JJJHHHH/CCR_EthicalSplit_Finetune

Model Card: Chinese Calligraphy Character Classifier (ResNet50-based)

Model Details

• Architecture: ResNet50 pretrained on ImageNet + custom classifier head
• Classes: 1200 Chinese calligraphy characters
• Input: 224x224 RGB images (grayscale converted to RGB)
• Framework: PyTorch

Intended Use

• Handwritten Chinese calligraphy OCR and recognition
• For research, cultural preservation, and academic purposes

Dataset

• EthicalSplit5508v3
• Train: 60,168 images | Val: 1,200 | Test: 1,200
• 1200 classes with fixed splits

Training

• Batch size: 64, Learning rate: 3e-5 with OneCycleLR scheduler
• Epochs: up to 50, early stopping enabled
• Optimizer: Adam with weight decay 1e-4
• Loss: Cross-entropy with label smoothing (0.1)

Performance

• Validation loss reduced from ~5.7 to ~1.06
• Test accuracy: ~88%+
• Model size: ~25M parameters

Limitations

• May underperform on unseen handwriting styles or poor image quality
• Uses RGB input; grayscale-specific training not applied
• Dataset biases may affect generalization

Ethical Considerations

• Dataset complies with ethical usage; no PII involved
• Intended for cultural and academic use only

Usage Example

model = ChineseClassifier(embed_dim=512, num_classes=1200, pretrainedEncoder=True, unfreezeEncoder=True)
checkpoint = torch.load("best_checkpoint.pth", map_location=device)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

transform = CalligraphyCharacterDataset.defaultTransform()
img = Image.open("path_to_image.jpg").convert("RGB")
input_tensor = transform(img).unsqueeze(0).to(device)

outputs = model(input_tensor)
pred_idx = torch.argmax(outputs, dim=1).item()
pred_char = idx2char[pred_idx]