Torchvision padding by mdabek-nvidia · Pull Request #6276 · NVIDIA/DALI

mdabek-nvidia · 2026-03-24T14:44:33Z

Category:

New feature

Description:

Torchvision pad operator implementation

Additional information:

Affected modules and functionalities:

Key points relevant for the review:

Tests:

Checklist

Documentation

DALI team only

Requirements

Implements new requirements
Affects existing requirements
N/A

REQ IDs: N/A

JIRA TASK: N/A

Signed-off-by: Marek Dabek <[email protected]>

Signed-off-by: Marek Dabek <[email protected]> Co-authored-by: Kamil Tokarski <[email protected]>

Signed-off-by: Marek Dabek <[email protected]>

…st iteration. (NVIDIA#6264) Signed-off-by: Michał Zientkiewicz <[email protected]>

- switch form launchpad API to ppa as it sometimes can be blocked Signed-off-by: Janusz Lisiecki <[email protected]>

…hreadPool per device. (NVIDIA#6254) * Use NewThreadPool in dynamic mode. * By default use only one instance of ThreadPool per device. * Add thread_pool argument to EvalContext constructor. Signed-off-by: Michal Zientkiewicz <[email protected]>

…IA#6268) - Replace deprecated trusted=yes with proper GPG keyring for the deadsnakes PPA Signed-off-by: Janusz Lisiecki <[email protected]>

- arg_helper_test.cc: pass local TensorListShape<0> variables via std::move() into SetupData() since they are not used after the call (CIDs 25217386, 25217389, 25217390, 26094597) - arg_helper_test.cc: use `const auto &` instead of `auto` when binding the return value of arg.get() (which returns const TLV&) to avoid unnecessary copies (CIDs 25217387, 25217388, 25217391) - roi_image_decoder.h: pass `shape` parameter via std::move() into RoiFromCropWindowGenerator() in all three GetRoi() overrides (WithCropAttr, WithSliceAttr, WithRandomCropAttr), since shape is not used after the call (CIDs 25217545, 25217546, 25217547) - thread_pool_base.h: add explicit noexcept to ~ThreadPoolBase() and ~TaskBulkAdd() to document intent that exceptions cause termination (CIDs 25783968, 25783975) - workspace_policy.h: add explicit noexcept to ~AOT_WS_Policy() to document intent that exceptions cause termination (CID 26095288) - thread_pool_base.cc: fix lock order reversal in WaitOrRunTasks() and Run() - release mtx_ before acquiring sem_ to maintain consistent order (sem_ before mtx_) matching all other acquisition sites, eliminating potential for ORDER_REVERSAL deadlock (CID 25783972) - new_thread_pool.cc: pass `name` parameter via std::move() into name_ member initializer since it is not used after the call (CID 26190478) Signed-off-by: Janusz Lisiecki <[email protected]> Signed-off-by: Michał Zientkiewicz <[email protected]> Co-authored-by: Michał Zientkiewicz <[email protected]>

…NVIDIA#6270) * Add quiet argument to RandomBBoxCrop to suppress crop failure warning When using mosaic augmentation or other pipelines where failing to find a valid crop window is expected, the repeated warning at bbox_crop.cc:794 is noisy and misleading. The new quiet=True argument suppresses it. Fixes NVIDIA#5695. * Add test for quiet=True suppressing crop failure warning

The fallback to the legacy CropMirrorNormalize implementation is the correct and intended path for video (4D) tensors and other non-2D-image inputs, since the optimized kernels only support 3D HWC/CHW tensors. The warning incorrectly implied something was wrong. Fixes NVIDIA#6267.

Center crop operators implementation Signed-off-by: Marek Dabek <[email protected]> Co-authored-by: Kamil Tokarski <[email protected]>

Signed-off-by: Marek Dabek <[email protected]>

mdabek-nvidia · 2026-03-24T14:47:26Z

@greptileai please review

mdabek-nvidia · 2026-03-24T14:48:21Z

!build

dali-automaton · 2026-03-24T14:50:34Z

CI MESSAGE: [46880686]: BUILD STARTED

greptile-apps · 2026-03-24T15:00:15Z

Greptile Summary

This PR adds Pad, GaussianBlur, ColorJitter, and Grayscale torchvision-compatible operators to DALI's experimental torchvision module, along with their functional counterparts. It also refactors shared layout-detection helpers (get_HWC_from_layout_pipeline, get_HWC_from_layout_dynamic) into operator.py and removes duplication in resize.py and centercrop.py.

Pad (_PadBase hierarchy) is the primary new feature. It correctly uses axis_names="WH" with (W, H)-ordered anchor/shape arguments in the pipeline mode, and explicit axes in the dynamic functional mode — the axis-ordering concern from prior review rounds is resolved.
Grayscale and ColorJitter look correct; the VerificationHue fix from prior rounds (uniform scalar→(-float(hue), float(hue)) conversion) is in place.
GaussianBlur — VerifyKernel.verify still contains isinstance(kernel_size, int, Sequence) with three positional arguments at line 37. Python's isinstance() only accepts two arguments, so this raises TypeError for every call, making GaussianBlur construction fail for all valid kernel sizes. A prior review thread flagged this exact issue and received a "Fixed" reply, but the fix does not appear in the current code. This renders the GaussianBlur class non-functional and will cause several tests in test_tv_gaussian_blur.py to fail.
One minor cleanup: error messages in VerifyKernel contain the typo sequenece (should be sequence).

Confidence Score: 3/5

Not ready to merge — GaussianBlur is broken due to a TypeError in VerifyKernel that was declared fixed but remains in the code.
The core Pad implementation and the Grayscale/ColorJitter operators look correct, and prior axis-ordering and hue-validation concerns are resolved. However, VerifyKernel.verify still has the isinstance(kernel_size, int, Sequence) three-argument call that raises TypeError for every valid input, rendering GaussianBlur completely non-functional. This breaks the primary user path for that operator and will cause test failures in test_tv_gaussian_blur.py. A single targeted fix is needed before the GaussianBlur path is safe to merge.
dali/python/nvidia/dali/experimental/torchvision/v2/gaussian_blur.py — VerifyKernel.verify (line 37) requires attention.

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/pad.py	New padding operator implementing constant, edge, reflect, and symmetric modes via DALI's fn.slice. The _PadBase._kernel correctly uses axis_names="WH" with (W, H)-ordered anchor/shape arguments, resolving the axis-ordering concern from prior review. Validation (non-negative padding, padding mode) looks correct.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/pad.py	Functional pad implementation correctly derives axes from the input type before calling the @adjust_input-decorated _pad helper. For HWC (PIL Image) axes=[-3, -2] and for CHW (Tensor) axes=[-2, -1] are passed to ndd.slice, making the anchor (-top, -left) in (H, W) order consistent with those axes.
dali/python/nvidia/dali/experimental/torchvision/v2/gaussian_blur.py	VerifyKernel.verify still contains isinstance(kernel_size, int, Sequence) with three positional arguments (line 37) — Python's isinstance() only accepts two, so this raises TypeError for every input, making GaussianBlur construction fail for all valid kernel sizes. The prior review thread flagged this and marked it "Fixed", but the fix does not appear in the current code.
dali/python/nvidia/dali/experimental/torchvision/v2/color.py	ColorJitter and Grayscale operators newly added. VerificationHue now consistently converts scalar hue to (-float(hue), float(hue)) for both int and float inputs, addressing the prior review concern. VerificationBCS and Grayscale logic look correct.
dali/test/python/torchvision/test_tv_pads.py	Comprehensive tests covering all four padding modes, single-side and multi-side padding, constant fill value, PIL images, and both CPU/GPU variants. Correctness is verified by exact equality against torchvision reference outputs.
dali/test/python/torchvision/test_tv_gaussian_blur.py	Tests for GaussianBlur will be affected by the broken VerifyKernel.verify (TypeError from isinstance with 3 args). In particular, the invalid-params test for ((3,3),-1.0) expects ValueError but would get TypeError, and valid-input tests for GaussianBlur construction would also fail with TypeError.

Class Diagram

%%{init: {'theme': 'neutral'}}%%
classDiagram
    class Operator {
        +arg_rules: list
        +preprocess_data
        +device: str
        +__call__(data_input)
        +_kernel(data_input)*
        +verify_args(**kwargs)
    }

    class _PadBase {
        +border_type: str
        +pad_left: int
        +pad_top: int
        +pad_right: int
        +pad_bottom: int
        +fill
        +get_padding(padding)$
        +_kernel(data_input)
    }

    class PadConstant {
        +border_type = "pad"
    }
    class PadEdge {
        +border_type = "clamp"
    }
    class PadReflect {
        +border_type = "reflect_101"
    }
    class PadSymmetric {
        +border_type = "reflect_1001"
    }

    class Pad {
        +pad: _PadBase
        +__init__(padding, fill, padding_mode, device)
        +__call__(data_input)
    }

    class GaussianBlur {
        +kernel_size
        +sigma
        +_kernel(data_input)
    }

    class ColorJitter {
        +brightness
        +contrast
        +saturation
        +hue
        +_kernel(data_input)
    }

    class Grayscale {
        +num_output_channels: int
        +_kernel(data_input)
    }

    Operator <|-- _PadBase
    _PadBase <|-- PadConstant
    _PadBase <|-- PadEdge
    _PadBase <|-- PadReflect
    _PadBase <|-- PadSymmetric
    Pad o-- _PadBase : delegates to
    Operator <|-- GaussianBlur
    Operator <|-- ColorJitter
    Operator <|-- Grayscale

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

dali/python/nvidia/dali/experimental/torchvision/v2/pad.py

dali/python/nvidia/dali/experimental/torchvision/v2/color.py

dali/python/nvidia/dali/experimental/torchvision/v2/pad.py

dali-automaton · 2026-03-24T19:55:23Z

CI MESSAGE: [46880686]: BUILD PASSED

Signed-off-by: Marek Dabek <[email protected]>

mdabek-nvidia · 2026-03-25T09:07:44Z

@greptileai please re-review

dali/python/nvidia/dali/experimental/torchvision/v2/gaussian_blur.py

Signed-off-by: Marek Dabek <[email protected]>

mdabek-nvidia and others added 21 commits March 19, 2026 13:30

Center crop operator

e09d04b

Signed-off-by: Marek Dabek <[email protected]>

Review fixes

1b043c0

Signed-off-by: Marek Dabek <[email protected]>

Review fixes

f30e282

Signed-off-by: Marek Dabek <[email protected]>

Apply suggestion from @stiepan

3337aa9

Signed-off-by: Marek Dabek <[email protected]> Co-authored-by: Kamil Tokarski <[email protected]>

Torchvision ColorJitter and Grayscale implementations

ebbf863

Signed-off-by: Marek Dabek <[email protected]>

Review fixes

d8c5cd4

Signed-off-by: Marek Dabek <[email protected]>

Review fixes

4dc37b5

Signed-off-by: Marek Dabek <[email protected]>

Review fixes

e146f9f

Signed-off-by: Marek Dabek <[email protected]>

Review fixes

6a0038d

Signed-off-by: Marek Dabek <[email protected]>

SequenceOperator: Do not keep thread pool and output order from the 1…

26410ba

…st iteration. (NVIDIA#6264) Signed-off-by: Michał Zientkiewicz <[email protected]>

Change a way deadsnakes ppa is accessed (NVIDIA#6263)

820848a

- switch form launchpad API to ppa as it sometimes can be blocked Signed-off-by: Janusz Lisiecki <[email protected]>

Improve deadsnakes PPA key handling in aarch64-linux Dockerfile (NVID…

66af962

…IA#6268) - Replace deprecated trusted=yes with proper GPG keyring for the deadsnakes PPA Signed-off-by: Janusz Lisiecki <[email protected]>

Torchvision API - center crop operator (NVIDIA#6266)

cfff99d

Center crop operators implementation Signed-off-by: Marek Dabek <[email protected]> Co-authored-by: Kamil Tokarski <[email protected]>

Torchvision ColorJitter and Grayscale implementations

df4d028

Signed-off-by: Marek Dabek <[email protected]>

Review fixes

caf8b4b

Signed-off-by: Marek Dabek <[email protected]>

Gaussian blur operator

6ddc68a

Signed-off-by: Marek Dabek <[email protected]>

Review fixes

ef0bf89

Signed-off-by: Marek Dabek <[email protected]>

mdabek-nvidia force-pushed the torchvision_pad branch from 71f53ca to 0c38e81 Compare March 24, 2026 14:46

greptile-apps bot reviewed Mar 24, 2026

View reviewed changes

dali/python/nvidia/dali/experimental/torchvision/v2/pad.py Show resolved Hide resolved

dali/python/nvidia/dali/experimental/torchvision/v2/color.py Show resolved Hide resolved

dali/python/nvidia/dali/experimental/torchvision/v2/pad.py Show resolved Hide resolved

mdabek-nvidia added 2 commits March 25, 2026 09:45

Review fixes

455f51e

Signed-off-by: Marek Dabek <[email protected]>

Torchvision ColorJitter and Grayscale implementations

f8fdd8b

Signed-off-by: Marek Dabek <[email protected]>

mdabek-nvidia added 3 commits March 25, 2026 09:49

Review fixes

a836fe9

Signed-off-by: Marek Dabek <[email protected]>

Gaussian blur operator

7db19e7

Signed-off-by: Marek Dabek <[email protected]>

Torchvision Pad operators

5c29882

Signed-off-by: Marek Dabek <[email protected]>

mdabek-nvidia force-pushed the torchvision_pad branch from 0c38e81 to d6f47fc Compare March 25, 2026 09:05

mdabek-nvidia marked this pull request as ready for review March 25, 2026 13:10

greptile-apps bot reviewed Mar 25, 2026

View reviewed changes

dali/python/nvidia/dali/experimental/torchvision/v2/gaussian_blur.py Show resolved Hide resolved

mdabek-nvidia force-pushed the torchvision_pad branch from d6f47fc to 9f1e43b Compare March 25, 2026 13:45

Review fixes

d9dee09

Signed-off-by: Marek Dabek <[email protected]>

mdabek-nvidia force-pushed the torchvision_pad branch from 9f1e43b to d9dee09 Compare March 25, 2026 13:47

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Torchvision padding#6276

Torchvision padding#6276
mdabek-nvidia wants to merge 27 commits intoNVIDIA:mainfrom
mdabek-nvidia:torchvision_pad

mdabek-nvidia commented Mar 24, 2026

Uh oh!

mdabek-nvidia commented Mar 24, 2026

Uh oh!

mdabek-nvidia commented Mar 24, 2026

Uh oh!

dali-automaton commented Mar 24, 2026

Uh oh!

greptile-apps bot commented Mar 24, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dali-automaton commented Mar 24, 2026

Uh oh!

mdabek-nvidia commented Mar 25, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

mdabek-nvidia commented Mar 24, 2026

Category:

Description:

Additional information:

Affected modules and functionalities:

Key points relevant for the review:

Tests:

Checklist

Documentation

DALI team only

Requirements

Uh oh!

mdabek-nvidia commented Mar 24, 2026

Uh oh!

mdabek-nvidia commented Mar 24, 2026

Uh oh!

dali-automaton commented Mar 24, 2026

Uh oh!

greptile-apps bot commented Mar 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 3/5

Important Files Changed

Class Diagram

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dali-automaton commented Mar 24, 2026

Uh oh!

mdabek-nvidia commented Mar 25, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

greptile-apps bot commented Mar 24, 2026 •

edited

Loading