Running PTB LM (10k vocabulary) benchmark with tensor2tensor

t2t_lmptb_10k.sh

#!/bin/bash

set -e

source activate tensorflow

# See what problems, models, and hyperparameter sets are available.
# You can easily swap between them (and add new ones).
#t2t-trainer --registry_help

PROBLEM="lmptb_10k"
MODEL="attention_lm"
HPARAMS="attention_lm_base"

DATA_DIR="./data"
TMP_DIR="/tmp/t2t_datagen"
TRAIN_DIR="attn_lm"

VALID_SRC="$TMP_DIR/simple-examples/data/ptb.valid.txt"

mkdir -p $DATA_DIR $TMP_DIR $TRAIN_DIR

echo "Parameters:"
printf "* Problem:\t$PROBLEM\n"
printf "* Model:\t$MODEL\n"
printf "* HParams:\t$HPARAMS\n"
printf "* Data:\t$DATA_DIR\n"
printf "* Tmp:\t$TMP_DIR\n"
printf "* Train:\t$TRAIN_DIR\n"
printf "* Cuda device:\t$CUDA_VISIBLE_DEVICES\n"

generate(){
  echo "Generating data..."
  t2t-datagen \
    --data_dir=$DATA_DIR \
    --tmp_dir=$TMP_DIR \
    --num_shards=100 \
    --problem=$PROBLEM
}

train(){
  echo "Training..."
  python3 tensor2tensor/bin/t2t-trainer \
          --model="attention_lm" \
          --problems="lmptb_10k" \
          --hparams_set="attention_lm_base" \
          --output_dir="$TRAIN_DIR" \
          --data_dir="$DATA_DIR"
}



decode(){
  echo "Decoding..."
  DECODE_FILE="$VALID_SRC"

  BEAM_SIZE=4
  ALPHA="0.6"

  t2t-trainer \
    --data_dir="$DATA_DIR" \
    --problems="$PROBLEM" \
    --model=$MODEL \
    --hparams_set="$HPARAMS" \
    --output_dir="$TRAIN_DIR" \
    --train_steps=0 \
    --eval_steps=0 \
    --decode_from_file="$DECODE_FILE" \
    --decode_beam_size="$BEAM_SIZE" \
}

action="$1"
echo "$action"
if [ -z "$action" ]; then
  echo "no action. please add one of the following command [gen, train, gentrain, decode]"
elif [ $action = "gen" ]; then
  generate
elif [ $action = "train" ]; then
  train
elif [ $action = "gentrain" ]; then
  generate
  train
elif [ $action = "decode" ]; then
  decode
fi

is this decode function the intended way to predict based on an input?