Skip to content

troubleshooting.md

Latest commit

 

History

History
165 lines (100 loc) · 3.38 KB

troubleshooting.md

File metadata and controls

165 lines (100 loc) · 3.38 KB

🛠️ Troubleshooting & FAQ

This guide covers common issues when using AnomaVision and how to fix them.

1. Installation Issues

❌ Problem: CUDA mismatch (torch not compiled with CUDA enabled)

Cause: Installed PyTorch version doesn’t match your CUDA toolkit. Fix: Reinstall PyTorch with the correct CUDA version:

# Example for CUDA 12.1
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Or use Poetry:

poetry install --extras "cu121"

❌ Problem: onnxruntime not found

Cause: ONNX Runtime not installed. Fix:

pip install onnxruntime-gpu onnxruntime-tools

❌ Problem: OpenVINO not working on Windows

Cause: Missing Intel OpenVINO dependencies. Fix:

pip install openvino

Check OpenVINO installation guide if errors persist.

2. Dataset Issues

❌ Problem: FileNotFoundError: train/good not found

Cause: Dataset not in MVTec structure. Expected Layout:

dataset/
└── bottle/
    ├── train/good/
    └── test/broken_large/

Fix: Reorganize folders into the correct structure.

❌ Problem: Some images are ignored

Cause: Non-image files (e.g., .txt, .DS_Store) in dataset. Fix: Remove or filter out invalid files.

3. Training Issues

❌ Problem: Training is very slow on CPU

Fixes:

  • Use smaller image size (resize: [128,128])
  • Reduce batch_size
  • Switch to GPU (--device cuda) if available

❌ Problem: CUDA out of memory during training

Fixes:

  • Reduce batch_size
  • Use smaller backbone (resnet18 instead of wide_resnet50)
  • Enable memory-efficient mode (where supported)

4. Inference Issues

❌ Problem: RuntimeError: Input size mismatch

Cause: Model expects a fixed input size. Fix: Resize input images to the same size used during training (resize / crop_size).

❌ Problem: No anomalies detected (all scores low)

Causes & Fixes:

  • Threshold too high → lower --thresh
  • Wrong normalization → ensure dataset uses same mean/std as training
  • Wrong class name → check --class_name or config.yml

5. Export Issues

❌ Problem: Unsupported operator during ONNX export

Fixes:

  • Upgrade to latest PyTorch and ONNX
  • Try lower opset (e.g., --opset 16)
  • If still failing → use TorchScript export

❌ Problem: Quantized ONNX too large / inaccurate

Fixes:

  • Use --quantize-dynamic for lightweight INT8
  • If accuracy drops, stick to fp16

6. Visualization Issues

❌ Problem: Heatmaps look empty

Causes & Fixes:

  • Wrong normalization → check norm_mean & norm_std
  • viz_alpha too low → increase to 0.7
  • Threshold too high → lower --thresh

7. General FAQ

Q: Which backbones are supported? A: Currently resnet18 and wide_resnet50. More will be added in future.

Q: How to deploy on edge devices without Python? A: Use the C++ ONNX runtime provided in /docs/cpp/.

Q: Can I train on custom datasets? A: Yes, as long as the dataset follows MVTec-style folder structure.

✅ With this guide, you should be able to quickly solve most common problems. If an issue persists, please open a GitHub Issue with details.