openvinotoolkit/anomalib

An anomaly detection library comprising state-of-the-art algorithms and features such as experiment management, hyper-parameter optimization, and edge inference.

anomalib's Issues

🧪 Refactor nightly/pre-merge tests to follow the new refactor design

Restructure tests to follow the same folder structure as the repo.

Depends on

• PR #66
• #87

Code attribution

Hi,

Your implementation of PatchCore coreset selection:

anomalib/anomalib/models/patchcore/model.py

Line 212 in 3613a6e

def create_coreset(

is VERY similar to my implementation, which I have designed and coded from scratch:
https://github.com/rvorias/ind_knn_ad/blob/57de628b334f9dd62316ac7770db6e963b71ee8e/indad/utils.py#L36

Please elaborate on this.

More details:
this repo:

        select_idx = 0
        last_item = sample_set[select_idx : select_idx + 1]
        coreset_idx = [torch.tensor(select_idx).to(embedding.device)]  # pylint: disable=not-callable
        min_distances = torch.linalg.norm(sample_set - last_item, dim=1, keepdims=True)

        for _ in range(sample_count - 1):
            distances = torch.linalg.norm(sample_set - last_item, dim=1, keepdims=True)  # broadcast
            min_distances = torch.minimum(distances, min_distances)  # iterate
            select_idx = torch.argmax(min_distances)  # select

            last_item = sample_set[select_idx : select_idx + 1]
            min_distances[select_idx] = 0
            coreset_idx.append(select_idx)

        coreset_idx = torch.stack(coreset_idx)
        self.memory_bank = embedding[coreset_idx]

my repo:

    select_idx = 0
    last_item = z_lib[select_idx:select_idx+1]
    coreset_idx = [torch.tensor(select_idx)]
    min_distances = torch.linalg.norm(z_lib-last_item, dim=1, keepdims=True)
    # The line below is not faster than linalg.norm, although i'm keeping it in for
    # future reference.
    # min_distances = torch.sum(torch.pow(z_lib-last_item, 2), dim=1, keepdims=True)
    
    ...
    
    for _ in tqdm(range(n-1), **TQDM_PARAMS):
        distances = torch.linalg.norm(z_lib-last_item, dim=1, keepdims=True) # broadcasting step
        # distances = torch.sum(torch.pow(z_lib-last_item, 2), dim=1, keepdims=True) # broadcasting step
        min_distances = torch.minimum(distances, min_distances) # iterative step
        select_idx = torch.argmax(min_distances) # selection step

        # bookkeeping
        last_item = z_lib[select_idx:select_idx+1]
        min_distances[select_idx] = 0
        coreset_idx.append(select_idx.to("cpu"))

    return torch.stack(coreset_idx)

Inference with OpenCV

Hello,
Thank you for providing such a fantastic repo!

I was curious if you could show a tutorial on how to use a custom trained model on a live stream webcam with OpenCV. Perhaps with torch and with openvino?

import cv2

cap = cv2.VideoCapture(0)

# Check if the webcam is opened correctly
if not cap.isOpened():
    raise IOError("Cannot open webcam")

while True:
    ret, frame = cap.read()
    
    # infer with custom model
    # show results on output_frame
    
    cv2.imshow('output', output_frame)

    k = cv2.waitKey(1)
    if k == 27:
        break

cap.release()
cv2.destroyAllWindows()

Thank you in advance 😃

Refactor DFKDE NormalityModel to contain the feature extractor

Refactor according to comments on PR #122

This would probably work, but I feel like the model would still need another refactoring. Ideally DfkdeModel should be the normality model that contains feature_extractor, pca and kde modules. I guess this would be scope of another PR, but something to keep in mind.
Alternatively, there could be a better naming than NormalityModel if we want to keep NormalityModel. For instance, in DFM, there SingleClassGaussian. Again, in another PR

Reference to comment: #122 (comment)

Match up with the latest torch version

Is your feature request related to a problem? Please describe.
Currently, it supports torch==1.8.1. Due to some compatibility issue, I couldn't use this library locally conveniently with the other latest versions locally. So, I think it would be better to support torch>=1.8.1

https://github.com/openvinotoolkit/anomalib/blob/development/requirements/base.txt

Use console scripts to replace `python tools/train.py`

To run training or evaluation, we call a python script such as python tools/train.py --model padim. With console scripts, it would be possible to run the same code via anomalib fit --model padim. To achieve this we need to modify setup.py. For more information, see this

Docs fail in Actions

CI build passes but docs fail. Would be good to check.

About model convert

Is there any supplement about .ckpt converted to .onnx, .bin or .xml?

Diverged metrics in PatchCore: dropping from 0.99 to 0.44 and 0.03 is rather critical?

diverged metrics: dropping from 0.99 to 0.44 and 0.03 is rather critical?
log images: nice :)
Also padim dropped in performance, but not as crazy.
Here's a patchcore result of "good" parts:

this is padim:
DATALOADER:0 TEST RESULTS
{'image_AUROC': 0.7589669823646545,
'image_F1': 0.8787878751754761,
'pixel_AUROC': 0.9781586527824402,
'pixel_F1': 0.22379672527313232}

Originally posted by @sequoiagrove in #67 (comment)

OpenVINO and NNCF compression callbacks cannot be configured from `config.yaml`.

Current CLI refactor cannot configure openvino and nncf callbacks from the config.yaml file.

Refactor Anomalib CLI

Current anomalib CLI implements before_instantiate_classes method. However, the implementation is not clean and needs refactoring.

Tasks

• Use console scripts to replace python tools/train.py

Refactor: Modify all callbacks to return `AnomalyModule` instead of `pl_module`

Ensure that all callbacks return AnomalyModule instead of pl_module. Otherwise we have to use cast statements. Eg-
compress.py uses pl_module = cast(AnomalyModule, pl_module) to placate mypy. This should be changed to ensure that we always pass AnomalyModule.

Implement predict_dataloader in MVTecDataModule

IndexError: tensors used as indices must be long, byte or bool tensors

Describe the bug
When training PatchCore, I'm getting the following error:
IndexError: tensors used as indices must be long, byte or bool tensors
caused by the line

anomalib/anomalib/models/components/dimensionality_reduction/random_projection.py

Line 85 in c33d25a

components[i, c_idx] = data.double()

To Reproduce
Steps to reproduce the behavior:
Run the PatchCore algorithm on CPU (maybe on GPU too)

The problem can be solved by casting to c_idx to c_idx.long()

Heavy Augmentations + More Epochs

Hello!
I'm wondering if anyone has experimented with heavy augmentations and therefore increasing the number of training epochs.
If so, did you get any performance gains and/or information about conflicting augmentations?

Thanks!

Update Pre-commit Documentation

• Pre-commit documentation refers to the old requirements_dev.txt format. This needs to be updated to reflect requirements/dev.txt
• Also, command for running it on all files should be added for reference.

1 epoch

Is it meant to be only one epoch in training?
Your config files state 1 epoch, is that just as quick example?
I tried to train PADIM for 10 on mvtec leather and wood and the metrics stay the same anyway, so it seems nothing as gained by training more.
Lightning module also warn that there is no optimizer so I guess train only finds the correct thresholds and that takes 1 epoch.

Rename normalization to cdf

When we created this normalisation, we thought this would be the only method we would use. However, we implemented more methods such as min-max, due to which we need to explicitly rename this method to cdf normalisation

On Test Model load callback does not load recently created model

Model is loading default model from config file

• While training model multiple times, new model file version gets created however the old model gets loaded for testing and results are evaluated with the old model. Any change in data or code does not get reflected in the results.

To Reproduce
Steps to reproduce the behavior:

1. Train a model with subset of images
2. Retrain with complete dataset
3. Evaluate performance

Always reorder auc inputs after upgrading to pytorch 1.9.x

In the AUROC metric computation, we should ideally always reorder the inputs. Currently the lack of support for stable sorting in pytorch 1.8.1 prevents us from doing so (more detailed explanation here). Stable sorting was added in pytorch 1.9.0, so once we upgrade our requirements to that version, we should be able to always reorder the inputs using the stable=True option of torch.sort().

patchcore torch.pairwise_distance OOM on larger datasets

I get CUDA memory blowing when doing larger datasets

return torch.pairwise_distance(x1, x2, p, eps, keepdim)
RuntimeError: CUDA out of memory. Tried to allocate 724.00 MiB (GPU 0; 11.00 GiB total capacity; 7.62 GiB already allocated; 190.31 MiB free; 8.77 GiB reserved in total by PyTorch)

Is there anything that could be done here? like moving this computation to CPU?

How to train in a custom dataset?

I am having a bit of trouble training on a custom dataset. I reproduced the MVTec structure but I still get some errors like:

IndexError: too many indices for tensor of dimension 0

Ganomaly inference.py error

AttributeError: 'GanomalyLightning' object has no attribute 'model'

Add GANomaly metrics to READMEs

PR #70 does not add GANomaly metrics to the readmes.
It will be more convenient to add it after PR #17 is merged.

The READMEs should follow a similar structure as DFM and DFKDE

Tasks

• GANomaly readme
• Repository readme

Check whether `best_model` or `latest_model` is used for `validate` and `test` methods in `tools/train.py`

Describe the bug

• POTENTIAL BUG
• We need to check which model weights are used for validation and test in tools/train.py. Is it the best model or the latest model?

To Reproduce
Steps to reproduce the behavior:

1. Debug python tools/train.py by monitoring ModelCheckpoint callback.

Expected behavior

• Best model should be used for validate and test methods.

Fix NNCF Issues.

OpenvinoCallback and NNCFCallback implementations are quite similar. They only differ in terms of type of the quantization.

Tasks

• Refactor Compression and NNCF callbacks could be merged into a single callback
• Design an error handling mechanism to pick up the errors caused by nncf 2.1.0 vs 2.2.0. Warn the user to manually upgrade the nncf to 2.2.0
• configure openvino and nncf callbacks from the config.yaml file.
• change the requirements in openvino.txt - nncf>=2.1.0
• Modify the nightly tests to handle nncf tests.
• Refactor NNCF config.

CUDA out of Memory when training `STFPM`

Got this error when training STFPM on pill category.
Occurred when training from the benchmarking script.

Error

output = self.trainer.call_hook('validation_step_end', *args, **kwargs)                                                                        
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/pytorch_lightning/trainer/trainer.py", line 1230, in call_hook           
    output = hook_fx(*args, **kwargs)                                                                                                              
  File "/home/ashwin/anomalib/anomalib/core/model/anomaly_module.py", line 105, in validation_step_end                                             
    self.pixel_metrics(val_step_outputs["anomaly_maps"].flatten(), val_step_outputs["mask"].flatten().int())
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/collections.py", line 110, in forward
    return {k: m(*args, **m._filter_kwargs(**kwargs)) for k, m in self.items()}
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/collections.py", line 110, in <dictcomp>
    return {k: m(*args, **m._filter_kwargs(**kwargs)) for k, m in self.items()}
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/metric.py", line 205, in forward
    self._forward_cache = self.compute()
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/metric.py", line 367, in wrapped_func
    self._computed = compute(*args, **kwargs)
  File "/home/ashwin/anomalib/anomalib/core/metrics/optimal_f1.py", line 38, in compute
    precision, recall, thresholds = self.precision_recall_curve.compute() 
File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/metric.py", line 367, in wrapped_func
    self._computed = compute(*args, **kwargs)
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/classification/precision_recall_curve.py", line 148, in comp
ute
    return _precision_recall_curve_compute(preds, target, self.num_classes, self.pos_label)
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/functional/classification/precision_recall_curve.py", line 2
60, in _precision_recall_curve_compute
    return _precision_recall_curve_compute_single_class(preds, target, pos_label, sample_weights)
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/functional/classification/precision_recall_curve.py", line 1
40, in _precision_recall_curve_compute_single_class
    fps, tps, thresholds = _binary_clf_curve(
  File "/home/ashwin/miniconda3/envs/anomalib/lib/python3.8/site-packages/torchmetrics/functional/classification/precision_recall_curve.py", line 3
6, in _binary_clf_curve
    desc_score_indices = torch.argsort(preds, descending=True)
RuntimeError: CUDA out of memory. Tried to allocate 5.82 GiB (GPU 0; 23.70 GiB total capacity; 9.28 GiB already allocated; 2.61 GiB free; 19.41 GiB
 reserved in total by PyTorch)

Relevant Config

• Image size = 256

A few questions

Hi，there are some codes which make me confused, hope to get your answer.

1. In padim/config.yaml, what does nncf(belongs to optimization) mean? and what does function update_nncf_config mean?
2. If i want to use multiple gpus to train, how should i set? and there are many parameters in trainer section which i wanna figure out each is used for and be clear which can be tuned or not.

Resize bug in inference

• When input the width and height of the image are not equal, sizes of input arguments that do not match will occur

Line 152 in ./anomalib/deploy/inferencers/torch.py should be anomaly_map = cv2.resize(anomaly_map, (meta_data["image_shape"][1], meta_data["image_shape"][0]))

meta_data["image_shape"] is (image_width, image_height)
-# cv2.resize(img, (image_height，image_width ),interpolation=cv2.INTER_CUBIC)

Refactor benchmarking script

• Add grid search instead of nested for loops
• Run gpu and cpu training in parallel as they don't share resources.

Setup codecov with github actions.

anomaly scores computed by patchcore model not used

The patchcore anomaly map generator computes the image-level anomaly scores, but we don't seem to use them: https://github.com/openvinotoolkit/anomalib/blob/development/anomalib/models/patchcore/model.py#L319

I think we should either use the scores and assign them to batch["pred_scores"], or not compute them at all.

ModuleNotFoundError: No module named 'anomalib.utils.callbacks'

Describe the bug
-anomalib package install error, seems the package released do not include all function of utils file

To Reproduce
Steps to reproduce the behavior:

1. Go to https://pypi.org/project/anomalib/0.2.4/#files download the file anomalib-0.2.4-py3-none-any.whl
2. pip install anomalib-0.2.4-py3-none-any.whl
3. cd anomalib
4. python tools/train.py
5. See error ModuleNotFoundError: No module named 'anomalib.utils.callbacks'

Expected behavior
-seems the package released on pypi.org do not include all function of the git file anomalib,such as utils file.

Support python 3.7

Is your feature request related to a problem? Please describe.
I mainly use anomalib on Google Colab but it only supports python=3.7 at the moment. There is an error when running the training command:
python -m tools.train --model_config_path path/to/config.yaml

The error was:

Traceback (most recent call last):
  File "/usr/lib/python3.7/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/usr/lib/python3.7/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/content/anomalib/tools/train.py", line 66, in <module>
    train()
  File "/content/anomalib/tools/train.py", line 55, in train
    model = get_model(config)
  File "/content/anomalib/anomalib/models/__init__.py", line 61, in get_model
    module = import_module(f"anomalib.models.{config.model.name}.model")
  File "/usr/lib/python3.7/importlib/__init__.py", line 127, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
  File "<frozen importlib._bootstrap>", line 983, in _find_and_load
  File "<frozen importlib._bootstrap>", line 953, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
  File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
  File "<frozen importlib._bootstrap>", line 983, in _find_and_load
  File "<frozen importlib._bootstrap>", line 967, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 677, in _load_unlocked
  File "<frozen importlib._bootstrap_external>", line 728, in exec_module
  File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
  File "/content/anomalib/anomalib/models/stfpm/__init__.py", line 17, in <module>
    from .model import (  # noqa # pylint: disable=unused-import
  File "/content/anomalib/anomalib/models/stfpm/model.py", line 31, in <module>
    from anomalib.data.tiler import Tiler
  File "/content/anomalib/anomalib/data/tiler.py", line 19, in <module>
    from typing import Optional, Sequence, SupportsIndex, Tuple, Union
ImportError: cannot import name 'SupportsIndex' from 'typing' (/usr/lib/python3.7/typing.py)

Suggestion
AFAIK, SupportsIndex is used for type-hint of classes that have __item__ method, but the default value doesn't seem right and int is better.

After making the below modification:

diff --git a/anomalib/data/tiler.py b/anomalib/data/tiler.py
index 5a274c2..d061337 100644
--- a/anomalib/data/tiler.py
+++ b/anomalib/data/tiler.py
@@ -16,7 +16,7 @@
 
 from itertools import product
 from math import ceil
-from typing import Optional, Sequence, SupportsIndex, Tuple, Union
+from typing import Optional, Sequence, Tuple, Union
 
 import torch
 import torchvision.transforms as T
@@ -161,7 +161,7 @@ class Tiler:
         stride: Union[int, Sequence],
         remove_border_count: int = 0,
         mode: str = "padding",
-        tile_count: SupportsIndex = 4,
+        tile_count: int = 4,
     ) -> None:
 
         self.tile_size_h, self.tile_size_w = self.__validate_size_type(tile_size)

The code works well.

I did not have further investigation, e.g. checking dependencies, other models, etc; but I think supporting python 3.7 is possible with minimal changes. I ran with the STFMP model.

Upload the library to PYPI

This is to upload anomalib to pypi so that the library could be installed via pip.

We would have two options.

1. Basic install that doesn't have openvino. pip install anomalib
2. Full option with openvino. This would be something like pip install anomalib["full"] or something similar.

Whether to install openvino within the basic version is still a question, and needs further discussion.

Create CONTRIBUTING guidelines.

the repo is missing the contributing guidelines. It should show how contribution to the repo could be made.

It should also note that feature extraction-based methods now need to avoid performing validation before collecting the features from the entire training set.

Verify that best (not latest) model is used for testing after early stopping

In train.py, we call trainer.test right after the model has been trained with trainer.train. It is unclear if this leads to correct behaviour when early stopping is applied. Since we do not explicitly load the best model before calling trainer.test, it could be the case that we simply use the latest model. This should be investigated.

Inference Bug - Too many variables to unpack

Hello, thank you for the great repo!

Describe the bug
I cannot run inference.

To Reproduce
Steps to reproduce the behavior:
conda install openvino-ie4py-ubuntu20 -c intel
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge

python tools/inference.py \
    --model_config_path anomalib/models/padim/config.yaml \
    --weight_path results/padim/mvtec/leather/weights/model.ckpt \
    --image_path datasets/MVTec/leather/test/color/000.png

• If applicable, add screenshots to help explain your problem.

Traceback (most recent call last):
  File "tools/inference.py", line 90, in <module>
    infer()
  File "tools/inference.py", line 78, in infer
    output = inference.predict(image=args.image_path, superimpose=True)
  File "anomalib/anomalib/core/model/inference.py", line 90, in predict
    anomaly_map, pred_score = self.post_process(predictions, meta_data=meta_data)
ValueError: too many values to unpack (expected 2)

Hardware and Software Configuration

• OS: [Ubuntu, 20]
• NVIDIA Driver Version [470.57.02]
• CUDA Version [e.g. 11.4]
• CUDNN Version [e.g. v11.4.120]
• OpenVINO Version [Optional e.g. v2021.4.2]

Additional context
I also noticed that when I do inference from a ckpt file it still imports openvino- i expected this not to since it should run purely in pytorch.

e

Hey There!

Who can help me to modify : all_in_one_block.py script
to avoid this warning, to happen:

C:\Users\libro\AppData\Local\Programs\Python\Python38\lib\site-packages\FrEIA\modules\all_in_one_block.py:123: UserWarning: Soft permutation will take a very long time to initialize with 2048 feature channels. Consider using hard permutation instead.
warnings.warn(("Soft permutation will take a very long time to initialize "

I would like to switch to hard permutation mode, but i dont know how to do it.

Thank you so much.

Update README Badges

Due to refactor in workflows, the coverage badge in README is broken. Add badges for nightly and pre-merge

version is hard-coded in `setup.py`. This should be read from either `VERSION.txt` file or in `anomalib.__init__.py`

Add FastFlow, a new Anomaly Detector

Is your feature request related to a problem? Please describe.

• New sota: https://paperswithcode.com/sota/anomaly-detection-on-mvtec-ad
• paper: https://arxiv.org/pdf/2111.07677v2.pdf

Get rid of task_type parameter

The task_type parameter currently dictates which types of ground truth data is passed to the model. When the task_type parameter is set to segmentation, the ground truth pixel masks are passed in addition to the image-level labels. This might no longer be needed. Instead we can just pass whatever ground truth information is available, and let the PL module handle the evaluation based on what type of ground truth data it receives with the batches.

Add `codecov` to CI

Switch to CSV Logger

The repo implements csv logging as a custom callback. Since Pytorch Lightning already has a logger for this, the callback should be replaced with the logger.

🧪 Increase test coverage for Pre-Merge and Nightly

PR #66 splits the tests to nightly and pre-merge. The coverage threshold has been reduced to 71% and 88% for nightly and pre-merge jobs, respectively. It should be brought the coverage back to 90%+.

PatchCore only support 224x224 input, and 28x28 feature maps.

Add proper logging and remove `print` statements with `logger` via `logger = logging.getLogger("pytorch_lightning")`

Describe the bug

• Currently print statements are used to log to console. Instead, logger should be used.
• Polish out pl loggings.

Expected behavior

• Modules should get access to the pytorch lightning logger via logger = logging.getLogger("pytorch_lightning") and log to the console.

Hardware and Software Configuration

• OS: [Ubuntu, OD] - any
• NVIDIA Driver Version [470.57.02] - any
• CUDA Version [e.g. 11.4] - any
• CUDNN Version [e.g. v11.4.120] - any
• OpenVINO Version [Optional e.g. v2021.4.2] - any

Additional context

• N/A

PatchCore: Training only on good images and then inference on good and bad images

First of all, great work! I found this work very interesting! I have a question related to PatchCore implementation.

When there is both training, and testing dataset (images and ground truth) available, then the PatchCore works good and I think during the validation step, it is using the testing data (GT and images) to calculate coreset subsampling.

But imagine a realworld scenario when only good images are available to teach the anomaly detection algorithm, and no defective images are available (GT and images), then how to train the model? I've tried limit_val_batches and limit_test_batches to zero, so during training, no validation and testing are performed. But with this setup, during inference time, the model is producing this error

raceback (most recent call last):
  File "tools/inference.py", line 118, in <module>
    infer()
  File "tools/inference.py", line 101, in infer
    output = inference.predict(image=args.image_path, superimpose=True)
  File "/notebooks/anomalib/anomalib/deploy/inferencers/base.py", line 90, in predict
    predictions = self.forward(processed_image)
  File "/notebooks/anomalib/anomalib/deploy/inferencers/torch.py", line 118, in forward
    return self.model(image)
  File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/notebooks/anomalib/anomalib/models/components/base/anomaly_module.py", line 75, in forward
    return self.model(batch)
  File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/notebooks/anomalib/anomalib/models/patchcore/model.py", line 170, in forward
    patch_scores = self.nearest_neighbors(embedding=embedding, n_neighbors=9)
  File "/notebooks/anomalib/anomalib/models/patchcore/model.py", line 238, in nearest_neighbors
    distances = torch.cdist(embedding, self.memory_bank, p=2.0)  # euclidean norm
  File "/opt/conda/lib/python3.8/site-packages/torch/functional.py", line 1119, in cdist
    return _VF.cdist(x1, x2, p, None)  # type: ignore
RuntimeError: cdist only supports at least 2D tensors, X2 got: 1D

Any guidance will be highly appreciated!

🧪 Refactor nightly thresholds.

Current Behaviour

Nightly job checks whether the tests produce results above a particular threshold. This might not be the best way to handle the tests.

Proposed Behaviour

Run the tests multiple times and report mean and standard deviation. These values should be compared for the nightly job.