CorrNet on RWTH-PHOENIX-Weather 2014 (reproducible)

0) Install

python3 -m venv /data/venv/corrnet
source /data/venv/corrnet/bin/activate
pip install -r requirements.txt

Tip:after install, verify

python - <<'PY'
import torch; print("torch", torch.__version__, "cuda", torch.cuda.is_available())
PY

1) Dataset (not included)

Download RWTH-PHOENIX-Weather 2014 [download link]and create a symlink:

Example

ln -s /PATH/phoenix2014-release ./dataset/phoenix2014

2) Preprocess

cd ./preprocess
python dataset_preprocess.py --process-image --multiprocessing

3) Train

source /data/venv/corrnet/bin/activate
export DECODE_MODE=max          # greedy for speed during training/val
mkdir -p work_dir/baseline_res18_run40
python main.py --config ./configs/baseline.yaml --device 0  

And you can generate plot for training process

python parse_and_plot.py --log your_log_path --outdir target_dir_path

4) Test

source /data/venv/corrnet/bin/activate
export DECODE_MODE=max
mkdir -p work_dir/baseline_res18_test
python main.py --config ./configs/test.yaml --device 0

tips: use your own model_weight.pt in test.yaml