[Torchvision API] Remove Torchvision's dependency - InterpolationMode

[mdabek-nvidia]

Category:

Refactoring

Description:

Remove currently the only Torchvision's dependency - InterpolationMode to make DALI Torchvision standalone solution:

• Introduce internal _enums.py
• Remove InterpolationMode dependencies from crops, resize and tests

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: DALI-4713

[mdabek-nvidia]

@greptileai review

[greptile-apps]

Greptile Summary

This PR removes DALI's only remaining torchvision import by introducing a DALI-native InterpolationMode enum (matching torchvision's name→value mapping) and a _normalize_enum_like_interpolation_mode helper that coerces foreign enum instances—including torchvision's—without importing torchvision at all.

• v2/_enums.py (new): defines InterpolationMode with the same seven string values as torchvision's enum, plus the private _normalize_enum_like_interpolation_mode that tries name-based then value-based lookup against the local enum, raising a descriptive ValueError for anything unrecognised.
• resize.py / randomcrop.py / functional/*.py: swap the from torchvision.transforms import InterpolationMode import for the local one; validate_interpolation (typo fix from validate_interpoliation) now calls normalize_interpolation first so torchvision enum instances pass through transparently.
• Tests: test_tv_interpolation_mode.py (new) verifies enum parity with torchvision, DALIInterpType mapping, normalization of torchvision-like enums, and error paths; test_tv_randomresizedcrop.py migrates most call sites to use the DALI-native enum.

Confidence Score: 5/5

Safe to merge; all changes are import swaps, a typo fix, and a new normalisation helper with no effect on existing operator behaviour.

The refactoring is mechanical: every torchvision InterpolationMode import is replaced by the DALI-native enum, and the normalisation helper is idempotent and well-tested. The typo rename (validate_interpoliation → validate_interpolation) is verified to have no remaining callers under the old name. No data-path logic changed.

No files require special attention.

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/_enums.py	New file: defines DALI-native InterpolationMode enum (same values as torchvision's) and _normalize_enum_like_interpolation_mode helper to coerce foreign enum instances without importing torchvision.
dali/python/nvidia/dali/experimental/torchvision/v2/resize.py	Swaps torchvision import for local _enums; renames validate_interpoliation→validate_interpolation (typo fix); validate_interpolation now normalizes before checking, enabling torchvision enum passthrough.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/resize.py	Swaps torchvision import; adds explicit normalize_interpolation call before verify_args so the local variable is an InterpolationMode member for the subsequent interpolation_modes dict lookup.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/crop.py	Import-only change: torchvision.InterpolationMode replaced by local _enums.InterpolationMode; resized_crop body unchanged.
dali/python/nvidia/dali/experimental/torchvision/v2/randomcrop.py	Import swapped to local enum; validate_interpoliation typo corrected to validate_interpolation in _ValidateRandomResizedCropInterpolation.verify.
dali/python/nvidia/dali/experimental/torchvision/init.py	Exports InterpolationMode at the package level via init.py and all, making it the public entry point for users who previously depended on torchvision's enum.
dali/test/python/torchvision/test_tv_interpolation_mode.py	New test file: covers enum value parity with torchvision, DALIInterpType mapping, normalization of torchvision-like enums, and error paths; uses private torchvision internal (_check_interpolation) for exact error-message comparison.
dali/test/python/torchvision/test_tv_randomresizedcrop.py	Migrates most InterpolationMode references from transforms.InterpolationMode to DALI's InterpolationMode; adds import of DALI's InterpolationMode.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["User passes interpolation\n(int, DALI InterpolationMode,\nor torchvision-like enum)"] --> B{isinstance int?}
    B -- Yes --> C["int_to_interpolation_mode lookup\n(PIL codes 0-5)"]
    C -- KeyError --> D["ValueError: not a valid PIL code"]
    C -- Found --> E["InterpolationMode member"]
    B -- No --> F["_normalize_enum_like_interpolation_mode"]
    F --> G{Already InterpolationMode?}
    G -- Yes --> E
    G -- No --> H["Try InterpolationMode[obj.name]"]
    H -- KeyError --> I["Try InterpolationMode(obj.value)"]
    I -- ValueError --> J["ValueError: unrecognised input"]
    H -- Found --> E
    I -- Found --> E
    E --> K["validate_interpolation\n(check supported modes)"]
    K --> L["interpolation_modes dict\n→ DALIInterpType"]
    L --> M["fn.resize / ndd.resize"]

Loading

_{Reviews (8): Last reviewed commit: "Imoroved InterpolationMode error handlin..." | Re-trigger Greptile}

[greptile-apps]

Greptile Summary

This refactoring removes the sole Torchvision import dependency from DALI's experimental Torchvision API layer by introducing a DALI-native InterpolationMode enum in _enums.py and a compatibility bridge (normalize_enum_like_interpolation_mode) that converts torchvision-style enum objects to the local enum without an explicit import.

• _enums.py mirrors every member of torchvision.transforms.InterpolationMode with the same string values, enabling duck-type lookup by .name or .value so existing torchvision enum objects still work transparently.
• Resize.normalize_interpolation is updated to call the bridge at the end (instead of a bare return), and validate_interpoliation now normalizes internally before its membership checks; the functional entry-points (functional/resize.py, functional/crop.py) and RandomResizedCrop pick up the new DALI enum via updated imports.
• Tests update two assertion lines to use the DALI-native enum but continue to pass torchvision enums as inputs, providing backward-compatibility coverage.

Confidence Score: 4/5

Safe to merge; the compatibility bridge is correct and all callers normalize before using the result as a dict key.

The enum values in _enums.py mirror torchvision's exactly, name- and value-based duck-type lookups are tested implicitly by existing tests, and all code paths that use the normalized result as an interpolation_modes dict key are guarded by validate_interpoliation raising before the lookup. The remaining comments are cosmetic (helper naming, typo, missing smoke-test for the DALI-native path).

_enums.py (internal helper naming) and the test file (no end-to-end test using the new DALI-native enum as input).

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/_enums.py	New file introducing DALI-native InterpolationMode enum with a compatibility bridge function; bridge has a silent passthrough for unrecognized types (caught downstream by validation).
dali/python/nvidia/dali/experimental/torchvision/v2/resize.py	Switches import from torchvision; adds normalize_interpolation call inside validate_interpoliation and updates normalize_interpolation to call the bridge. Double normalization is redundant but harmless due to idempotency.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/resize.py	Adds normalize_interpolation call before verify_args; normalize is also called internally inside validate_interpoliation — double call is safe due to idempotency.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/crop.py	Import-only change; resized_crop already had normalize_interpolation call before this PR.
dali/python/nvidia/dali/experimental/torchvision/v2/randomcrop.py	Import-only change; no functional behavior differences.
dali/python/nvidia/dali/experimental/torchvision/init.py	Exports the new InterpolationMode enum from the package's public API surface.
dali/test/python/torchvision/test_tv_randomresizedcrop.py	Updates two assertions to use DALI's InterpolationMode; most test cases still pass torchvision enums as inputs (intentional backward-compatibility check), but no end-to-end test exercises the DALI-native enum as the primary input.

Sequence Diagram

sequenceDiagram
    participant User
    participant Resize/RRC as Resize / RandomResizedCrop
    participant normalize_interpolation as normalize_interpolation()
    participant normalize_bridge as normalize_enum_like_interpolation_mode()
    participant validate as validate_interpoliation()
    participant DALI as fn.resize / fn.random_resized_crop

    User->>Resize/RRC: interpolation (TV enum, DALI enum, or int)
    Resize/RRC->>normalize_interpolation: interpolation
    alt int input
        normalize_interpolation-->>Resize/RRC: InterpolationMode via int_to_interpolation_mode
    else non-int input
        normalize_interpolation->>normalize_bridge: interpolation
        alt already InterpolationMode
            normalize_bridge-->>normalize_interpolation: pass-through
        else has matching .name
            normalize_bridge-->>normalize_interpolation: InterpolationMode[name]
        else has matching .value
            normalize_bridge-->>normalize_interpolation: InterpolationMode(value)
        else unrecognized
            normalize_bridge-->>normalize_interpolation: original object (caught downstream)
        end
        normalize_interpolation-->>Resize/RRC: InterpolationMode enum
    end
    Resize/RRC->>validate: normalized interpolation
    validate->>normalize_interpolation: re-normalizes internally
    validate-->>Resize/RRC: raises NotImplementedError or ValueError if unsupported
    Resize/RRC->>DALI: interpolation_modes[InterpolationMode] → DALIInterpType

Loading

_{Reviews (1): Last reviewed commit: "Removed Torchvision dependency - Interpo..." | Re-trigger Greptile}

[mdabek-nvidia]

!build

[dali-automaton]

CI MESSAGE: [53516601]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [53516601]: BUILD PASSED