refactor(data): unify Perlin noise helpers and expose synthetic blend factor. by geeky33 · Pull Request #3447 · open-edge-platform/anomalib

geeky33 · 2026-03-20T23:06:05Z

Behavior

Unifies Perlin noise generation (including GLASS-friendly options), shares threshold rescaling across DSR and SuperSimpleNet, fixes internal padding when pad_to_power_of_2=False so odd sizes don’t break, and exposes synthetic_blend_factor on image datamodules so synthetic anomaly intensity can be set from config/API (forwarded into SyntheticAnomalyDataset / make_synthetic_dataset).

Description

This PR consolidates Perlin-related logic in generate_perlin_noise (configurable scale exponent range, optional power-of-two padding, generate_perlin_noise_glass, apply_perlin_threshold_rescale), reuses that helper in DSR and SuperSimpleNet instead of duplicated code, and threads blend_factor through synthetic dataset creation. Image datamodules gain synthetic_blend_factor (default (0.01, 0.2)), with example/docs YAML updates and unit tests.

Fixes #(issue) #3413

Snapshots for supporting the PR :

Changes

New feature (non-breaking): synthetic_blend_factor on datamodules / synthetic pipeline
Refactor: shared Perlin utilities and model anomaly generators
Bug fix: padding / shape correctness when not using power-of-two padding
Tests: test_perlin.py; updates to test_synthetic.py, test_folder.py
Documentation: config snippets + examples/configs/data/folder.yaml

Checklist

I have made the necessary updates to the documentation (where applicable).
I have written tests that support these changes (where applicable).
My PR title follows conventional commit format.

…ehavior Signed-off-by: geeky33 <aaryap1204@gmail.com>

geeky33 · 2026-03-20T23:08:35Z

Hello, @ashwinvaidya17 @rajeshgangireddy
Could you please review the PR ?

Thank you
Aarya.

Copilot

Pull request overview

This PR refactors synthetic anomaly generation by centralizing Perlin noise utilities (including GLASS-compatible defaults), reusing shared threshold-rescaling logic across models, and exposing a configurable synthetic_blend_factor on image datamodules so synthetic anomaly intensity can be controlled via config/API.

Changes:

Consolidates Perlin helpers (default/GLASS scale ranges, optional power-of-two padding, threshold rescaling) and reuses them in DSR and SuperSimpleNet anomaly generators.
Threads blend_factor through make_synthetic_dataset / SyntheticAnomalyDataset and forwards it from image datamodules via synthetic_blend_factor.
Adds/updates unit tests and updates example/docs YAML snippets to show synthetic_blend_factor.

Reviewed changes

Copilot reviewed 27 out of 27 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
`src/anomalib/data/utils/generators/perlin.py`	Adds shared Perlin utilities, GLASS defaults, and PerlinAnomalyGenerator compatibility knobs.
`src/anomalib/data/utils/generators/__init__.py`	Re-exports new Perlin helpers/constants from generators package.
`src/anomalib/data/utils/__init__.py`	Re-exports new Perlin helpers/constants from data utils top-level API.
`src/anomalib/data/utils/synthetic.py`	Adds `blend_factor` plumbing into synthetic dataset creation and dataset class.
`src/anomalib/data/datamodules/base/image.py`	Adds `synthetic_blend_factor` to datamodule and forwards to synthetic split creation.
`src/anomalib/data/datamodules/image/{folder,bmad,btech,datumaro,kaputt,kolektor,mpdd,mvtecad,mvtecad2,mvtec_loco,realiad,tabular,vad,visa}.py`	Threads `synthetic_blend_factor` through concrete image datamodules into base class.
`src/anomalib/models/image/dsr/anomaly_generator.py`	Replaces duplicated Perlin threshold-rescale logic with shared helper.
`src/anomalib/models/image/supersimplenet/anomaly_generator.py`	Replaces duplicated Perlin threshold-rescale logic with shared helper.
`tests/unit/data/utils/test_perlin.py`	Adds unit tests for shared Perlin utilities and GLASS helper.
`tests/unit/data/utils/test_synthetic.py`	Adds/updates tests to ensure blend factor is exposed and configurable.
`tests/unit/data/datamodule/image/test_folder.py`	Tests that datamodule forwards `synthetic_blend_factor` into synthetic split creation.
`examples/configs/data/folder.yaml`	Documents new `synthetic_blend_factor` config field.
`docs/source/snippets/.../normal_and_synthetic.yaml`	Updates docs snippets to include `synthetic_blend_factor`.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-03-20T23:10:05Z

@@ -117,6 +122,7 @@ def __init__(
        self.val_split_mode = ValSplitMode(val_split_mode) if val_split_mode else ValSplitMode.NONE
        self.val_split_ratio = val_split_ratio or 0.5
        self.seed = seed
+        self.synthetic_blend_factor = synthetic_blend_factor


synthetic_blend_factor is documented/configured as a YAML list (e.g. [0.01, 0.2]), which will typically arrive here as a list/OmegaConf ListConfig, not a tuple. Downstream, PerlinAnomalyGenerator only treats tuple as a range, so passing a list can silently ignore the configured range. Consider normalizing synthetic_blend_factor in __init__ (e.g., coerce 2-item list-like values to tuple[float, float]) or broadening downstream checks to accept Sequence.

Copilot · 2026-03-20T23:10:05Z

    # Ensure dimensions are powers of 2 for proper noise generation
    def nextpow2(value: int) -> int:
        return int(2 ** torch.ceil(torch.log2(torch.tensor(value))).int().item())

-    pad_h = nextpow2(height)
-    pad_w = nextpow2(width)
+    pad_h_base = nextpow2(height) if pad_to_power_of_2 else height
+    pad_w_base = nextpow2(width) if pad_to_power_of_2 else width
+    pad_h = ((pad_h_base + scalex - 1) // scalex) * scalex
+    pad_w = ((pad_w_base + scaley - 1) // scaley) * scaley


The comment # Ensure dimensions are powers of 2 for proper noise generation is no longer accurate now that pad_to_power_of_2 can be False and the padding is also rounded to a multiple of the chosen scale. Updating the comment (and/or docstring) would help avoid confusion about what invariants are actually required for the algorithm.

Copilot · 2026-03-20T23:10:05Z

+    denominator = perlin_noise.max() - perlin_noise.min()
+    if denominator == 0:
+        return perlin_noise
+
+    perlin_noise = (perlin_noise - perlin_noise.min()) / denominator


apply_perlin_threshold_rescale recomputes perlin_noise.min() multiple times. Since this runs on tensors (often on GPU), it would be cleaner (and avoids extra reductions) to compute min_val/max_val once, reuse them for denominator and normalization, and then rescale.

Suggested change

denominator = perlin_noise.max() - perlin_noise.min()

if denominator == 0:

return perlin_noise

perlin_noise = (perlin_noise - perlin_noise.min()) / denominator

min_val = perlin_noise.min()

max_val = perlin_noise.max()

denominator = max_val - min_val

if denominator == 0:

return perlin_noise

perlin_noise = (perlin_noise - min_val) / denominator

rajeshgangireddy · 2026-03-24T09:21:17Z

Hi @geeky33
Thanks for the changes.
Please do note that currently due to many PRs, it might take some time before we start reviewing this PR.

geeky33 · 2026-03-24T11:36:35Z

Hi @rajeshgangireddy
No issues,
Thank you.

Copilot

Pull request overview

Copilot reviewed 27 out of 27 changed files in this pull request and generated 2 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-03-24T12:29:19Z

        anomaly_source_path: Path | str | None = None,
        probability: float = 0.5,
        blend_factor: float | tuple[float, float] = (0.2, 1.0),
        rotation_range: tuple[float, float] = (-90, 90),
+        *,
+        perlin_pad_to_power_of_2: bool = True,
+        perlin_scale_exponent_range: tuple[int, int] = DEFAULT_PERLIN_SCALE_EXPONENT_RANGE,
+        perlin_rescale_below_threshold: bool = True,


PerlinAnomalyGenerator only samples a range when blend_factor is a tuple; if a config/API passes a 2-item list (common from YAML/omegaconf), the later beta selection falls through to the fallback constant (0.5), effectively ignoring the configured range. Consider normalizing blend_factor to a tuple in __init__ (e.g., accept Sequence[float] length-2) or updating the type guards so both list and tuple ranges are handled.

Copilot · 2026-03-24T12:29:19Z

+    def __init__(
+        self,
+        augmentations: Transform | None,
+        source_samples: DataFrame,
+        dataset_name: str,
+        blend_factor: float | tuple[float, float] = (0.01, 0.2),
+    ) -> None:


SyntheticAnomalyDataset.__init__ now accepts blend_factor, but the class docstring/Args section above still documents only augmentations, source_samples, and dataset_name. Please update the docstring to describe blend_factor (fixed float vs sampled range) so public API docs stay accurate.

geeky33 requested review from ashwinvaidya17, rajeshgangireddy and samet-akcay as code owners March 20, 2026 23:06

Copilot AI review requested due to automatic review settings March 20, 2026 23:06

Copilot started reviewing on behalf of geeky33 March 20, 2026 23:06 View session

refactor(data): expose synthetic blend factor and centralize perlin b…

89ea805

…ehavior Signed-off-by: geeky33 <aaryap1204@gmail.com>

geeky33 force-pushed the anomaly branch from ee378ee to 89ea805 Compare March 20, 2026 23:07

Copilot AI reviewed Mar 20, 2026

View reviewed changes

Merge branch 'main' into anomaly

7534b30

Merge branch 'open-edge-platform:main' into anomaly

ba590db

Copilot AI review requested due to automatic review settings March 24, 2026 12:22

Copilot started reviewing on behalf of geeky33 March 24, 2026 12:23 View session

Copilot AI reviewed Mar 24, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

refactor(data): unify Perlin noise helpers and expose synthetic blend factor.#3447

refactor(data): unify Perlin noise helpers and expose synthetic blend factor.#3447
geeky33 wants to merge 3 commits intoopen-edge-platform:mainfrom
geeky33:anomaly

geeky33 commented Mar 20, 2026

Uh oh!

geeky33 commented Mar 20, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

rajeshgangireddy commented Mar 24, 2026

Uh oh!

geeky33 commented Mar 24, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Mar 24, 2026

Uh oh!

Copilot AI Mar 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

geeky33 commented Mar 20, 2026

Behavior

Description

Changes

Checklist

Uh oh!

geeky33 commented Mar 20, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

rajeshgangireddy commented Mar 24, 2026

Uh oh!

geeky33 commented Mar 24, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Copilot AI Mar 24, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 24, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants