README.md

---
license: cc-by-4.0
library_name: pytorch
inference: true
tags:
- image-to-image
- style-transfer
- cyclegan
- vision-transformer
- mvit
- sketch-synthesis
- hand-drawn
---

# MViT-Stroke-CycleGAN - Pretrained Weights

Pretrained weights for **MViT-Stroke-CycleGAN**, a CycleGAN variant using **Mobile Vision Transformer (MViT)** for hand-drawn stroke synthesis.

> 🔗 Code: [GitHub Repository](https://github.com/HJ-Peng/MViT-Stroke-GAN)

---

## 📦 Checkpoint

| File | Description |
|------|-------------|
| `MViT-Stroke-GAN.pth` | Trained model weights. |

---

## ⚠️ Inference: Use `train()` Mode Only

Do **not** use `model.eval()` — it causes severe artifacts due to MViT's design.

✅ Correct:
```python
# Load the model and weights
model = MViTCycleGANModel(3, 3)  # or your model class
model.netG_A.load_state_dict(torch.load("Stroke_MViTGAN.pth", map_location="cpu")["netG_A"])

model.train()  # Keep training mode

def disable_dropout(m):
    if isinstance(m, torch.nn.Dropout):
        m.eval()
    if isinstance(m, torch.nn.Dropout2d):
        m.eval()

model.apply(disable_dropout)

with torch.no_grad():
    output = model.netG_A(input_tensor)
```

---

## 📦 Installation

```bash
pip install torch torchvision timm
```

---

## 📄 Citation

If you use this model or weights in your work, please cite:

```bibtex
@misc{mvit_stroke_cyclegan_2025,
  author = {},
  title = {},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://github.com/HJ-Peng/MViT-Stroke-GAN}}
}
```