fashion-clip-b32

Fashion CLIP (ViT-B-32)

This model is a fine-tuned version of OpenAI's ViT-B-32 on an e-commerce fashion catalog containing 14,833 products across 23 categories for Visual Product Search, "Shop the Look", and Cross-Modal Semantic Retrieval.

Model Details

• Base Architecture: ViT-B-32 (Vision Transformer Base, 32x32 patch size)
• Pretrained Weights: openai
• Embedding Dimension: 512
• Image Input Resolution: 224 × 224
• Framework: OpenCLIP / PyTorch
• Dataset Size: 14,833 image-text pairs across 23 fashion categories

Evaluation Performance

• Recall@5 (Top-5 Retrieval): ~85.0%
• Recall@1 (Exact Match): ~55.5%
• Validation Loss: 0.1146

Supported Categories

Casual Shirts, Formal Shirts, Formal Pants, Jeans, Men Cargos, Printed Hoodies, Printed T-shirts, Solid T-shirts, Blazers, Jackets, Denim Jackets, Sports Jackets, Long Pants, Short Pants, Dresses (Gaun), Sweaters, Polo Shirts, Skirts (Rok), Coats (Mantel), etc.

How to Use with OpenCLIP

import torch
import open_clip
from PIL import Image

# 1. Load model and preprocessing transforms directly from Hugging Face Hub
model, _, preprocess = open_clip.create_model_and_transforms('hf-hub:Panavath/fashion-clip-b32')
tokenizer = open_clip.get_tokenizer('ViT-B-32')

model.eval()

# 2. Embed an image
image = preprocess(Image.open('path_to_clothing.jpg')).unsqueeze(0)
with torch.no_grad():
    image_features = model.encode_image(image)
    image_features /= image_features.norm(dim=-1, keepdim=True)

# 3. Embed a text query
text = tokenizer(['black tactical cargo pants', 'formal striped button down shirt'])
with torch.no_grad():
    text_features = model.encode_text(text)
    text_features /= text_features.norm(dim=-1, keepdim=True)

# 4. Compute cosine similarity
similarity = (image_features @ text_features.T).squeeze(0)
print('Similarity scores:', similarity)