Torchvision gaussian blur by mdabek-nvidia · Pull Request #6275 · NVIDIA/DALI

mdabek-nvidia · 2026-03-24T14:20:24Z

Category:

New feature

Description:

Implementation of Torchvision's GaussianBlur operator

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 <mdabek@nvidia.com>

Signed-off-by: Marek Dabek <mdabek@nvidia.com> Co-authored-by: Kamil Tokarski <kamiltokarski04@gmail.com>

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Signed-of-by: Marek Dabek <mdabek@nvidia.com>

mdabek-nvidia · 2026-03-24T14:42:19Z

@greptileai please review

mdabek-nvidia · 2026-03-24T14:42:33Z

!build

dali-automaton · 2026-03-24T14:45:49Z

CI MESSAGE: [46880386]: BUILD STARTED

greptile-apps · 2026-03-24T14:53:54Z

Greptile Summary

This PR adds a GaussianBlur operator (pipeline mode) and a gaussian_blur functional (dynamic mode) that mirror torchvision's v2.GaussianBlur API, including correct odd-kernel validation, uniform sigma sampling for range inputs, and DALI-typed input/output handling. It also cleans up type hints across the existing functional wrappers (centercrop, color, flips, resize) to use TensorLike/DeviceLike instead of torch.Tensor | PIL.Image. All previously flagged issues (inverted int-check, broken imports, tuple-sigma misinterpretation, missing odd-kernel guard) have been resolved.

Confidence Score: 5/5

Safe to merge; all prior blocking issues are resolved and remaining findings are minor test housekeeping.

All P0/P1 issues from previous review rounds have been addressed. The only remaining findings are P2 test-file style concerns (wrong comment shape, stale citation tokens, commented-out dead code). These do not affect runtime correctness or the operator contract.

dali/test/python/torchvision/test_tv_gaussian_blur.py — minor comment and dead-code cleanup needed.

Important Files Changed

Filename	Overview
dali/python/nvidia/dali/experimental/torchvision/v2/gaussian_blur.py	New GaussianBlur operator with correct validation, odd-kernel check, and uniform sigma sampling via fn.random.uniform for range inputs; all previously flagged issues resolved.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/gaussian_blur.py	Functional dynamic-mode wrapper; correctly calls verify_args, promotes scalar sigma to a tuple, uses ndd.random.uniform for range sampling, and extracts scalar for the degenerate equal-bounds case.
dali/test/python/torchvision/test_tv_gaussian_blur.py	Comprehensive test coverage for PIL, tensor, and GPU paths; minor issues: wrong comment shape, stale citation artifacts, and commented-out test/assertions that should be restored or tracked.
dali/python/nvidia/dali/experimental/torchvision/init.py	Registers GaussianBlur in the public all and module import; straightforward and correct.
dali/python/nvidia/dali/experimental/torchvision/v2/functional/init.py	Adds gaussian_blur import and all entry; clean and correct.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    Input["Input (PIL / torch.Tensor)"]
    AdjustInput["adjust_input decorator"]
    VerifyArgs["GaussianBlur.verify_args"]
    subgraph PipelineMode["Pipeline Mode"]
        SigmaRange1{"sigma range?"}
        RandomUniform1["fn.random.uniform(range=sigma)"]
        ScalarSigma1["sigma = sigma[0]"]
        FnGaussianBlur["fn.gaussian_blur"]
    end
    subgraph DynamicMode["Dynamic Mode"]
        SigmaRange2{"sigma range?"}
        RandomUniform2["ndd.random.uniform(range=sigma)"]
        ScalarSigma2["sigma = sigma[0]"]
        NddGaussianBlur["ndd.gaussian_blur"]
    end
    Output["Output (PIL / torch.Tensor)"]
    Input --> AdjustInput --> VerifyArgs
    VerifyArgs --> PipelineMode
    VerifyArgs --> DynamicMode
    SigmaRange1 -- yes --> RandomUniform1 --> FnGaussianBlur
    SigmaRange1 -- no --> ScalarSigma1 --> FnGaussianBlur
    SigmaRange2 -- yes --> RandomUniform2 --> NddGaussianBlur
    SigmaRange2 -- no --> ScalarSigma2 --> NddGaussianBlur
    FnGaussianBlur --> Output
    NddGaussianBlur --> Output

_{Reviews (12): Last reviewed commit: "Merge branch 'main' into torchvision_gau..." | Re-trigger Greptile}

dali-automaton · 2026-03-24T19:32:08Z

CI MESSAGE: [46880386]: BUILD PASSED

mdabek-nvidia · 2026-03-25T08:47:03Z

@greptileai please review

mdabek-nvidia · 2026-03-25T08:52:54Z

!build

dali-automaton · 2026-03-25T08:55:27Z

CI MESSAGE: [46954308]: BUILD STARTED

dali-automaton · 2026-03-25T12:54:34Z

CI MESSAGE: [46954308]: BUILD PASSED

dali-automaton · 2026-03-31T09:51:11Z

CI MESSAGE: [47327945]: BUILD PASSED

dali-automaton · 2026-03-31T10:31:39Z

CI MESSAGE: [47329835]: BUILD PASSED

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Center crop operators implementation Signed-off-by: Marek Dabek <mdabek@nvidia.com> Co-authored-by: Kamil Tokarski <kamiltokarski04@gmail.com>

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

banasraf · 2026-04-02T08:24:59Z

+
+        if not isinstance(kernel_size, (int, Sequence)):
+            raise ValueError(
+                f" Kernel size should be an int or a sequenece of ints, got {kernel_size}"


Suggested change

f" Kernel size should be an int or a sequenece of ints, got {kernel_size}"

f"Kernel size should be an int or a sequence of ints, got {kernel_size}"

banasraf · 2026-04-02T08:25:12Z

+
+        if isinstance(kernel_size, Sequence) and any(not isinstance(v, int) for v in kernel_size):
+            raise ValueError(
+                f" Kernel size should be an int or a sequenece of ints, got {kernel_size}"


Suggested change

f" Kernel size should be an int or a sequenece of ints, got {kernel_size}"

f"Kernel size should be an int or a sequence of ints, got {kernel_size}"

banasraf · 2026-04-02T08:29:55Z

+    kernel_size: int | Sequence[int],
+    sigma: int | float | Sequence[float] = (0.1, 2.0),
+    device: Literal["cpu", "gpu"] = "cpu",
+) -> Image.Image | torch.Tensor:


should these types annotations match the actual function body (ndd typesor what happens after annotating withadjust_input`?

Answering both comments, the one I am responding to and to @rostan-t 's one .
The logic for annotations (here and in every other functional API call) is that it is additional information for users to what is being supported by the API, especially that every input is converted to ndd.Tensor or ndd.Batch by the adjust_input operator.

banasraf · 2026-04-02T08:30:43Z

+        dali_blur = Compose([GaussianBlur(kernel_size=kernel_size, sigma=sigma)])
+        _ = dali_blur(tens)
+
+    """ Both do not raise ValueError


why do we have this commented out?

The expectation was that invalid kernel/sigma will cause a ValueError exception, but it did not even in the original code. I am keeping it as a documentation of an oddity.

…VIDIA#6262) * Improve call stack depth handling for error tracebacks in dynamic mode * Rework exception_tester to automatically infer expected line and message --------- Signed-off-by: Rostan Tabet <rtabet@nvidia.com>

Add uniform_sample option to VideoReaderDecoder Adds a `uniform_sample` argument that samples exactly `sequence_length` frames uniformly across the full video (or ROI) using linspace with std::round rounding. Each video produces one sample per epoch; `stride` and `step` are ignored (with a DALI_WARN if `pad_mode` is also set). `sequence_length` must be >= 1 when `uniform_sample=True`. To avoid a per-sample heap allocation, VideoSampleDesc::frame_idxs_ changes from std::vector<int> to span<const int>. Owned storage moves to all_frame_idxs_ (vector<vector<int>>) in the loader, which is populated once in PrepareMetadataImpl and stable for the loader's lifetime. The copy in ReadSample becomes a cheap 16-byte span copy. all_frame_idxs_ is declared before samples_ so its lifetime encloses the spans, preventing dangling references on destruction. New tests: - test_uniform_sample_basic: verifies frame indices match expected linspace - test_uniform_sample_file_list_roi: ROI with non-zero start_frame via file_list - test_uniform_sample_sequence_length_zero_raises: sequence_length=0 raises - test_uniform_sample_stride_step_ignored: stride/step have no effect Signed-off-by: Joaquin Anton Guirao <janton@nvidia.com>

- cudaDeviceSynchronize returns cudaErrorStreamCaptureUnsupported when any stream on the device is being captured. DLTensorGraveyard now puts pending deletions back in the queue and retries after a short wait instead of propagating the error. Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>

* Torchvision ColorJitter and Grayscale implementations Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

mdabek-nvidia · 2026-04-02T11:17:37Z

!build

dali-automaton · 2026-04-02T11:31:55Z

CI MESSAGE: [47536002]: BUILD STARTED

dali-automaton · 2026-04-02T15:11:58Z

CI MESSAGE: [47536002]: BUILD PASSED

mdabek-nvidia and others added 10 commits March 19, 2026 13:30

Center crop operator

e09d04b

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Review fixes

1b043c0

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Review fixes

f30e282

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Apply suggestion from @stiepan

3337aa9

Signed-off-by: Marek Dabek <mdabek@nvidia.com> Co-authored-by: Kamil Tokarski <kamiltokarski04@gmail.com>

Torchvision ColorJitter and Grayscale implementations

ebbf863

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Review fixes

d8c5cd4

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Review fixes

4dc37b5

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Review fixes

e146f9f

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Review fixes

6a0038d

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Review fixes

f53d4ec

Signed-of-by: Marek Dabek <mdabek@nvidia.com>

mdabek-nvidia force-pushed the torchvision_gaussian_blur branch from 6056187 to 6ddc68a Compare March 24, 2026 14:25

github-advanced-security AI found potential problems Mar 24, 2026

View reviewed changes

Comment thread dali/python/nvidia/dali/experimental/torchvision/v2/color.py Fixed

mdabek-nvidia force-pushed the torchvision_gaussian_blur branch from 5521fe5 to ef0bf89 Compare March 24, 2026 14:31

greptile-apps Bot reviewed Mar 24, 2026

View reviewed changes

mdabek-nvidia force-pushed the torchvision_gaussian_blur branch from afc4838 to 455f51e Compare March 25, 2026 08:46

mdabek-nvidia marked this pull request as ready for review March 25, 2026 08:52

dali-automaton assigned banasraf, rostan-t and mzient Mar 25, 2026

mdabek-nvidia force-pushed the torchvision_gaussian_blur branch 2 times, most recently from ea898cb to abdd2e6 Compare March 25, 2026 17:42

mdabek-nvidia and others added 5 commits April 1, 2026 10:37

Review fixes and validation renaming

0eb7bd7

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Torchvision API - center crop operator (NVIDIA#6266)

7fd89d9

Center crop operators implementation Signed-off-by: Marek Dabek <mdabek@nvidia.com> Co-authored-by: Kamil Tokarski <kamiltokarski04@gmail.com>

Torchvision ColorJitter and Grayscale implementations

453d18b

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Review fixes

2bf954c

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Gaussian blur operator

842dbea

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

mdabek-nvidia force-pushed the torchvision_gaussian_blur branch from 26edf14 to 946223a Compare April 2, 2026 07:51

Review fixes

b940dc6

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

mdabek-nvidia force-pushed the torchvision_gaussian_blur branch from 08fd716 to b940dc6 Compare April 2, 2026 07:57

mdabek-nvidia unassigned mzient Apr 2, 2026

banasraf reviewed Apr 2, 2026

View reviewed changes

rostan-t reviewed Apr 2, 2026

View reviewed changes

rostan-t and others added 5 commits April 2, 2026 11:38

Torchvision API - ColorJitter and Grayscale operators (NVIDIA#6272)

5652054

* Torchvision ColorJitter and Grayscale implementations Signed-off-by: Marek Dabek <mdabek@nvidia.com>

Gaussian blur - review fixes

feb2508

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

mdabek-nvidia force-pushed the torchvision_gaussian_blur branch 3 times, most recently from 9a2fbb5 to 73b0d56 Compare April 2, 2026 11:10

Improve type hinting in functional API

52c9006

Signed-off-by: Marek Dabek <mdabek@nvidia.com>

mdabek-nvidia force-pushed the torchvision_gaussian_blur branch from 73b0d56 to 52c9006 Compare April 2, 2026 11:14

Merge branch 'main' into torchvision_gaussian_blur

7eb7bc1

banasraf approved these changes Apr 2, 2026

View reviewed changes

rostan-t approved these changes Apr 2, 2026

View reviewed changes

mdabek-nvidia merged commit fa4e029 into NVIDIA:main Apr 2, 2026
7 checks passed

	f" Kernel size should be an int or a sequenece of ints, got {kernel_size}"
	f"Kernel size should be an int or a sequence of ints, got {kernel_size}"

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!

Uh oh!

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

Uh oh!

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: 5/5

Important Files Changed

Flowchart

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dali-automaton commented Mar 24, 2026

Uh oh!

mdabek-nvidia commented Mar 25, 2026

Uh oh!

mdabek-nvidia commented Mar 25, 2026

Uh oh!

dali-automaton commented Mar 25, 2026

Uh oh!

dali-automaton commented Mar 25, 2026

Uh oh!

dali-automaton commented Mar 31, 2026

Uh oh!

dali-automaton commented Mar 31, 2026

Uh oh!

banasraf Apr 2, 2026

Choose a reason for hiding this comment

Uh oh!

banasraf Apr 2, 2026

Choose a reason for hiding this comment

Uh oh!

banasraf Apr 2, 2026

Choose a reason for hiding this comment

Uh oh!

mdabek-nvidia Apr 2, 2026

Choose a reason for hiding this comment

Uh oh!

banasraf Apr 2, 2026

Choose a reason for hiding this comment

Uh oh!

mdabek-nvidia Apr 2, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mdabek-nvidia commented Apr 2, 2026

Uh oh!

dali-automaton commented Apr 2, 2026

Uh oh!

dali-automaton commented Apr 2, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

mdabek-nvidia commented Mar 24, 2026 •

edited

Loading

greptile-apps Bot commented Mar 24, 2026 •

edited

Loading