Skip to main content
This quickstart guide will help you train and evaluate your first PatchCore model on the MVTec AD dataset.

​
Prerequisites

Make sure you have:
  • Python 3.8 or higher
  • CUDA-compatible GPU (recommended)
  • At least 11GB of GPU memory

​
Quick Setup

1

Install Dependencies

Clone the repository and install required packages:
git clone https://github.com/amazon-science/patchcore-inspection.git
cd patchcore-inspection
pip install -r requirements.txt

click>= 8.0.3
faiss-cpu
matplotlib>= 3.5.0
pillow>= 8.4.0
pretrainedmodels>= 0.7.4
torch>= 1.10.0
scikit-image>= 0.18.3
scikit-learn>= 1.0.1
scipy>= 1.7.1
torchvision>= 0.11.1
tqdm>= 4.62.3
2

Download MVTec AD Dataset

Download the MVTec AD dataset from the official website and extract it.Your dataset structure should look like:
mvtec/
├── bottle/
│   ├── ground_truth/
│   ├── test/
│   │   ├── good/
│   │   ├── broken_large/
│   │   └── ...
│   └── train/
│       └── good/
├── cable/
└── ...
The dataset contains 15 categories: bottle, cable, capsule, carpet, grid, hazelnut, leather, metal_nut, pill, screw, tile, toothbrush, transistor, wood, and zipper.
3

Set Environment Variable

Set the PYTHONPATH to include the source directory:
export PYTHONPATH=src:$PYTHONPATH
4

Train Your First Model

Train PatchCore on a single category (e.g., “bottle”) using WideResNet50:
datapath=/path/to/mvtec

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 $datapath
  • --gpu 0: Use GPU 0
  • --seed 0: Set random seed for reproducibility
  • --save_patchcore_model: Save the trained model
  • -b wideresnet50: Use WideResNet50 backbone (pretrained on ImageNet)
  • -le layer2 -le layer3: Extract features from layer2 and layer3
  • --faiss_on_gpu: Run similarity search on GPU for speed
  • --pretrain_embed_dimension 1024: Backbone feature dimension
  • --target_embed_dimension 1024: Final PatchCore embedding dimension
  • --anomaly_scorer_num_nn 1: Use 1 nearest neighbor for scoring
  • --patchsize 3: Local aggregation neighborhood size
  • -p 0.1: Keep 10% of patches via coreset subsampling
  • --resize 256 --imagesize 224: Resize to 256, then center crop to 224

​
Expected Output

During training, you’ll see output similar to: 
Evaluating dataset [mvtec_bottle] (1/1)...
Training models (1/1)
Embedding test data with models (1/1)
Computing evaluation metrics.
instance_auroc: 1.000
full_pixel_auroc: 0.988
anomaly_pixel_auroc: 0.988
Expected Performance (WideResNet50 baseline on 224x224 images):
  • Instance AUROC: 99.2% (image-level anomaly detection)
  • Pixel-wise AUROC: 98.1% (defect localization)
  • PRO Score: 94.4%

​
Train on All Categories

To train on all 15 MVTec AD categories: 
datapath=/path/to/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/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 "${dataset_flags[@]}" mvtec $datapath

​
Model Output

After training, your results will be saved in: 
results/MVTecAD_Results/IM224_WR50_L2-3_P01_D1024-1024_PS-3_AN-1_S0/
├── models/
│   └── mvtec_bottle/
│       ├── nnscorer_search_index.faiss
│       └── patchcore_params.pkl
└── results.csv

​
Evaluate a Trained Model

To evaluate your trained model: 
datapath=/path/to/mvtec
loadpath=results/MVTecAD_Results/IM224_WR50_L2-3_P01_D1024-1024_PS-3_AN-1_S0

python bin/load_and_evaluate_patchcore.py --gpu 0 --seed 0 evaluated_results \
  patch_core_loader -p $loadpath/models/mvtec_bottle --faiss_on_gpu \
  dataset --resize 256 --imagesize 224 -d bottle mvtec $datapath

​
Next Steps

Installation Guide

Detailed installation instructions and troubleshooting

Model Architecture

Learn how PatchCore works under the hood

Training Options

Explore advanced training configurations

API Reference

Complete API documentation

​
Common Issues

GPU Memory Error? If you encounter out-of-memory errors:
  • Reduce --imagesize (e.g., from 224 to 192)
  • Reduce batch size with --batch_size 1
  • Use CPU FAISS by removing --faiss_on_gpu
If you get “FAISS not found”, install the GPU version:
pip uninstall faiss-cpu
pip install faiss-gpu

Build docs developers (and LLMs) love

Get started for freeTalk to us