This project contains all source files, as well as results/logs, for our submission. All sample commands below should work on Linux (including WSL2 on Windows) or Mac.

1. create a subfolder data and put train/val/test image folders under data folder.

$ tree -L 1 data
data
├── test
├── train
└── validation

2. install all depedencies:

$ pip install -r requirements.txt

3. preprocess images with the following command:

• preprocessed results are stored as data/timm_preprocessed_*.npy files.

python preprocess_timm_v2.py

4. transform data with the following command:

• transformed data results are stored as data/distance_timm_preprocessed_*.npy files.

python transform_data.py

5. train the model with the following command.

• model checkpoints for all epochs are stored under subfolders data/checkpoints/.

$ python train_v9.py

6. to reproduce results sumbitted to the public leaderboard,

• run the following command which will load the checkpoints from step 4) to generate results.
• validation results are stored in data/validation.txt file which is compressed into data/result.zip.

python validate-timm.py

7. to reproduce results sumbitted to the private leaderboard,

• run the following command which will load the checkpoints from step 3) to generate results.
• test results are stored in data/test.txt file which is compressed into data/test-result.zip.

python validate-timm.py -n test