This repository contains the code used for the live conding workshop : "Vector database for an advanced search engine"
In order to have the fully functioning app, you'll need to download the H&M product pictures from the kaggle H&M challenge and place it in the "data/images" directory a the root of the cloned repo.
Each picture, containing the product_id in its name must be store in a subdirectory containing the three first digits of the product_id
It should look like this:
You can find on this repo, the 105k images embeddeg using fashion clip. You'll need to install git-lfs to be able to pull them:
data/dict_ids_embeddings_full.pickle
You can also use
python gc_db/embedding/embedder.py
It will embed all the images found in
IMAGES_PATH: str = str(ROOT_DIR / './data/images')
Clone the repository
pip install .
make run
Or if you want to use HNSW lib
make run-hnsw
You can reach me at [email protected]
- Kaggle H&M personalized recommendations https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations
- Fashion Clip model by Patrick John Chia https://github.com/patrickjohncyh/fashion-clip
- mattmdjaga's segformer model used for clothe segmentation https://github.com/mattmdjaga/segformer_b2_clothes