Training a Single Model

Overview

Training a single-backbone PatchCore model is the fastest way to get started. This guide uses WideResNet50 as the backbone, which provides excellent performance while remaining computationally efficient.

Recommended for:

First-time users learning PatchCore
Development and experimentation
Production systems with speed/memory constraints
Achieving 99.2% AUROC on MVTec AD

Quick Start

Train PatchCore on all MVTec AD categories with recommended settings:

Complete Training Command

# Set dataset path
datapath=/path/to/mvtec

# Define all 15 MVTec categories
datasets=('bottle' 'cable' 'capsule' 'carpet' 'grid' 'hazelnut' 
          'leather' 'metal_nut' 'pill' 'screw' 'tile' 'toothbrush' 
          'transistor' 'wood' 'zipper')

# Create dataset flags
dataset_flags=($(for dataset in "${datasets[@]}"; do echo '-d '$dataset; done))

# Run training
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

Don’t forget to:

Set PYTHONPATH before running: export PYTHONPATH=src
Replace /path/to/mvtec with your actual dataset path
Ensure MVTec AD is properly set up (see MVTec Setup)

Command Structure

The PatchCore training command has four main sections:

Global Arguments
Backbone Config
Sampler Config
Dataset Config

Global Settings

python bin/run_patchcore.py \
  --gpu 0                    # GPU device ID (use 0 for first GPU)
  --seed 0                   # Random seed for reproducibility
  --save_patchcore_model     # Save trained model to disk
  --log_group experiment_name # Name for this experiment
  --log_project project_name  # Project folder name
  results                    # Output directory

Key Parameters:

--gpu: Which GPU to use (supports multi-GPU with --gpu 0 --gpu 1)
--seed: Controls random initialization for reproducibility
--save_patchcore_model: Must include to save the model for inference
results: Base directory where outputs are stored

PatchCore Configuration

patch_core \
  -b wideresnet50           # Backbone network architecture
  -le layer2                # Extract from layer2
  -le layer3                # Extract from layer3
  --faiss_on_gpu            # Use GPU for nearest neighbor search
  --pretrain_embed_dimension 1024  # Backbone feature dimension
  --target_embed_dimension 1024    # Final embedding dimension
  --anomaly_scorer_num_nn 1        # Number of nearest neighbors
  --patchsize 3                    # Local aggregation size

Critical Choices:

Backbone (-b): Network pretrained on ImageNet
Layers (-le): Which layers to extract features from
Dimensions: Balance between memory and performance

Coreset Sampling

sampler \
  -p 0.1                    # Sample 10% of features
  approx_greedy_coreset     # Sampling algorithm

Sampling Methods:

approx_greedy_coreset: Fast approximate coreset (recommended)
greedy_coreset: Exact coreset (slower, slightly better)
identity: No sampling (uses all features, high memory)

Percentage (-p):

0.1 (10%): Good for development, faster training
0.01 (1%): Production setting, minimal accuracy loss

Dataset Configuration

dataset \
  --resize 256              # Initial resize dimension
  --imagesize 224           # Final crop size (input to backbone)
  -d bottle                 # Include bottle category
  -d cable                  # Include cable category
  [...more categories...]   # Add more with -d
  mvtec                     # Dataset type
  $datapath                 # Path to dataset root

Image Processing:

Load image → 2. Resize to 256×256 → 3. Center crop to 224×224

Categories: Specify each with -d category_name

Step-by-Step Training

Let’s train on a single category first, then scale to all 15:

Set Environment

Setup

# Navigate to repository
cd /path/to/patchcore

# Set Python path
export PYTHONPATH=src

# Verify dataset
ls /path/to/mvtec/bottle/train/good

Train on Single Category

Start with one category (bottle) to verify setup:

Single Category Training

datapath=/path/to/mvtec

python bin/run_patchcore.py \
  --gpu 0 --seed 0 --save_patchcore_model \
  --log_group WR50_bottle_test \
  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

Expected output:

Computing support features...
Subsampling...
Inferring...
instance_auroc: 1.000
full_pixel_auroc: 0.984
anomaly_pixel_auroc: 0.984

Scale to All Categories

Once verified, train on all 15 categories:

All 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_baseline \
  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

Training will process each category sequentially. Total time: ~1-2 hours on GPU.

Check Results

After training completes:

View Results

# Navigate to results directory
cd results/MVTecAD_Results/IM224_WR50_baseline_0

# View summary metrics
cat results.csv

# Check saved models
ls -lh models/

Each category will have a dedicated model directory.

Understanding Output Files

After training with --save_patchcore_model, you’ll see:

Output Structure

results/
└── MVTecAD_Results/
    └── IM224_WR50_baseline_0/
        ├── models/
        │   ├── mvtec_bottle/
        │   │   ├── nnscorer_search_index.faiss  # Nearest neighbor index
        │   │   └── patchcore_params.pkl         # Model parameters
        │   ├── mvtec_cable/
        │   ├── mvtec_capsule/
        │   └── [...]
        └── results.csv                           # Performance summary

What's in each file?

nnscorer_search_index.faiss:

FAISS nearest-neighbor search index
Contains memory bank of training features
Size: 50-500 MB depending on sampling percentage

patchcore_params.pkl:

Model configuration (backbone, layers, dimensions)
Small file (~1 KB)
Needed to load model for inference

results.csv:

Performance metrics for each category
Columns: dataset_name, instance_auroc, full_pixel_auroc, anomaly_pixel_auroc

All CLI Arguments Explained

Here’s every argument used in the training command:

int

default:"0"

GPU device ID. Use --gpu 0 for first GPU, --gpu 1 for second, etc.

int

default:"0"

Random seed for reproducibility. Same seed = same results.

flag

Save trained model to disk. Required if you want to use the model later.

string

Experiment name. Creates a subdirectory with this name in results.

string

default:"project"

Project folder name. Organizes experiments into projects.

string

required

Backbone network architecture. Options:

wideresnet50 (recommended)
wideresnet101
resnet50, resnet101
See Backbones for all options

string

required

Which layers to extract features from. Common choices:

WideResNet: layer2, layer3
ResNet: layer2, layer3
DenseNet: features.denseblock2, features.denseblock3

int

default:"1024"

Dimensionality of features from backbone layers. Standard: 1024.

int

default:"1024"

Final embedding dimension after aggregation. Lower = less memory.

int

default:"3"

Neighborhood size for local aggregation. Typical: 3 or 5.

int

default:"1"

Number of nearest neighbors for anomaly scoring. Usually 1, 3, or 5.

flag

Use GPU for FAISS nearest neighbor search. Faster but uses GPU memory.

float

default:"0.1"

Coreset sampling percentage (0.01 = 1%, 0.1 = 10%).

int

default:"256"

Initial resize dimension before center crop.

int

default:"224"

Final image size after center crop (input to backbone).

string

required

MVTec category names. Repeat for each category: -d bottle -d cable ...

Configuration Presets

Try these proven configurations for different use cases:

# Quick training for testing (~5-10 min/category)
python bin/run_patchcore.py --gpu 0 --seed 0 --save_patchcore_model \
  --log_group dev_fast 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

# Expected: ~99% AUROC, 1-2 GB model size

# Baseline production config from paper
# Instance AUROC: 99.2%, Pixel AUROC: 98.1%, PRO: 94.4%
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 \
  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 0.01 approx_greedy_coreset \
dataset --resize 256 --imagesize 224 "${dataset_flags[@]}" mvtec $datapath

# Better localization, higher resolution
# Instance AUROC: 99.3%, Pixel AUROC: 97.8%, PRO: 94.3%
python bin/run_patchcore.py --gpu 0 --seed 22 --save_patchcore_model \
  --log_group IM320_WR50_L2-3_P001_D1024-1024_PS-3_AN-1_S22 \
  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 0.01 approx_greedy_coreset \
dataset --resize 366 --imagesize 320 "${dataset_flags[@]}" mvtec $datapath

# Optimized for pixel-level segmentation
# Instance AUROC: 99.0%, Pixel AUROC: 98.4%, PRO: 94.2%
python bin/run_patchcore.py --gpu 0 --seed 39 --save_patchcore_model \
  --log_group IM320_WR50_L2-3_P001_D1024-1024_PS-5_AN-3_S39 \
  results \
patch_core -b wideresnet50 -le layer2 -le layer3 --faiss_on_gpu \
  --pretrain_embed_dimension 1024 --target_embed_dimension 1024 \
  --anomaly_scorer_num_nn 3 --patchsize 5 \
sampler -p 0.01 approx_greedy_coreset \
dataset --resize 366 --imagesize 320 "${dataset_flags[@]}" mvtec $datapath

Monitoring Training Progress

During training, you’ll see progress output:

Training Output

Command line arguments: bin/run_patchcore.py --gpu 0 ...
Evaluating dataset [mvtec_bottle] (1/15)...
Computing support features...: 100%|████████████| 209/209 [00:45<00:00]
Subsampling...: 100%|████████████████████████| 1679/1679 [01:23<00:00]
Inferring...: 100%|██████████████████████████| 63/63 [00:12<00:00]
Computing evaluation metrics.
instance_auroc: 1.000
full_pixel_auroc: 0.984
anomaly_pixel_auroc: 0.984
Saving PatchCore data.

-----

Evaluating dataset [mvtec_cable] (2/15)...
[...continues for each category...]

Troubleshooting

CUDA out of memory error

Problem: GPU doesn’t have enough memory.Solutions:

Reduce batch size: dataset [...] --batch_size 1 mvtec $datapath
Lower image resolution: Use 224 instead of 320
Reduce target embedding dimension: --target_embed_dimension 512
Use CPU for FAISS: Remove --faiss_on_gpu

ModuleNotFoundError: No module named 'patchcore'

Problem: Python can’t find the patchcore module.Solution:

export PYTHONPATH=src
# Or run with:
env PYTHONPATH=src python bin/run_patchcore.py [...]

Training very slow

Problem: Training takes much longer than expected.Causes & Solutions:

Not using GPU: Verify with nvidia-smi, add --faiss_on_gpu
High sampling percentage: Use -p 0.01 instead of -p 0.1
Too many workers: Reduce --num_workers 4 in dataset config

Next Steps

Ensemble Models

Combine multiple backbones for 99.6% AUROC

Configuration Reference

Deep dive into every parameter

Get Started

Core Concepts

Training

Inference

Model Zoo

Overview

Quick Start

Command Structure

Step-by-Step Training

Understanding Output Files

All CLI Arguments Explained

Configuration Presets

Monitoring Training Progress

Troubleshooting

Next Steps

Ensemble Models

Configuration Reference

​Overview

​Quick Start

​Command Structure

​Step-by-Step Training

​Understanding Output Files

​All CLI Arguments Explained

​Configuration Presets

​Monitoring Training Progress

​Troubleshooting

​Next Steps

Ensemble Models

Configuration Reference

Overview

Quick Start

Command Structure

Step-by-Step Training

Understanding Output Files

All CLI Arguments Explained

Configuration Presets

Monitoring Training Progress

Troubleshooting

Next Steps