thangkt/visionqc-efficientad-medium-pcb1

VisionQC EfficientAD-medium for VisA/pcb1

EfficientAD-medium anomaly detection checkpoint trained with Anomalib 2.6.0 on the pcb1 category of VisA.

Results

The reported test split is independent from the validation split used for early stopping and checkpoint selection.

Metric	Value
Validation image AUROC	0.8936
Test image AUROC	0.9364
Test image F1	0.8785
Test pixel AUROC	0.9883
Test pixel F1	0.6095
Test pixel AUPRO	0.8686

Training configuration

• Model: EfficientAD-medium
• Dataset: VisA/pcb1
• Training images: 904 normal
• Validation: 100 images
• Test: 50 normal and 50 anomalous images
• Batch size: 32 per GPU
• Devices: 2 Tesla T4 GPUs
• Effective batch size: 64
• Precision: FP16 mixed precision
• Early stopping: validation image AUROC, patience 20, minimum delta 0.001
• Best epoch: 30
• Best global step: 465

This uses a local BatchedEfficientAd wrapper to permit batched DDP training without modifying Anomalib core. The architecture itself remains EfficientAD-medium. Results are not directly comparable to the official batch-size-1 EfficientAD baseline.

Files

• model-best.ckpt: best Lightning checkpoint selected by validation image AUROC
• config.yaml: complete VisionQC training configuration
• metrics.json: test metrics generated after restoring the best checkpoint

Loading

Download the checkpoint:

hf download thangkt/visionqc-efficientad-medium-pcb1 model-best.ckpt \
  --local-dir weights/efficientad-medium-pcb1

Use it with the VisionQC project and Anomalib 2.6.0:

visionqc evaluate \
  --checkpoint weights/efficientad-medium-pcb1/model-best.ckpt

Links

• Source: https://github.com/pnthang04/VisionQC
• Multi-GPU training PR: https://github.com/pnthang04/VisionQC/pull/1
• Model repository: https://huggingface.co/thangkt/visionqc-efficientad-medium-pcb1