Torchvision custom operator documentation

[mdabek-nvidia]

Category:

Other (Documentation)

Description:

Adding Implementing custom operator chapter to Torchvision documentation

Additional information:

Affected modules and functionalities:

Key points relevant for the review:

Check if new chapter could be clearly understood and if user is able to implement a custom operator based on this information.

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: DALI-4809

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[greptile-apps]

Confidence Score: 3/5

The PR is not ready to merge because both custom-operator examples fail when followed or executed, and an explicit documentation typo rule remains violated.

Findings

1.  Custom operator cannot run ▶
2.  Functional example returns None ▶
3.  Guidance relies on private APIs ▶
4.  Correct documentation typo ▶

Summary

This PR adds an object-oriented TrimBorder walkthrough, an abbreviated explanation of the Torchvision Operator base class, and a functional Dynamic Mode example. However, neither example is currently executable from the notebook, and the guidance exposes private implementation classes as extension points.

• Adds documentation for implementing custom Torchvision-compatible transformations.
• Demonstrates preprocessing, slicing, validation rules, and Dynamic Mode input adjustment.
• Requires corrections to imports/helpers and a real functional implementation before the chapter can serve as a working recipe.

_{Reviews (1) · Last reviewed commit: "Black fixes"}

[greptile-apps]

Custom operator cannot run

Running the notebook reaches class TrimBorder(Operator) without defining or importing Operator, so the example immediately raises NameError. The later code also uses unimported get_HWC_from_layout_pipeline, dali, and fn, and calls _stack_int64, which has no definition in the notebook or repository. Add the required imports and use a defined helper such as the fn.stack pattern used by existing operators so readers can run the documented implementation.

[greptile-apps]

Functional example returns None

This example uses names that the notebook does not import, and its ... body returns None. Once the imports are resolved, adjust_input unconditionally calls .evaluate() on that result, causing an AttributeError instead of applying a crop. Show and return the actual Dynamic Mode operation, as the existing crop implementation does with ndd.slice, so readers receive a usable example.

[greptile-apps]

Guidance relies on private APIs

The chapter tells users to derive validators from _ArgumentValidateRule, but that underscore-prefixed class, _DataValidateRule, and Operator are not exported by the public Torchvision package. This makes custom operators depend on private implementation details that may change without compatibility guarantees. Either expose a supported extension API or avoid presenting these private classes as user-facing extension points.

[greptile-apps]

Correct documentation typo

The new comment spells “transferred” as “transfered.” The repository's fix-typos directive requires typos in documentation and comments to be corrected before merging.

Suggested change

	" # Original input is transfered to GPU, before being preprocess_data.\n",
	" # Original input is transferred to GPU, before being preprocess_data.\n",

Rule Used: Fix typos in identifiers, comments, error messages, and docs. Common ones spotted in DALI history: sequenece, Allready, Whtether, milliseoncds, on-pair, Succesful, poitner, funciton, syncrhonized, deleteion, retruned, `worskpace... (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[jantonguirao]

The P1 items flagged by greptile should be fixed before merging:

• Missing imports / undefined _stack_int64 in the TrimBorder example (line 936)
• The crop functional example returns None/Ellipsis, which breaks adjust_input (line 1076)