peternara/1st-place-solution-in-google-universal-image-embedding

1st Place Solution in Google Universal Image Embedding

1st Solution in Google Universal Image Embedding

Shihao Shao[[email protected]]

Qinghua Cui[[email protected]]

1st place code in Google Universal Image Embedding is now avaliable!🔥🔥🔥

If you have any further question, Please kindly e-mail us.

The detailed solution can be found here.

Data preparation

Please download the following datasets:

Products-10K
Shopee
MET Artwork Dataset
Alibaba goods
H&M Personalized Fashion
GPR1200
GLDv2-Full
DeepFashion - Consumer-to-shop Clothes Retrieval Benchmark part

Place them into ROOT_PATH/DATA, and run:

python datasets_processing.py

The preprocessed datasets will present in ROOT_PATH/data_preprocess

Pretrained weights preparation

Put vit-h-14-laion2b_s32b_b79k.pth into the pretrained_weights/

Training

Install open_clip_280,

cd open_clip_280
pip install -e .

Run the following commands in order:

python -m torch.distributed.run --nproc_per_node=4 --master_addr 127.0.1.0 --master_port 10000 training*.py

where * should end with s*.pth, please run from s1 to s13.

Please note that when you finish s9, run:

cd open_clip_280_overlap
pip install -e .

then run the rest of s*.