#!/bin/bash DATA_PATH=$HOME/.cache/paddle/dataset/wmt16 if [ ! -d $DATA_PATH ] ; then python -c 'import paddle;paddle.dataset.wmt16.train(10000, 10000, "en")'\ '().next()' tar -zxf $DATA_PATH/wmt16.tar.gz -C $DATA_PATH fi train(){ python -u train.py \ --src_vocab_fpath $DATA_PATH/en_10000.dict \ --trg_vocab_fpath $DATA_PATH/de_10000.dict \ --special_token '<s>' '<e>' '<unk>' \ --train_file_pattern $DATA_PATH/wmt16/train \ --val_file_pattern $DATA_PATH/wmt16/val \ --use_token_batch True \ --batch_size 2048 \ --sort_type pool \ --pool_size 10000 \ --enable_ce True \ weight_sharing False \ pass_num 20 \ dropout_seed 10 } train | python _ce.py
