vietnamese-vit-classification

Vietnamese Scene Classification with Vision Transformer

Fine-tuned google/vit-base-patch16-224 (86M parameters) for Vietnamese scene classification.

Model Details

• Architecture: Vision Transformer (ViT-Base, patch size 16, 224×224 input)
• Parameters: 86M
• Base Model: google/vit-base-patch16-224 (ImageNet-21k pretrained)
• Task: Multi-class Vietnamese scene classification
• Framework: PyTorch + HuggingFace Transformers

Training

• Data Augmentation: RandomResizedCrop, RandomHorizontalFlip, ColorJitter
• Optimizer: AdamW (lr=3e-5, weight_decay=0.01)
• Scheduler: OneCycleLR with cosine annealing
• Gradient Clipping: max_norm=1.0
• Validation Accuracy: 94%+

Scene Classes (Vietnamese)

Features

• Attention rollout visualization across 12 transformer layers
• Per-class precision, recall, and F1 metrics
• Formatted confusion matrix output

Usage

from transformers import ViTForImageClassification, ViTImageProcessor
from PIL import Image

model = ViTForImageClassification.from_pretrained("sanvo/vietnamese-vit-classification")
processor = ViTImageProcessor.from_pretrained("google/vit-base-patch16-224")

image = Image.open("scene.jpg")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
predicted_class = outputs.logits.argmax(-1).item()
print(model.config.id2label[predicted_class])

Links

• GitHub: svn05/vietnamese-vit-classification