- Docker >= 19.03
- NVIDIA Container Toolkit
Build a docker image from Dockerfile
make docker-build
See this page.
Launch a docker container and then run a training command in this container
make train-model
The trained model is available at google drive. The validation score (F1) of this trained model is 60.3. To get the validation scores with this trained model, you can specify the directory as follows:
# download zip file from my google drive.
unzip trained_model.zip -d output
python main.py --eval_dir data/iFashion/img/validation --eval_file data/iFashion/json/tweak/validation.json --evaluation --checkpoint output/adabound_wd1e-06 --output_dir output/adabound_wd1e-06
This implementation is not intended to reproduce the results reported in Guo et al.'s paper. There are some differences between the configuration in the paper and this implementation.
reported | this repo. | |
---|---|---|
base model | Inception-{V1,BN,V3}, ResNet101 | ResNet152 |
loss | weighted BCE | unweighted BCE |
optimizer | RMSProp | Adabound |