Torchvision API RandomApply implementation

[mdabek-nvidia]

Category:

New feature

Description:

Torchvision API implementation of RandomApply operator

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 RandomApply operator to DALI's experimental torchvision-compatible API, allowing a list of transforms to be applied with a configurable probability p. The implementation follows the established RandomGrayscale pattern using fn.random.coin_flip with if convert: inside enable_conditionals=True pipelines.

• rand_apply.py: New RandomApply class that validates p via _ValidateIfZeroOneRange, stores the op list as _op_list, and uses a DALI coin-flip node as a pipeline-graph conditional to selectively apply the ops.
• test_tv_rand_apply.py: Tests covering p=0, p=1, multi-op, shape preservation, invalid probability, and a sanity probability-distribution test.
• __init__.py: Wires RandomApply into the module's public __all__.

Confidence Score: 5/5

Safe to merge; the implementation correctly follows the established RandomGrayscale pattern for DALI conditional execution.

The RandomApply operator faithfully mirrors the existing RandomGrayscale conditional pattern (fn.random.coin_flip + if convert: inside enable_conditionals=True) which is already working in production. The p=0 and p=1 boundary tests provide strong correctness coverage. The only remaining gap is missing deterministic seeds in the two probabilistic sanity tests, which does not affect the operator's runtime correctness.

test_tv_rand_apply.py: probabilistic tests should set a seed for reproducibility.

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/rand_apply.py	New RandomApply operator; follows the established RandomGrayscale conditional pattern correctly; p validation and _op_list storage are correct.
dali/test/python/torchvision/test_tv_rand_apply.py	Tests for p=0/1/multi-op are solid; probabilistic tests (test_random_apply_p, test_random_apply_p_sanity) run without a deterministic seed, violating test-set-seed.
dali/python/nvidia/dali/experimental/torchvision/init.py	Adds RandomApply import and all entry; straightforward.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["Compose.__call__(img)"] --> B["PipelineHWC / PipelineCHW.run()"]
    B --> C["_pipeline_function (enable_conditionals=True)"]
    C --> D["fn.external_source → DataNode"]
    D --> E["RandomApply.__call__(DataNode)"]
    E --> F["Operator.__call__: verify_data, device placement"]
    F --> G["RandomApply._kernel(DataNode)"]
    G --> H["fn.random.coin_flip(probability=p) → convert:DataNode"]
    H --> I{"if convert:\n(DALI conditional branch)"}
    I -- "True (at runtime)" --> J["for op in _op_list:\n  output = op(output)"]
    J --> K["Nested op._kernel(DataNode) × N"]
    K --> M["return output (transformed)"]
    I -- "False (at runtime)" --> N["return output (unchanged = data_input)"]
    M --> O["Pipeline output → PIL / Tensor"]
    N --> O

Loading

_{Reviews (7): Last reviewed commit: "Review fixes" | Re-trigger Greptile}

[mdabek-nvidia]

@greptileai please re-review

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [50671844]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [50671844]: BUILD FAILED

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [50675590]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [50675590]: BUILD PASSED

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [51158133]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [51158133]: BUILD PASSED