Delete trans_fairseq/scripts/run23.sh

trans_fairseq/scripts/run23.sh

@@ -1,179 +0,0 @@
-#! /usr/bin/bash
-set -eux
-
-##### FFN + RMSNorm 
-
-train_device=0,1,2,3,4,5,6,7
-eval_device=0
-# xzq-fairseq
-root_dir=$(dirname "$PWD")
-
-src_lang=${1:-"de"}
-tgt_lang=${2:-"en"}
-
-data_name=wmt23
-pair_lang=${src_lang}-${tgt_lang}
-task_name=${src_lang}2${tgt_lang}
-data_dir=$root_dir/data/$pair_lang/$data_name
-raw_data_dir=$data_dir/raw
-trainable_data_dir=$data_dir/trainable_data
-user_dir=$root_dir/my_dir
-
-## eval&decode param
-decode_max_tokens=4096
-beam=5
-nbest=1
-lenpen=1.0
-
-## common param
-criterion=label_smoothed_cross_entropy
-label_smoothing=0.1
-seed=42
-max_epoch=40
-keep_last_epochs=1
-keep_best_checkpoints=5 
-patience=5
-num_workers=8 
-
-# specified param
-# Global Batch=卡数*max-tokens*梯度累计,对于训练数据较大的语种(train-set几十M),global batch在 100k tokens以上较好
-arch=my_transformer_vaswani_wmt_en_de_big
-use_rmsnorm=1
-use_llama_ffn=1
-use_rope=0
-activation_fn=swish
-encoder_ffn_embed_dim=2732
-share_all_embeddings=1
-share_decoder_input_output_embed=1
-learing_rate=1e-3
-warmup=4000
-max_tokens=8192
-weight_decay=0.0
-dropout=0.3
-gradient_accumulation_steps=4
-
-model_dir=$root_dir/exps/$task_name/${conf_name}_${data_name}
-# 根据 消融组件 更改 目录名
-if [ $use_rmsnorm -eq 1 ]; then
-    model_dir=${model_dir}_norm
-fi
-if [ $use_llama_ffn -eq 1 ]; then
-    model_dir=${model_dir}_ffn
-    activation_fn=swish
-fi
-if [ $use_rope -eq 1 ] && [ $use_rmsnorm -eq 1 ] && [ $use_llama_ffn -eq 1 ]; then
-    model_dir=$root_dir/exps/$task_name/${conf_name}_${data_name}_all_nosinpos
-elif [ $use_rope -eq 1 ]; then
-    model_dir=${model_dir}_att_nosinpos
-fi
-mkdir -p $model_dir
-cp ${BASH_SOURCE[0]}  $model_dir
-
-gpu_num=`echo "$train_device" | awk '{split($0,arr,",");print length(arr)}'`
-export CUDA_VISIBLE_DEVICES=$train_device
-cmd="fairseq-train $trainable_data_dir \
---distributed-world-size $gpu_num -s $src_lang -t $tgt_lang \
---arch $arch \
---fp16 \
---optimizer adam --clip-norm 0.0 \
---lr-scheduler inverse_sqrt --warmup-init-lr 1e-07 --warmup-updates $warmup \
---lr $learing_rate --adam-betas  '(0.9, 0.98)' \
---weight-decay $weight_decay \
---dropout $dropout \
---criterion $criterion  --label-smoothing $label_smoothing \
---max-epoch $max_epoch \
---max-tokens $max_tokens \
---update-freq $gradient_accumulation_steps \
---user-dir $user_dir \
---activation-fn $activation_fn \
---encoder-ffn-embed-dim $encoder_ffn_embed_dim \
---seed $seed \
---num-workers $num_workers \
---no-epoch-checkpoints \
---keep-last-epochs $keep_last_epochs \
---keep-best-checkpoints $keep_best_checkpoints \
---patience $patience \
---no-progress-bar \
---log-interval 100 \
---task "translation" \
---ddp-backend no_c10d \
---save-dir $model_dir \
---tensorboard-logdir $model_dir"
-
-# add param
-if [ $share_all_embeddings -eq 1 ]; then
-cmd=${cmd}" --share-all-embeddings "
-fi
-if [ $share_decoder_input_output_embed -eq 1 ]; then
-cmd=${cmd}" --share-decoder-input-output-embed "
-fi
-if [ ${max_update:=0} -ne 0 ]; then
-cmd=${cmd}" --max-update $max_update"
-fi
-if [ $use_rmsnorm -eq 1 ]; then
-cmd=${cmd}" --use-rmsnorm "
-fi
-if [ $use_llama_ffn -eq 1 ]; then
-cmd=${cmd}" --use-llama-ffn "
-fi
-if [ $use_rope -eq 1 ]; then
-cmd=${cmd}" --use-rope --no-token-positional-embeddings "
-fi
-
-# run command
-cur_time=`date +"%Y-%m-%d %H:%M:%S"`
-echo "=============$cur_time===================" >> $model_dir/train.log
-cmd="nohup ${cmd} >> $model_dir/train.log 2>&1 &"
-
-eval $cmd
-
-wait
-
-### decode
-checkpoint_path=$model_dir/checkpoint_best.pt
-save_dir=$model_dir/decode_result
-
-mkdir -p $save_dir
-cp ${BASH_SOURCE[0]}  $save_dir
-
-declare -A gen_subset_dict
-gen_subset_dict=([test]=flores [test1]=wmt22 [test2]=wmt23)
-for gen_subset in ${!gen_subset_dict[*]}
-do
-    decode_file=$save_dir/decode_${gen_subset_dict[$gen_subset]}_beam${beam}_lenpen${lenpen}.$tgt_lang
-    pure_file=$save_dir/pure_decode_${gen_subset_dict[$gen_subset]}_beam${beam}_lenpen${lenpen}.$tgt_lang
-
-    CUDA_VISIBLE_DEVICES=$eval_device fairseq-generate \
-        $trainable_data_dir \
-        -s $src_lang -t $tgt_lang \
-        --user-dir $user_dir \
-        --gen-subset $gen_subset \
-        --path $checkpoint_path  \
-        --max-tokens $decode_max_tokens \
-        --beam $beam \
-        --nbest $nbest \
-        --lenpen $lenpen  \
-        --seed $seed \
-        --remove-bpe | tee $decode_file
-
-    ### eval
-    # purify file
-    grep ^H $decode_file  | LC_ALL=C sort -V | cut -f3- | perl $root_dir/mosesdecoder/scripts/tokenizer/detokenizer.perl -l $tgt_lang > $pure_file
-
-    eval_file=$model_dir/eval_${gen_subset_dict[$gen_subset]}.log
-    cur_time=`date +"%Y-%m-%d %H:%M:%S"`
-    echo "=============$cur_time===================" >> $eval_file
-    echo $checkpoint_path >> $eval_file
-    tail -n1 $decode_file >> $eval_file    # multi-bleu
-    # get score
-    src_file=$raw_data_dir/test.${gen_subset_dict[$gen_subset]}.$src_lang
-    ref_file=$raw_data_dir/test.${gen_subset_dict[$gen_subset]}.$tgt_lang
-    sacrebleu_file=$save_dir/sacrebleu.${gen_subset_dict[$gen_subset]}.beam${beam}_lenpen${lenpen}
-    comet22_file=$save_dir/comet22.${gen_subset_dict[$gen_subset]}.beam${beam}_lenpen${lenpen}
-    sacrebleu $ref_file -i $pure_file -w 2 >> $eval_file
-    comet-score -s $src_file -t $pure_file -r $ref_file --model $root_dir/wmt22-comet-da/checkpoints/model.ckpt | tee $comet22_file
-    echo "Comet22 Score" >> $eval_file
-    tail -n1 $comet22_file >> $eval_file    # 只取平均comet分
-
-    echo -e "decode finished! \n decode tokenized file in $decode_file \n detokenized file in $pure_file \n sacrebleu file in $eval_file"
-done