Steps for self-training PRPN:
- Training multiple PRPN instances:
python -u main.py --batch 64 --save trained_models/ --alpha 1.0 --epochs 35 --PRPN --force_binarize
- Store outputs of trained PRPN models (will be stored in the same dir as model_path):
python -u main.py --force_binarize --eval_on train --eval_only --load trained_models/${model_path} --batch 1 --PRPN
- Generate training data from PRPN model outputs:
python -u scripts/overlap.py <comma sep outputs generated in step 2>
- Co-train the multi-task model (from loaded pre-trained model + training outputs):
python -u main.py --batch 64 --PRPN
--shen --alpha 0.5 --save Fout20_12_${load_from}
--beta 0.5 --force_binarize --training_ratio 0.2 --load --train_from_pickle --training_method interleave