Below you can find a outline of how to reproduce my solution for the Humpback Whale Identification competition. If you run into any trouble with the setup/code or have any questions please contact me at [email protected]

HARDWARE REQUIREMENTS

The following specs were used to create the original solution

Ubuntu 16.04 LTS
128GB RAM (64GB minimum required to run keypoint matching)
Core i9-9980X CPU
8x GPU w/24GB memory each (required to run batch sizes in siamese network training)

note that even with the above HW, training may take 13-30+ days to complete due to exhaustive brute-force keypoint matching

SOFTWARE

python packages are detailed separately in requirements.txt

Python 3.5.1
CUDA 10.0
cuddn 7.3

DATA SETUP

mkdir -p whale/input
cd whale/input

• download the competition train/test files and folders into whale/input

unzip mask_predictions_test_4.zip
unzip mask_predictions_train_known_4.zip

• copy all images !=new_whale from train into the folder: train_specific
• copy all images from train into folder: val

Keypoint matching:

cd kps
python index_features_2kp.py --dataset ../input/train_specific --features-db train --mask-db ../input/mask_predictions_train_known_4
python index_features_2kp.py --dataset ../input/test --features-db test --mask-db ../input/mask_predictions_test_4
python kp_matching.py

Base model training

For this step training is done with close attention to training/val loss and final checkpoints need to be selected at the end of each training cycle and the filenames put back into the training files at the appropriate insertion points. One set of train and inference files are included here. To replicate the solution results, please substiture DenseNet121, Resnet50, and InceptionV3 as the basenetworks and run each for both train and inference (6 total runs: 3 networks x (train + inference).

cd ../base_network
python densenet121_200_train.py
python densenet121_422_train.py

Siamese model training

In this step as well, the output of the previous steps (which can't be predetermined due to checkpoint number and val_loss in file name) will need to be inserted into the py files before running One set of train and inference files are included here. To replicate the solution results, please substiture DenseNet121, Resnet50, and InceptionV3 as the basenetworks and run each for both train and inference (6 total runs: 3 networks x (train + inference).

cd ../siamese
python siamese_train.py
python siamese_inference.py

Ensemble

In this step as well, the output of the previous steps

cd ../ensemble
python ensemble.py