ShushanSS/DeepFashionRetrieval

Full Project with frontend backend

https://github.com/ShushanSS/TopK_Similar_Images_Retrieval

DeepFashion Retrieval — Fine-tuned ResNet50 Embedder

Fine-tuned ResNet50 image embedder for fashion item retrieval, trained with batch-hard triplet loss and a category-aware PK sampler on the DeepFashion In-shop Clothes Retrieval dataset.

Contents

File	Description
embedder_full_train_epoch_6.pt	Fine-tuned ResNet50 weights (epoch 6)
faiss_index.bin	FAISS index built over embedded gallery images (if included)
metadata.json / config.json	Embedding dim, class list, preprocessing params (if included)

Setup

pip install -r requirements.txt
python download_weights.py

This downloads the files above into ./weights in your project root.

Generating metadata locally

The metadata CSVs are not included in this repo since they're derived from the raw DeepFashion images, which you need to download separately.

1. Download the DeepFashion (In-shop Clothes Retrieval) high-res images.
2. Update DATASET_PATH in prepare_metadata.py to point to your local copy, preserving the original folder structure:

   <DATASET_PATH>/<gender>/<clothing_category>/<item_id>/<image>.jpg
   

3. Run:

   python prepare_metadata.py
   

   This produces original_metadata.csv, full_metadata.csv, and original_metadata_filtered.csv in your working directory.

Model details

• Backbone: ResNet50, fine-tuned with batch-hard triplet loss
• Sampler: CategoryAwarePKSampler
• Retrieval baseline (CLIP zero-shot): Hit Rate@5 = 0.546
• Fine-tuned ResNet50: Hit Rate@5 = 0.819 on controlled test set

Usage

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ShushanSS/DeepFashionRetrieval",
    local_dir="./weights",
    local_dir_use_symlinks=False,
)

License

MIT (update if different)