Seq2SeqTrainer (#6769)

Co-authored-by: Sam Shleifer <sshleifer@gmail.com>

examples/seq2seq/builtin_trainer/finetune.sh

@@ -0,0 +1,9 @@
+# the proper usage is documented in the README, you need to specify data_dir, output_dir and model_name_or_path
+# run ./builtin_trainer/finetune.sh --help to see all the possible options
+python finetune_trainer.py \
+    --learning_rate=3e-5 \
+    --fp16 \
+    --do_train --do_eval --do_predict --evaluate_during_training \
+    --predict_with_generate \
+    --n_val 1000 \
+    "$@"

examples/seq2seq/builtin_trainer/finetune_tpu.sh

@@ -0,0 +1,11 @@
+export TPU_NUM_CORES=8
+
+# the proper usage is documented in the README, you need to specify data_dir, output_dir and model_name_or_path
+# run ./builtin_trainer/finetune_tpu.sh --help to see all the possible options
+python xla_spawn.py --num_cores $TPU_NUM_CORES \
+    finetune_trainer.py \
+    --learning_rate=3e-5 \
+    --do_train --do_eval --evaluate_during_training \
+    --prediction_loss_only \
+    --n_val 1000 \
+    "$@"

examples/seq2seq/builtin_trainer/train_distil_marian_enro.sh

@@ -0,0 +1,23 @@
+export WANDB_PROJECT=distil-marian
+export BS=64
+export GAS=1
+export m=sshleifer/student_marian_en_ro_6_3
+export MAX_LEN=128
+python finetune_trainer.py \
+    --tokenizer_name $m --model_name_or_path $m \
+    --data_dir $ENRO_DIR \
+    --output_dir marian_en_ro_6_3 --overwrite_output_dir \
+    --learning_rate=3e-4 \
+    --warmup_steps 500 --sortish_sampler \
+    --fp16 \
+    --gradient_accumulation_steps=$GAS \
+    --per_device_train_batch_size=$BS --per_device_eval_batch_size=$BS \
+    --freeze_encoder --freeze_embeds \
+    --num_train_epochs=6 \
+    --save_steps 3000 --eval_steps 3000 \
+    --max_source_length $MAX_LEN --max_target_length $MAX_LEN --val_max_target_length $MAX_LEN --test_max_target_length $MAX_LEN \
+    --do_train --do_eval --do_predict --evaluate_during_training\
+    --predict_with_generate --logging_first_step \
+    --task translation --label_smoothing 0.1 \
+    --run_name marian_en_ro_6_3 \
+    "$@"

examples/seq2seq/builtin_trainer/train_distil_marian_enro_tpu.sh

@@ -0,0 +1,24 @@
+export WANDB_PROJECT=distil-marian
+export BS=64
+export m=sshleifer/student_marian_en_ro_6_3
+export MAX_LEN=128
+export TPU_NUM_CORES=8
+
+python xla_spawn.py --num_cores $TPU_NUM_CORES \
+    finetune_trainer.py \
+    --tokenizer_name $m --model_name_or_path $m \
+    --data_dir $ENRO_DIR \
+    --output_dir marian_en_ro_6_3 --overwrite_output_dir \
+    --learning_rate=3e-4 \
+    --warmup_steps 500 \
+    --per_device_train_batch_size=$BS --per_device_eval_batch_size=$BS \
+    --freeze_encoder --freeze_embeds \
+    --num_train_epochs=6 \
+    --save_steps 500 --eval_steps 500 \
+    --logging_first_step --logging_steps 200 \
+    --max_source_length $MAX_LEN --max_target_length $MAX_LEN --val_max_target_length $MAX_LEN --test_max_target_length $MAX_LEN \
+    --do_train --do_eval --evaluate_during_training \
+    --prediction_loss_only \
+    --task translation --label_smoothing 0.1 \
+    --run_name marian_en_ro_6_3 \
+    "$@"

examples/seq2seq/builtin_trainer/train_distilbart_cnn.sh

@@ -0,0 +1,26 @@
+export WANDB_PROJECT=distilbart-cnn
+export BS=32
+export GAS=1
+export m=sshleifer/student_cnn_12_6
+export tok=facebook/bart-large
+export MAX_TGT_LEN=142
+
+python finetune_trainer.py \
+    --model_name_or_path $m --tokenizer_name $tok \ 
+    --data_dir $CNN_DIR \
+    --output_dir distilbart-cnn-12-6 --overwrite_output_dir \
+    --learning_rate=3e-5 \
+    --warmup_steps 500 --sortish_sampler \
+    --fp16 \
+    --n_val 500 \
+    --gradient_accumulation_steps=$GAS \
+    --per_device_train_batch_size=$BS --per_device_eval_batch_size=$BS \
+    --freeze_encoder --freeze_embeds \
+    --num_train_epochs=2 \
+    --save_steps 3000 --eval_steps 3000 \
+    --logging_first_step \
+    --max_target_length $MAX_TGT_LEN --val_max_target_length $MAX_TGT_LEN --test_max_target_length $MAX_TGT_LEN \
+    --do_train --do_eval --do_predict --evaluate_during_training \
+    --predict_with_generate \
+    --run_name distilbart-cnn-12-6 \
+    "$@"

examples/seq2seq/builtin_trainer/train_mbart_cc25_enro.sh

@@ -0,0 +1,22 @@
+python finetune_trainer.py \
+    --model_name_or_path=facebook/mbart-large-cc25 \
+    --data_dir $ENRO_DIR \
+    --output_dir mbart_cc25_enro --overwrite_output_dir \
+    --learning_rate=3e-5 \
+    --warmup_steps 500 \ 
+    --fp16 \
+    --label_smoothing 0.1 \
+    --adam_eps 1e-06 \
+    --src_lang en_XX --tgt_lang ro_RO \
+    --freeze_embeds \
+    --per_device_train_batch_size=4 --per_device_eval_batch_size=4 \
+    --max_source_length 128 --max_target_length 128 \
+    --val_max_target_length 128 --test_max_target_length 128 \
+    --sortish_sampler \
+    --num_train_epochs 6 \
+    --save_steps 25000 --eval_steps 25000 --logging_steps 1000 \
+    --do_train --do_eval --do_predict --evaluate_during_training \
+    --predict_with_generate --logging_first_step 
+    --task translation \
+    --run_name mbart_en_ro \
+    "$@"