dino-vit-small-custom

DINO ViT-Small Custom Dataset

This model is a Vision Transformer (ViT) Small model trained using DINO (self-DIstillation with NO labels) on a custom dataset.

Model Details

• Architecture: ViT-Small (patch size 16)
• Pre-training Method: DINO
• Training Epochs: 500
• Output Dimension: 384
• Dataset Size: ~3000 images
• Base Model: WinKawaks/vit-small-patch16-224

Training Configuration

• Batch Size: 32
• Learning Rate: 0.005
• Teacher Temperature: 0.07
• Local Crops: 4
• Weight Decay: 0.04 → 0.4
• Optimizer: adamw

Training Results

• Final Loss: 2.9609
• Training Time: 5:16:13

Usage

from transformers import ViTModel
import torch

# Load the model
model = ViTModel.from_pretrained("odinson/dino-vit-small-custom")

# Use for feature extraction
model.eval()
with torch.no_grad():
    features = model(images).last_hidden_state
Training Curves
See the training plots in the repository for loss, learning rate, and weight decay curves.