Skip to main content

​
What is PatchCore Training?

PatchCore doesn’t use traditional neural network training with backpropagation. Instead, it extracts and stores a memory bank of feature representations from normal (defect-free) training images.
Training in PatchCore refers to:
  1. Extracting features from training images using a pretrained CNN backbone
  2. Performing local aggregation on patch features
  3. Subsampling the feature set using coreset selection
  4. Building a nearest-neighbor search index for anomaly detection

​
Training Workflow

1

Prepare MVTec AD Dataset

Download and organize the MVTec AD dataset with proper directory structure.See MVTec Setup for detailed instructions.
2

Choose Model Configuration

Select backbone network, layers to extract features from, and sampling parameters.
  • Single model: Use one backbone (e.g., WideResNet50) for faster training
  • Ensemble: Use multiple backbones for better performance
3

Run Training

Execute the training command. PatchCore will:
  • Load training images (only “good” samples)
  • Extract features from selected backbone layers
  • Apply coreset subsampling to reduce memory
  • Build FAISS nearest-neighbor index
4

Save Model

Use --save_patchcore_model flag to save:
  • Feature memory bank
  • FAISS search index
  • Model parameters (backbone, layers, dimensions)

​
Performance Characteristics

​
Training Time

PatchCore training is fast compared to traditional deep learning:
  • Single model (WideResNet50): ~5-10 minutes per MVTec category on GPU
  • Ensemble (3 backbones): ~15-30 minutes per category on GPU
  • Time scales with:
    • Image resolution (224x224 vs 320x320)
    • Number of training samples
    • Coreset sampling percentage

​
Memory Requirements

GPU Memory

  • 11GB recommended for most experiments
  • 16GB required for:
    • Large image sizes (320x320 or higher)
    • Ensemble models with 3+ backbones
    • Higher embedding dimensions

Disk Storage

  • Per model: 50-500 MB depending on:
    • Coreset sampling percentage (1% vs 10%)
    • Embedding dimensions
    • Number of training samples
  • 15 MVTec categories: 1-7 GB total

​
Hardware Recommendations

​
Minimum Requirements

Minimum Configuration
GPU: NVIDIA GPU with 11GB VRAM (e.g., RTX 2080 Ti, RTX 3060)
CPU: 4+ cores
RAM: 16 GB
Storage: 50 GB (dataset + models)
Recommended Configuration
GPU: NVIDIA GPU with 16GB+ VRAM (e.g., RTX 3090, A5000, V100)
CPU: 8+ cores
RAM: 32 GB
Storage: 100 GB SSD
Significantly large input images (>512x512) will require more GPU memory and may not fit on 11GB GPUs.

​
Key Configuration Decisions

​
Backbone Selection

The choice of backbone affects both performance and speed:
BackbonePerformanceSpeedMemory
WideResNet50GoodFastModerate
WideResNet101BetterSlowerHigher
Ensemble (3 models)BestSlowestHighest

​
Coreset Sampling Percentage

Controls the trade-off between memory usage and performance:
  • 10% (-p 0.1): Good for development and testing
  • 1% (-p 0.01): Recommended for production (minimal performance loss)
  • Lower percentages = less memory, faster inference, minimal accuracy impact
In the original paper, 1% coreset sampling achieved 99.2% AUROC on MVTec AD with WideResNet50.

​
Image Resolution

Higher resolution improves localization but increases compute:
  • 224x224: Baseline resolution, fastest training
  • 320x320: Better localization, ~1.5x slower
  • Higher: Possible but requires more GPU memory

​
Expected Results

Using the recommended configurations from single-model and ensemble-models:
Performance Metrics
Instance AUROC: 99.2%
Pixel-wise AUROC: 98.1%
PRO Score: 94.4%

Training time: ~1-2 hours (all 15 categories)
GPU memory: 8-10 GB
Model size: 1-2 GB total
Performance Metrics
Instance AUROC: 99.3%
Pixel-wise AUROC: 98.1%
PRO Score: 94.2%

Training time: ~3-5 hours (all 15 categories)
GPU memory: 10-12 GB
Model size: 4-7 GB total
Performance Metrics
Instance AUROC: 99.6%
Pixel-wise AUROC: 98.2%
PRO Score: 94.9%

Training time: ~5-8 hours (all 15 categories)
GPU memory: 12-15 GB
Model size: 5-10 GB total

​
Next Steps

Setup MVTec Dataset

Download and organize the MVTec AD benchmark dataset

Train Single Model

Train your first PatchCore model with WideResNet50

Configuration Guide

Deep dive into all training parameters

Ensemble Models

Combine multiple backbones for maximum performance

Build docs developers (and LLMs) love

Get started for freeTalk to us