🚀 feat(model): Add Dinomaly Model #2835
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…dation steps - Added detailed docstrings for the Dinomaly class and its methods. - Improved error handling in training and validation steps. - Updated pre-processor configuration to include crop size validation. - Refined output structure in the training step for clarity.
…zer configuration; enhance Gaussian kernel function
… SSLMetaArch implementation - Deleted `train.py`, `__init__.py`, and `ssl_meta_arch.py` files from the DINOv2 training module. - Removed unused imports and commented-out code in `vit_encoder.py`. - Streamlined the model loading process and eliminated unnecessary complexity in the architecture. - Ensured that the remaining code adheres to the latest standards and practices for clarity and maintainability.
…er and torch model classes
- Rearranged import statements for better organization and consistency. - Updated type hints to use the new syntax for optional types. - Simplified conditional checks and improved readability in various functions. - Enhanced logging messages for clarity during model loading and training. - Modified the `get_params_groups_with_decay` function to improve parameter handling. - Updated the `DinoV2Loader` class to streamline model loading and weight management. - Improved the `ViTill` class by refining feature processing and anomaly map calculations. - Adjusted the `simple_script.py` to utilize the new export types for model exporting. - Reduced the number of epochs in the training script for quicker testing.
… clarity and accuracy style: adjust training configuration in simple_script.py
…integration refactor: enhance training configuration and streamline model initialization in ViTill chore: add benchmark configuration and script for Padim model evaluation fix: update simple script for MVTecAD category and improve timing output
…related utilities refactor: update attention and drop path layers for improved efficiency and clarity
… timm library equivalents and clean up unused code
… clarity; remove unused layer files
…ding initialization
… handling - Added type hints and ClassVar annotations in model_loader.py for better clarity and type checking. - Enhanced error messages in model_loader.py to provide clearer guidance on model name and architecture issues. - Updated global_cosine_hm_percent and modify_grad functions in utils.py with type hints and improved gradient modification logic. - Improved documentation and type hints in vision_transformer.py, including detailed docstrings for methods and parameters. - Refined training configuration in lightning_model.py with type hints and assertions for better validation of input parameters. - Enhanced ViTill class in torch_model.py with static methods and type safety checks for architecture configuration. - General code cleanup and consistency improvements across all modified files.
…er; remove unused max_steps from training config
…oss function and dynamic training steps
…rectory handling in DinoV2Loader
…ep configuration in Dinomaly model
rajeshgangireddy
changed the title
There was a problem hiding this comment.
Pull Request Overview
This PR adds the Dinomaly model—a Vision Transformer-based anomaly detection approach—into anomalib, including both a raw PyTorch implementation and a Lightning training/inference module.
- Introduces
ViTillPyTorch model with encoder, bottleneck, decoder and map computation - Adds
DinomalyLightningModule for training, validation, optimizer/scheduler setup - Bundles DINOv2 transformer components, training utilities (loss, optimizer, scheduler), model loader, layer definitions, and example config
Reviewed Changes
Copilot reviewed 12 out of 12 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| src/anomalib/models/image/dinomaly/torch_model.py | Core ViTill model implementation & anomaly-map computation |
| src/anomalib/models/image/dinomaly/lightning_model.py | LightningModule wrapping ViTill for training/inference |
| src/anomalib/models/image/dinomaly/components/vision_transformer.py | DINOv2 Vision Transformer implementation |
| src/anomalib/models/image/dinomaly/components/training_utils.py | Loss, optimizer, scheduler utilities |
| src/anomalib/models/image/dinomaly/components/model_loader.py | Loader for pretrained DINOv2 weights |
| src/anomalib/models/image/dinomaly/components/layers.py | Attention, MLP, transformer block components |
| src/anomalib/models/image/dinomaly/components/init.py | Component exports |
| src/anomalib/models/image/dinomaly/init.py | Dinomaly module export |
| src/anomalib/models/image/init.py | Registered Dinomaly in image models |
| src/anomalib/models/init.py | Registered Dinomaly globally |
| examples/configs/model/dinomaly.yaml | Example YAML config for Dinomaly |
| src/anomalib/models/image/dinomaly/README.md | Model README with usage and benchmarks placeholder |
src/anomalib/models/image/dinomaly/components/training_utils.py
Outdated
Show resolved
Hide resolved
… for Dinomaly model;
samet-akcay
requested changes
Jul 17, 2025
There was a problem hiding this comment.
Thanks @rajeshgangireddy, overall the code is quite clean and easy to follow!
I've added some comments to discuss
src/anomalib/models/image/dinomaly/components/training_utils.py
Outdated
Show resolved
Hide resolved
src/anomalib/models/image/dinomaly/components/vision_transformer.py
Outdated
Show resolved
Hide resolved
| on increasingly difficult examples as training progresses. | ||
| """ | ||
| del args, kwargs # These variables are not used. | ||
| try: |
There was a problem hiding this comment.
why would there be a training error?
There was a problem hiding this comment.
Also is it possible to move this loss computation to the torch model by any chance?
There was a problem hiding this comment.
Nope. It is now removed.
Loss calculation for this model depends on an additional components - p which get's updated based on the number of trained epochs.
p (float): Percentage of well-reconstructed points to down-weight (0.0 to 1.0).
Higher values make training focus on fewer, harder examples. Default is 0.9.
That's why lightning model (which has access to this) is a good place to calculate loss.
Let me think, if we can move loss to torch model.
There was a problem hiding this comment.
Hmm tricky let's discuss this. This makes it valid to keep loss in lightning module. Let's discuss what should be the scope of torch model and what should be in the lightning model. I think we need to refactor all the models to ensure that our inferencers work correctly.
| max_epochs * len(self.trainer.datamodule.train_dataloader()), | ||
| ) | ||
|
|
||
| optimizer_config = TRAINING_CONFIG["optimizer"] |
There was a problem hiding this comment.
@ashwinvaidya17, can you confirm if this dict-based config is API/CLI friendly ?
samet-akcay
reviewed
Jul 18, 2025
samet-akcay
reviewed
Jul 18, 2025
samet-akcay
reviewed
Jul 18, 2025
Refactor Dinomaly model as per comments plus some small improvements. Attention class was used only for type checking - replaced it with attention from timm. Separated loss and optimiser into different files. Code readability improvements. Added explicitly pass the image size to create anomaly maps instead of a default. Docstring improvements. Removed unnecessary TRY CATCH blocks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
📝 Description
Pending:
✨ Changes
Select what type of change your PR is:
✅ Checklist
Before you submit your pull request, please make sure you have completed the following steps:
For more information about code review checklists, see the Code Review Checklist.
Anomaly Maps
###Metrics
<style> </style>*NOTE : The Pixel_F1 are slighly off as the paper publishes F1_Max and I benchmarked F1Score. I am running some benchmarks to get F1_Max and will update the table.
Overall, it looks like the difference is less than 2%