• train.sh

envs_dir= export PATH=$envs_dir/bin:$PATH base_dir= codes_dir=$base_dir/codes/fairseq datas=$1 seed=1111 max_tokens=4096 dropout=0.1 attention_heads=8 embed_dim=512 ffn_embed_dim=2048 encoder_layer=6 decoder_layer=6 src=en tgt=ch train_dataset_dir= export CUDA_VISIBLE_DEVICES=0,1,2,3 hf_pen=0 warmup_updates=4000 ratio=1 lr=2e-04 update_freq=2 echo "begin lr ${lr}" save_dir=$base_dir/ mkdir -p $save_dir echo "begin training based on hf_len is ${hf_pen} batch_size is ${batch_size}" python $codes_dir/train.py $train_dataset_dir
--task translation_rrhf
--arch hf_transformer_base
--restore-file $base_dir/models/checkpoints-nist-en-ch-subword-32000-test/average.pt
--dataset-impl='raw_json'
--source-lang $src
--target-lang $tgt
--save-dir $save_dir
--optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0
--lr-scheduler inverse_sqrt --lr ${lr} --max-epoch 15
--reset-optimizer
--reset-dataloader
--disable-validation
--batch-size 8
--fp16
--num-workers 0
--warmup-updates ${warmup_updates} --warmup-init-lr '1e-07'
--criterion hf_label_smoothed_cross_entropy --label-smoothing 0.1 --hf-score-penalty ${hf_pen} --hf-loss-ratio ${ratio}
--seed ${seed}
--max-tokens ${max_tokens} --update-freq ${update_freq}
--dropout ${dropout} --relu-dropout 0.1 --attention-dropout 0.1
--decoder-attention-heads ${attention_heads} --encoder-attention-heads ${attention_heads}
--decoder-embed-dim ${embed_dim} --encoder-embed-dim ${embed_dim}
--decoder-ffn-embed-dim ${ffn_embed_dim} --encoder-ffn-embed-dim ${ffn_embed_dim}
--encoder-layers ${encoder_layer} --decoder-layers ${decoder_layer}