arguments.md

Runners

• --in-path - path to data structured as in docs/structure.md
• --out-path - path to save logs and checkpoints
• --name - experiment name
• --time - set True to add date and time to experiment name
• --n-gpus - number of visible GPUs
• --batch-size - batch size
• --n_epochs - number of training epochs; one epoch is 1000 batches
• --backbone - backbone architecture (efficient_net_b* with * in {0, 1, 2, 3, 4, 5, 6, 7}, mobile_net_v2 or vgg19)
• --source-domain - index of source domain; domains are enumerated as ('real', 'infograph', 'quickdraw', 'sketch', 'clipart', 'painting')
• --source-domains - comma-separated list of source domain indexes
• --target-domain - index of target domain
• --phase - name of data split; train or test for multi-source track; labeled or unlabeled for semi-supervised; all can be used for both tracks

MixMatch runners

• --loss-weight - target loss weight
• --temperature - distribution sharpening parameter
• --alpha - beta distribution parameter

Scripts

convert_to_tfrecords.py

• --path - path to data
• --size - number of images per .tfrecord file

download.py

• --path - path to data

ensemble.py

• --out-path - path to load and save predictions
• --in-names - comma-separated names of experiments
• --out-name - current experiment name
• --weights - comma-separated weights for probabilities

make_submission.py

• --in-path - path to data
• --out-path - path to load and save predictions
• --clipart-name - clipart experiment name
• --painting-name - painting experiment name
• --out-name - current experiment name
• --track - 0: multi-source, 1: semi-supervised