README.md

metadata

title: IPAD VAD Training
emoji: 🏭
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit
tags:
  - video-anomaly-detection
  - industrial-ai
  - computer-vision
  - pytorch
  - swin-transformer
duplicated_from: null
short_description: Train IPAD video anomaly detection models on ZeroGPU
hardware: zero-a100

🏭 IPAD: Industrial Process Anomaly Detection Training

Train state-of-the-art video anomaly detection models on industrial datasets using ZeroGPU.

Features

• 🚀 ZeroGPU Training: Free access to NVIDIA H200 GPUs
• 📊 12 Industrial Devices: Synthetic training data for various equipment
• 🎯 Baseline Model: Video Swin Transformer + Periodic Memory Module
• 💾 Automatic Checkpointing: Save and upload trained models to HF Hub
• 📈 Real-time Monitoring: Track training progress with Gradio interface

Quick Start

1. Download Dataset: Click "Download Dataset" to fetch the 8.3GB IPAD dataset from HF Hub
2. Quick Test: Verify GPU access with the quick test (no dataset required)
3. Train Baseline: Select a device and start training with default hyperparameters

Dataset

The IPAD dataset contains:

• 16 Industrial Devices (12 synthetic + 4 real)
• 597,979 Video Frames
• 39 Anomaly Classes
• Frame-level Annotations

View Dataset →

Model Architecture

Input (16 frames, 256×256)
    ↓
Video Swin Transformer Encoder
    ↓
Periodic Memory Module (2000-dim)
    ↓
I3D Decoder
    ↓
Reconstructed Output

Baseline Results

Device	AUC (%)	Device	AUC (%)
S01	69.5	S07	60.6
S02	63.9	S08	85.6
S03	70.6	S09	71.2
S04	58.3	S10	62.2
S05	86.2	S11	60.9
S06	61.2	S12	67.1
Avg	68.6

Citation

@article{liu2024ipad,
  title={IPAD: Industrial Process Anomaly Detection Dataset},
  author={Liu, Jinfan and Yan, Yichao and Li, Junjie and others},
  journal={arXiv preprint arXiv:2404.15033},
  year={2024}
}

Links

• Paper
• Original Code
• Dataset