Skip to main content

​
Overview

This repository provides pretrained PatchCore models that achieve state-of-the-art performance on MVTec AD, with up to 99.6% image-level AUROC, 98.4% pixel-level AUROC, and >95% PRO score. Each model is trained on one of the 15 MVTec AD subdatasets and can be directly loaded for inference without retraining.

​
Available Model Configurations

WideResNet50 Baseline

Standard single-backbone model optimized for 224x224 images
  • Image AUROC: 99.2%
  • Pixel AUROC: 98.1%
  • PRO Score: 94.4%

Ensemble Model

Multi-backbone ensemble with WR101/ResNext101/DenseNet201
  • Image AUROC: 99.6%
  • Pixel AUROC: 98.2%
  • PRO Score: 94.9%

​
Model Structure

Pretrained models are organized in the following directory structure: 
models/
├── IM224_WR50_L2-3_P01_D1024-1024_PS-3_AN-1/
│   └── models/
│       ├── mvtec_bottle/
│       │   ├── nnscorer_search_index.faiss
│       │   └── patchcore_params.pkl
│       ├── mvtec_cable/
│       ├── mvtec_capsule/
│       └── ... (15 categories total)
├── IM224_Ensemble_L2-3_P001_D1024-384_PS-3_AN-1/
│   └── models/
│       └── ... (same 15 categories)

​
Model Files

Each category model contains two files:
  • nnscorer_search_index.faiss - FAISS index for nearest neighbor search
  • patchcore_params.pkl - Serialized PatchCore parameters and memory bank

​
MVTec AD Categories

Pretrained models are available for all 15 MVTec AD object and texture categories: Object Categories:
  • bottle
  • cable
  • capsule
  • hazelnut
  • metal_nut
  • pill
  • screw
  • toothbrush
  • transistor
  • zipper
Texture Categories:
  • carpet
  • grid
  • leather
  • tile
  • wood

​
Loading Pretrained Models

​
Evaluation Script

Use the provided evaluation script to load and test pretrained models: 
datapath=/path/to/mvtec/data
loadpath=/path/to/pretrained/models
modelfolder=IM224_WR50_L2-3_P001_D1024-1024_PS-3_AN-1
savefolder=evaluated_results/$modelfolder

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

​
Loading Individual Models

To load a single category model: 
python bin/load_and_evaluate_patchcore.py --gpu 0 --seed 0 results \
  patch_core_loader -p models/IM224_WR50_L2-3_P01_D1024-1024_PS-3_AN-1/models/mvtec_bottle \
  --faiss_on_gpu \
  dataset --resize 366 --imagesize 320 -d bottle mvtec /path/to/mvtec

​
Model Configurations

​
WideResNet50 Baseline (IM224)

Configuration:
  • Backbone: Wide ResNet-50
  • Feature layers: layer2, layer3
  • Input size: 224x224 (resized from 256x256)
  • Coreset sampling: 10%
  • Embedding dimensions: 1024 → 1024
  • Patch size: 3
  • Nearest neighbors: 1
Model ID: IM224_WR50_L2-3_P01_D1024-1024_PS-3_AN-1

​
Ensemble Model (IM224)

Configuration:
  • Backbones: Wide ResNet-101, ResNeXt-101, DenseNet-201
  • Feature layers:
    • WR101: layer2, layer3
    • ResNeXt101: layer2, layer3
    • DenseNet201: denseblock2, denseblock3
  • Input size: 224x224
  • Coreset sampling: 1%
  • Embedding dimensions: 1024 → 384
  • Patch size: 3
  • Nearest neighbors: 1
Model ID: IM224_Ensemble_L2-3_P001_D1024-384_PS-3_AN-1

​
Higher Resolution Models (IM320)

Models trained on 320x320 images achieve similar or better performance:
  • WR50 (IM320): 99.3% Image AUROC, 97.8% Pixel AUROC
  • Ensemble (IM320): 99.6% Image AUROC, 98.2% Pixel AUROC
Higher resolution models (320x320) require more GPU memory but can provide better localization accuracy.

​
Inference Requirements

​
Hardware

  • GPU Memory: 11GB recommended (varies with image size)
  • GPU: CUDA-compatible GPU for FAISS acceleration
  • CPU: Any modern multi-core processor

​
Software Dependencies

pip install torch torchvision
pip install faiss-gpu  # or faiss-cpu
pip install timm
pip install scikit-learn
pip install Pillow
See requirements.txt for complete dependency list.

​
Model Performance

For detailed per-category performance metrics, see the Performance Benchmarks page.

​
Training Custom Models

To train your own PatchCore model with the same configuration: 
python bin/run_patchcore.py --gpu 0 --seed 0 --save_patchcore_model \
  --log_group IM224_WR50_L2-3_P01_D1024-1024_PS-3_AN-1_S0 \
  --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 -d bottle mvtec /path/to/mvtec
See Training Guide for complete training instructions.

​
Output Format

After evaluation, results are saved as: 
evaluated_results/
├── results.csv           # Performance metrics for all categories
└── segmentation_images/  # Visual anomaly maps (if enabled)
The results.csv contains:
  • instance_auroc - Image-level AUROC
  • full_pixel_auroc - Pixel-level AUROC
  • full_pro - PRO score
  • anomaly_pixel_auroc - AUROC on anomalous pixels only
  • anomaly_pro - PRO on anomalous regions only
All pretrained models use ImageNet-pretrained backbone weights for feature extraction.

Build docs developers (and LLMs) love

Get started for freeTalk to us