AIMeetsBeautyChallenge

This is the source code of Perfect Half Million Beauty Product Image Recognition Challenge.

Installation

• Env: python3.6
• Package:

pytorch 1.1
h5py
faiss
pandas
tensorboardX
tqdm

Get started

1. train.py

python train.py --model-name resnet152 --dataset /data/dataset/AIMeetsBeauty --epochs 30 --batch-size 64 --lr 0.0001

2. get_feature.py

python get_feature.py --model-name resnet152 --dataset /data/dataset/AIMeetsBeauty --model-ckpt ./ckpts/resnet152.pkl --output-feature-path ./feature/resnet152.hdf5

3. predict.py

python predict.py --model-name resnet152 --model-ckpt ./ckpts/resnet152.pkl --feature-path ./feature/resnet152.hdf5 --test-path ./val_2019 --output-result-path ./result/resnet152.csv

4. evaluation.py

python evaluation.py --prediction-path ./result/predictions.csv --label-path ./val_2019.csv

5. ensemble_predict.py

Options: Using multiple models to perform ensemble predictions, which need to change the loading path of model checkpoint file and feature hdf5 file.

python ensemble_predict.py --prediction-path ./result/predictions_att.csv --label-path ./val_2019.csv