Please refer to original paper Unsupervised Data Augmentation for Consistency Training (NIPS 2020) and code google-research/uda.
bash train.sh 0
TASK="IMDB"
CUDA_VISIBLE_DEVICES=$1 python run_classifier.py \
--data_dir "./data/${TASK}/" \
--model_name_or_path bert-base-uncased \
--output_dir "./output/tmp" \
--num_train_steps 10000 \
--learning_rate 2e-05 \
--warmup_steps 1000 \
--loss_weight 1.0 \
--softmax_temp 0.85 \
--tsa_mode 'linear' \
--do_train \
--do_eval \
--eval_steps 100 \
--save_steps 1000
| Model | # labeled samples (train) | # unlabeled samples (train) | # test samples | Error rate (from paper) | Error rate (re-implementation) |
|---|---|---|---|---|---|
| BERT | 20 | 69972 | 25000 | 6.50 | TBA |
| UDA | 20 | 69972 | 25000 | 4.20 | TBA |