---
license: mit
tags:
- computer-vision
- defect-detection
- industrial-inspection
- vit
- unet
- grad-cam
- knowledge-distillation
- quantization
library_name: pytorch
---

# Industrial Vision Agent — Model Weights

Trained checkpoints for an agentic multimodal quality-inspection system on metal
castings (7,347 images, binary defect classification + weakly-supervised segmentation).

| File | Architecture | Role | Verified metric |
|---|---|---|---|
| `vit_tiny_patch16_224_best.pth` | ViT-Tiny (timm), fine-tuned | Defect classifier | Test F1 0.9967 / Recall 0.9956 |
| `resnet18_best.pth` | ResNet18, linear-probed head | Grad-CAM localization backbone | Test F1 0.924 |
| `mobilenetv3_small_050_best.pth` | MobileNetV3-small distilled student | Edge classifier (94.9% fewer params) | Test F1 0.991 |
| `unet_pseudo_best.pth` | U-Net trained on Grad-CAM pseudo-masks | Segmentation (224x224 mask) | Val Dice 0.706 |
| `metrics.json` | — | Full training/test results | — |

INT8 dynamic quantization of the ViT shrinks it 71.7% (22.2 -> 6.3 MB).

## Quick load

```python
import torch, timm
from huggingface_hub import hf_hub_download

weights = hf_hub_download("kaushik3009/industrial-vision-agent", "vit_tiny_patch16_224_best.pth")
model = timm.create_model("vit_tiny_patch16_224", pretrained=False, num_classes=2)
model.load_state_dict(torch.load(weights, map_location="cpu"))
```

Labels: 0 = OK, 1 = Defective. Input: RGB, resized 224x224, ImageNet-normalized.

Full pipeline, agent code, and docs: https://github.com/kaushik-3009/industrial-vision-agent