digital-breakthrough-hack

First place solution of Russian Railways hackathon

Note: to reproduce solution you should have a machine with at least 2x3090

Download data

train, test

Data dir tree structure

-- data
    |-- test [1000 entries exceeds filelimit, not opening dir]
    |-- train
    |   |-- images [8203 entries exceeds filelimit, not opening dir]
    |   `- mask [8203 entries exceeds filelimit, not opening dir]

Prepare env

cd mmsegmentation/docker
docker build . -t rlh

Launch docker

Before launching the docker container, change /path/to/data and /path/to/src/ to the appropriate directories.

bash launch_docker.sh

Prepare training masks & split

cd /workspace/rlh
python3 create_masks.py
python3 make_splits.py

Data dir structure now

-- data
    |-- stratified_split
    |   |-- test.txt
    |   |-- train.txt
    |   `-- val.txt
    |-- test [1000 entries exceeds filelimit, not opening dir]
    |-- train
    |   |-- images [8203 entries exceeds filelimit, not opening dir]
    |   |-- mask [8203 entries exceeds filelimit, not opening dir]
    |   |-- meta.csv
    |   `-- new_mask [8203 entries exceeds filelimit, not opening dir]

Train model

cd /workspace/rlh/mmsegmentation
bash tools/dist_train.sh /workspace/rlh/configs/config_27_augs.py 2 --work-dir /workspace/rlh/work_dirs/exp37

Inference trained model

cd /workspace/rlh
python3 inference.py --exp exp37