catalyst-tutorial

catalyst

more generic classification pipeline

Data preparation

1. Download dataset

download-gdrive 1eAk36MEMjKPKL5j9VWLvNTVKk4ube9Ml artworks.tar.gz
extract-archive artworks.tar.gz &>/dev/null

2. Create DataFrame:

catalyst-data tag2label \
    --in-dir=./data/dataset \
    --out-dataset=./data/dataset.csv \
    --out-labeling=./data/labeling.json \
    --tag-column=class

3. Prepare train / val splits

python3 utils/prepare_splits.py \
    --df=./data/dataset.csv \
    --labeling=./data/labeling.json \
    --out-path=./data/

Training

1. CUDA_VISIBLE_DEVICES="<YOUR_DEVICES>" catalyst-dl run --configs training/configs/train_config.yml

While training

1. Tensorboard training logs are available at <logdir>/train_log
2. Tensorboard validation logs are available at <logdir>/valid_log

After training

1. Weights for final model (which are the best, according to <main metric> on validation) are available at <logdir>/checkpoints/best.pth