For the description of the solution, refer to this post.

This repository contains Dockerfile that covers all required dependencies.

By default mixed precision training and inference is used (see --fp16 flag in main.py) which is fully supported in Volta and Turing architectures.

python main.py --save <experiment name> <additional arguments>

will result in:

• saving checkpoint each epoch into <experiment name>.<epoch>
• saving best checkpoint (on local validation) into <experiment name>
• saving log from training into <experiment name>.log

For additional arguments and default values see python main.py --help or main.py file.

python main.py --mode predict --load <path to checkpoint> <additional arguments>

will result in:

• saving raw predictions into <path to checkpoint>.output or into <path to checkpoint>.output<pred suffix> if --pred-suffix is specified
• saving log into <path to checkpoint>.output.log

Raw predictions are pickled logits for each class and each test and validation image. To convert it into CSV submission or ensemble multiple raw predictions, run:

./make_submission.py -o <csv output file> <path to one or multiple raw predictions>

This script will also ensemble and print score on validation set if it's not empty (see --cv-number flag for main.py).

These examples assume data (in the same format as in Kaggle with extracted files from archives) to be in ../data.

python main.py -e 130 --pl-epoch 90 --lr cosine,1.5e-4,90,6e-5,150,0 --pl-size-func 0.6*x+0.4 --cv-number -1 --seed 0 --save /results/dn161_0
python main.py --mode predict --cv-number -1 --tta 8 --load /results/dn161_0.129
./make_submission.csv /results/dn161_0.129.output -o /results/submission_0.csv

results in submission with 0.99658 private score and 0.98826 public score.

python main.py -e 130 --pl-epoch 90 --lr cosine,1.5e-4,90,6e-5,150,0 --pl-size-func 0.6*x+0.4 --cv-number -1 --seed 1 --save results/dn161_1
python main.py --mode predict --cv-number -1 --tta 8 --load /results/dn161_1.129
./make_submission.csv /results/dn161_1.129.output -o /results/submission_1.csv

results in submission with 0.99623 private score and 0.98871 public score.

Ensembling both of them (./make_submission.csv /results/dn161_0.129.output /results/dn161_1.129.output -o /results/submission.csv) results in 0.99784 private score and 0.99187 public score.