tensor2tensor icon indicating copy to clipboard operation
tensor2tensor copied to clipboard

Reproduce English to German Machine Translation

Open sarthmit opened this issue 4 years ago • 0 comments

Description

I am unable to reproduce the english to german machine translation using either transformers or universal transformers. I am only able to get around 20 BLEU score for both on newstest2013.

Reproducing Code

PROBLEM=translate_ende_wmt32k
MODEL=transformer
HPARAMS=transformer_base_single_gpu

DATA_DIR=$HOME/tensor2tensor/t2t_data
TMP_DIR=/miniscratch/mittalsa/tmp/t2t_datagen
TRAIN_DIR=$HOME/tensor2tensor/t2t_train/$PROBLEM/$MODEL-$HPARAMS

train_steps=500000 # Total number of train steps for all Epochs
eval_steps=1000 # Number of steps to perform for each evaluation
save_checkpoints_steps=1000
schedule="continuous_train_and_eval"

t2t-trainer \
  --data_dir=$DATA_DIR \
  --problem=$PROBLEM \
  --model=$MODEL \
  --hparams_set=$HPARAMS \
  --schedule=$schedule\
  --output_dir=$TRAIN_DIR \
  --train_steps=$train_steps \
  --worker-gpu=1 \
  --eval_steps=$eval_steps

And I did model universal_transformer and hparams universal_transformer_base for the Universal Transformer part. Then for evaluation, I did

PROBLEM=translate_ende_wmt32k
MODEL=transformer
HPARAMS=transformer_base

DATA_DIR=$HOME/tensor2tensor/t2t_data
TMP_DIR=/miniscratch/mittalsa/tmp/t2t_datagen/dev
TRAIN_DIR=$HOME/tensor2tensor/t2t_train/$PROBLEM/$MODEL-$HPARAMS

DECODE_FILE=$TMP_DIR/newstest2013.en
REF_FILE=$TMP_DIR/newstest2013.de

t2t-translate-all \
  --data_dir=$DATA_DIR \
  --problem=$PROBLEM \
  --model $MODEL \
  --model_dir=$TRAIN_DIR \
  --output_dir=$TRAIN_DIR \
  --hparams_set=$HPARAMS \
  --output_dir=$TRAIN_DIR \
  --decode_hparams="beam_size=4,alpha=0.6" \
  --source=$DECODE_FILE \
  --translations_dir=t2t_translations/$PROBLEM/$MODEL-$HPARAMS \
  --t2t_usr_dir=$HOME

t2t-bleu \
  --translations_dir=t2t_translations/$PROBLEM/$MODEL-$HPARAMS \
  --reference=$REF_FILE \
  --event_dir=events/$PROBLEM/$MODEL-$HPARAMS

I am seeing the outputs for both transformer and universal transformer as around 20-21 BLEU points on newstest2013 data. What should I be changing to be able to reproduce both transformer and universal transformer results?

sarthmit avatar Feb 21 '21 19:02 sarthmit