod, and tightens size validation in _ValidateSize.

• resized_crop (functional): crops via the refactored _crop helper, validates dimensions through RandomCrop.verify_args and Resize.verify_args, then calls ndd.resize with aspect-ratio-aware target sizing via calculate_target_size_dynamic_mode.
• RandomResizedCrop (class): delegates directly to fn.random_resized_crop with new validation classes for scale/ratio ranges and interpolation, keeping the implementation thin.
• Both new operators and the updated test_tv_crop.py now use assert_raises(glob=...) patterns throughout.

Confidence Score: 5/5

Safe to merge; the new operators correctly implement the torchvision contract and delegate to well-tested DALI primitives.

The core crop-then-resize pipeline in resized_crop is straightforward and tested pixel-level against torchvision reference output. RandomResizedCrop is a thin wrapper over fn.random_resized_crop with clean validation. The only finding is a dead helper function that was defined but never wired in, which has no runtime impact.

crop.py — contains the unreachable _verify_crop_coordinate helper that should be removed.

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/functional/crop.py	Adds resized_crop functional: extracts _crop helper, validates via RandomCrop.verify_args + Resize.verify_args, then calls ndd.resize on the cropped result. One unused helper function _verify_crop_coordinate was added but is never called.
dali/python/nvidia/dali/experimental/torchvision/v2/randomcrop.py	Adds RandomResizedCrop operator delegating to fn.random_resized_crop. Validation classes for scale/ratio and interpolation are well-structured and consistent with existing patterns.
dali/python/nvidia/dali/experimental/torchvision/v2/resize.py	Adds normalize_interpolation classmethod for PIL int-code to InterpolationMode conversion, int_to_interpolation_mode map, and additional size positivity/length checks in _ValidateSize. Clean changes used by both new operators.
dali/test/python/torchvision/test_tv_randomresizedcrop.py	New test file covering shape, correctness vs torchvision (identity crop), interpolation modes, antialias, and all validation error paths with assert_raises(glob=...) patterns.
dali/test/python/torchvision/test_tv_resized_crop.py	New test file with comprehensive pixel-level comparison against torchvision for tensor inputs and shape-level comparison for PIL inputs; covers crop positions, padding, interpolation modes, antialias, and all validation paths.
dali/test/python/torchvision/test_tv_crop.py	Updated existing crop tests to use assert_raises(glob=...) patterns for all error-path assertions, satisfying the test-edge-cases rule.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["resized_crop(inpt, top, left, height, width, size, ...)"] --> B["@adjust_input\n(PIL ↔ DALI tensor)"]
    B --> C["_validate_crop_params\n(tensor: _validate_integer_param\nPIL: _round_pil_box)"]
    C --> D["RandomCrop.verify_args\n(size=(height,width))"]
    D --> E["Resize.normalize_interpolation\n(int → InterpolationMode)"]
    E --> F["Resize.verify_args\n(size, interpolation)"]
    F --> G["_crop(inpt, top, left, h, w)\nndd.slice with out_of_bounds pad"]
    G --> H["Resize.calculate_target_size_dynamic_mode\n(height,width) + size_normalized"]
    H --> I["ndd.resize(cropped, size=(target_h,target_w))"]

    J["RandomResizedCrop.__init__(size, scale, ratio, ...)"] --> K["Resize.normalize_interpolation"]
    K --> L["Operator.__init__ → arg_rules:\n_ValidateSizeDescriptor\n_ValidateCropSize\n_ValidateRandomResizedCropScaleRatio\n_ValidateRandomResizedCropInterpolation"]
    L --> M["_kernel → fn.random_resized_crop\n(size, random_area=scale,\nrandom_aspect_ratio=ratio)"]

Loading

_{Reviews (4): Last reviewed commit: "Flake8 fixes - rebase artifacts" | Re-trigger Greptile}

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[mdabek-nvidia]

@greptileai re-review