25th solution for the Kagge Competition https://www.kaggle.com/c/humpback-whale-identification
Look at requirements.sh. There are listed all requirements. After creating new conda evv, you can just run sh requirements.sh
Download data to data folder. Then apply script data_cropping.py for train and test dataset.
For training first of all change path in main file. There are two variable to change: train_df and train_folder.
You can also change batch-size (default = 16, which need 2x 11 GB).
You can then just run
python main_cosine_bb_final.py <folder_name>
There would be saved two checkpoints. The last one and with best-acc. As this script learn on entire dataset, I'm using last one (checkpoint_base).
Here the code is for only 'without-new-whale'. To create the submission, firstly change folder to data in the file.
Then
eval_bb_se.py <path_to_checkpoint> <encoder_npy_file> <X_test>
For a 'new-whale', I have just found threshold which gave 28%-29%.