Skip to content


Towards Total Recall in Industrial Anomaly Detection

This repository contains the implementation for PatchCore as proposed in Roth et al. (2021),

It also provides various pretrained models that can achieve up to 99.6% image-level anomaly detection AUROC, 98.4% pixel-level anomaly localization AUROC and >95% PRO score (although the later metric is not included for license reasons).


For questions & feedback, please reach out to!

Quick Guide

First, clone this repository and set the PYTHONPATH environment variable with env PYTHONPATH=src python bin/ To train PatchCore on MVTec AD (as described below), run

datapath=/path_to_mvtec_folder/mvtec datasets=('bottle' 'cable' 'capsule' 'carpet' 'grid' 'hazelnut'
'leather' 'metal_nut' 'pill' 'screw' 'tile' 'toothbrush' 'transistor' 'wood' 'zipper')
dataset_flags=($(for dataset in "${datasets[@]}"; do echo '-d '$dataset; done))

python bin/ --gpu 0 --seed 0 --save_patchcore_model \
--log_group IM224_WR50_L2-3_P01_D1024-1024_PS-3_AN-1_S0 --log_online --log_project MVTecAD_Results results \
patch_core -b wideresnet50 -le layer2 -le layer3 --faiss_on_gpu \
--pretrain_embed_dimension 1024  --target_embed_dimension 1024 --anomaly_scorer_num_nn 1 --patchsize 3 \
sampler -p 0.1 approx_greedy_coreset dataset --resize 256 --imagesize 224 "${dataset_flags[@]}" mvtec $datapath

which runs PatchCore on MVTec images of sizes 224x224 using a WideResNet50-backbone pretrained on ImageNet. For other sample runs with different backbones, larger images or ensembles, see

Given a pretrained PatchCore model (or models for all MVTec AD subdatasets), these can be evaluated using


datasets=('bottle'  'cable'  'capsule'  'carpet'  'grid'  'hazelnut' 'leather'  'metal_nut'  'pill' 'screw' 'tile' 'toothbrush' 'transistor' 'wood' 'zipper')
dataset_flags=($(for dataset in "${datasets[@]}"; do echo '-d '$dataset; done))
model_flags=($(for dataset in "${datasets[@]}"; do echo '-p '$loadpath'/'$modelfolder'/models/mvtec_'$dataset; done))

python bin/ --gpu 0 --seed 0 $savefolder \
patch_core_loader "${model_flags[@]}" --faiss_on_gpu \
dataset --resize 366 --imagesize 320 "${dataset_flags[@]}" mvtec $datapath

A set of pretrained PatchCores are hosted here: add link. To use them (and replicate training), check out and

In-Depth Description


Our results were computed using Python 3.8, with packages and respective version noted in requirements.txt. In general, the majority of experiments should not exceed 11GB of GPU memory; however using significantly large input images will incur higher memory cost.

Setting up MVTec AD

To set up the main MVTec AD benchmark, download it from here: Place it in some location datapath. Make sure that it follows the following data tree:

|-- bottle
|-----|----- ground_truth
|-----|----- test
|-----|--------|------ good
|-----|--------|------ broken_large
|-----|--------|------ ...
|-----|----- train
|-----|--------|------ good
|-- cable
|-- ...

containing in total 15 subdatasets: bottle, cable, capsule, carpet, grid, hazelnut, leather, metal_nut, pill, screw, tile, toothbrush, transistor, wood, zipper.

"Training" PatchCore

PatchCore extracts a (coreset-subsampled) memory of pretrained, locally aggregated training patch features:


To do so, we have provided bin/, which uses click to manage and aggregate input arguments. This looks something like

python bin/ \
--gpu <gpu_id> --seed <seed> # Set GPU-id & reproducibility seed.
--save_patchcore_model # If set, saves the patchcore model(s).
--log_online # If set, logs results to a Weights & Biases account.
--log_group IM224_WR50_L2-3_P01_D1024-1024_PS-3_AN-1_S0 --log_project MVTecAD_Results results # Logging details: Name of the run & Name of the overall project folder.

patch_core  # We now pass all PatchCore-related parameters.
-b wideresnet50  # Which backbone to use.
-le layer2 -le layer3 # Which layers to extract features from.
--faiss_on_gpu # If similarity-searches should be performed on GPU.
--pretrain_embed_dimension 1024  --target_embed_dimension 1024 # Dimensionality of features extracted from backbone layer(s) and final aggregated PatchCore Dimensionality
--anomaly_scorer_num_nn 1 --patchsize 3 # Num. nearest neighbours to use for anomaly detection & neighbourhoodsize for local aggregation.

sampler # We now pass all the (Coreset-)subsampling parameters.
-p 0.1 approx_greedy_coreset # Subsampling percentage & exact subsampling method.

dataset # We now pass all the Dataset-relevant parameters.
--resize 256 --imagesize 224 "${dataset_flags[@]}" mvtec $datapath # Initial resizing shape and final imagesize (centercropped) as well as the MVTec subdatasets to use.

Note that contains exemplary training runs to achieve strong AD performance. Due to repository changes (& hardware differences), results may deviate slightly from those reported in the paper, but should generally be very close or even better. As mentioned previously, for re-use and replicability we have also provided several pretrained PatchCore models hosted at add link - download the folder, extract, and pass the model of your choice to bin/ which showcases an exemplary evaluation process.

During (after) training, the following information will be stored:

|PatchCore model (if --save_patchcore_model is set)
|-- models
|-----|----- mvtec_bottle
|-----|-----------|------- nnscorer_search_index.faiss
|-----|-----------|------- patchcore_params.pkl
|-----|----- mvtec_cable
|-----|----- ...
|-- results.csv # Contains performance for each subdataset.

|Sample_segmentations (if --save_segmentation_images is set)

In addition to the main training process, we have also included Weights-&-Biases logging, which allows you to log all training & test performances online to Weights-and-Biases servers ( To use that, include the --log_online flag and provide your W&B key in > --log_wandb_key.

Finally, due to the effectiveness and efficiency of PatchCore, we also incorporate the option to use an ensemble of backbone networks and network featuremaps. For this, provide the list of backbones to use (as listed in /src/anomaly_detection/ with -b <backbone and, given their ordering, denote the layers to extract with -le idx.<layer_name>. An example with three different backbones would look something like

python bin/ --gpu <gpu_id> --seed <seed> --save_patchcore_model --log_group <log_name> --log_online --log_project <log_project> results \

patch_core -b wideresnet101 -b resnext101 -b densenet201 -le 0.layer2 -le 0.layer3 -le 1.layer2 -le 1.layer3 -le 2.features.denseblock2 -le 2.features.denseblock3 --faiss_on_gpu \

--pretrain_embed_dimension 1024  --target_embed_dimension 384 --anomaly_scorer_num_nn 1 --patchsize 3 sampler -p 0.01 approx_greedy_coreset dataset --resize 256 --imagesize 224 "${dataset_flags[@]}" mvtec $datapath

When using --save_patchcore_model, in the case of ensembles, a respective ensemble of PatchCore parameters is stored.

Evaluating a pretrained PatchCore model

To evaluate a/our pretrained PatchCore model(s), run

python bin/ --gpu <gpu_id> --seed <seed> $savefolder \
patch_core_loader "${model_flags[@]}" --faiss_on_gpu \
dataset --resize 366 --imagesize 320 "${dataset_flags[@]}" mvtec $datapath

assuming your pretrained model locations to be contained in model_flags; one for each subdataset in dataset_flags. Results will then be stored in savefolder. Example model & dataset flags:

model_flags=('-p', 'path_to_mvtec_bottle_patchcore_model', '-p', 'path_to_mvtec_cable_patchcore_model', ...)
dataset_flags=('-d', 'bottle', '-d', 'cable', ...)

Expected performance of pretrained models

While there may be minor changes in performance due to software & hardware differences, the provided pretrained models should achieve the performances provided in their respective results.csv-files. The mean performance (particularly for the baseline WR50 as well as the larger Ensemble model) should look something like:

Model Mean AUROC Mean Seg. AUROC Mean PRO
WR50-baseline 99.2% 98.1% 94.4%
Ensemble 99.6% 98.2% 94.9%


If you use the code in this repository, please cite

      title={Towards Total Recall in Industrial Anomaly Detection},
      author={Karsten Roth and Latha Pemula and Joaquin Zepeda and Bernhard Schölkopf and Thomas Brox and Peter Gehler},


See CONTRIBUTING for more information.


This project is licensed under the Apache-2.0 License.