run_full_experiment.sh

#!/bin/bash
# Run this before if needed : sed -i -e 's/\r$//' run_full_experiment.sh
# Finally, run this command : sh run_full_experiment.sh\

# UNLABELED-SIZE (6000-35,000) ---> 20000

# NUM_PIVOTS (100, 200,300, 400, 500) ---> 100

# NUM_PRE_TRAIN_EPOCHS (20, 40, 60) ---> 20
# UNFROZEN_BERT_LAYERS (1, 2, 3, 5, 8, 12) ---> [2, 8]
# PIVOT_PROB (0.1, 0.3, 0.5,0.8) ---> 0.5
# NON_PIVOT_PROB (0.1, 0.3, 0.5,0.8) ---> 0.1

# PRE_TRAINED_EPOCH
# CNN_OUT_CHANNELS (16, 32, 64) ---> 32
# CNN_FILTER_SIZE (7, 9, 11) --->  9
# BATCH_SIZE (32, 64) ---> 32
# NUM_TRAIN_EPOCHS ---> 20

DATA_DIR=stancedata
MODEL_DIR=stancemodels
DOBASELINE=False

# before running - delete the right model dir
# feminist - atheism - hillary
#for MODEL in feminist_to_atheism atheism_to_feminist feminist_to_hillary hillary_to_feminist atheism_to_hillary hillary_to_atheism
for MODEL in feminist_to_hillary
do
  SRC_DOMAIN="${MODEL%_to_*}" # split model name according to '_to_' and take the prefix
  TRG_DOMAIN="${MODEL#*_to_}" # split model name according to '_to_' and take the suffix
  MODELS_DIR=${MODEL_DIR}/${MODEL}

  # Step 1 - Select pivot features
  # Pivot selection params
  NUM_PIVOTS=100
  PIV_MN_ST=20
  LOG_NAME="log/pivot.log"

  if [ $DOBASELINE == False ]
  then
    python utils/pivot_selection.py \
    --pivot_num=${NUM_PIVOTS} \
    --pivot_min_st=${PIV_MN_ST} \
    --src=${DATA_DIR}/${SRC_DOMAIN} \
    --dest=${DATA_DIR}/${TRG_DOMAIN} \
    --log_name=${LOG_NAME}
  fi

  # Step 2 - Run pivot-based finetuning on a pre-trained BERT
  # Finetuning params
  PIVOT_PROB=0.5
  NON_PIVOT_PROB=0.1
  NUM_PRE_TRAIN_EPOCHS=40
  SAVE_FREQ=10
  UNFROZEN_BERT_LAYERS=8

  mkdir -p ${MODELS_DIR}

  OUTPUT_DIR_NAME=${MODELS_DIR}
  PIVOTS_PATH=${DATA_DIR}/pivots/${MODEL}/${NUM_PIVOTS}_bi

  if [ $DOBASELINE == True ]
  then
    INIT_OUTPUT_EMBEDS=False
  else
    INIT_OUTPUT_EMBEDS=True
  fi

  python perl_pretrain.py \
   --src_domain=${DATA_DIR}/${SRC_DOMAIN} \
   --trg_domain=${DATA_DIR}/${TRG_DOMAIN} \
   --pivot_path=${PIVOTS_PATH} \
   --output_dir=${OUTPUT_DIR_NAME} \
   --num_train_epochs=${NUM_PRE_TRAIN_EPOCHS} \
   --save_every_num_epochs=${SAVE_FREQ} \
   --pivot_prob=${PIVOT_PROB} \
   --non_pivot_prob=${NON_PIVOT_PROB} \
   --num_of_unfrozen_bert_layers=${UNFROZEN_BERT_LAYERS} \
   --init_output_embeds=${INIT_OUTPUT_EMBEDS} \
   --do_baseline=${DOBASELINE} \
   --train_output_embeds

  # Step 3 - Train a classifier on source domain labeled data then predict and evaluate on target domain.
  # Supervised task params
  PRE_TRAINED_EPOCH=${NUM_PRE_TRAIN_EPOCHS}
  CNN_OUT_CHANNELS=32
  BATCH_SIZE=32
  CNN_FILTER_SIZE=9
  NUM_TRAIN_EPOCHS=10

  mkdir -p 5-fold-hyper-tune
  mkdir 5-fold-hyper-tune/${MODEL}/

  TEMP_DIR=${MODELS_DIR}/temp
  mkdir -p ${TEMP_DIR}/
  mkdir -p 5-fold-hyper-tune/${MODEL}/


  for FOLD_NUM in 1 2 3 4 5
  do
    cp ${MODELS_DIR}/pytorch_model${PRE_TRAINED_EPOCH}.bin ${TEMP_DIR}

    python supervised_task_learning.py \
    --in_domain_data_dir=${DATA_DIR}/${SRC_DOMAIN}/ \
    --cross_domain_data_dir=${DATA_DIR}/${TRG_DOMAIN}/ \
    --do_train \
    --output_dir=${TEMP_DIR}/ \
    --load_model \
    --model_name=pytorch_model${PRE_TRAINED_EPOCH}.bin \
    --cnn_window_size=${CNN_FILTER_SIZE} \
    --cnn_out_channels=${CNN_OUT_CHANNELS} \
    --learning_rate=5e-5 \
    --train_batch_size=${BATCH_SIZE} \
    --num_train_epochs=${NUM_TRAIN_EPOCHS} \
    --save_according_to=acc \
    --use_fold \
    --fold_num=${FOLD_NUM} \
    --write_log_for_each_epoch

    COPY_FROM_PATH=${TEMP_DIR}/pytorch_model${PRE_TRAINED_EPOCH}.bin-final_eval_results.txt

    COPY_TO_PATH=5-fold-hyper-tune/${MODEL}/ep-${PRE_TRAINED_EPOCH}_ch-${CNN_OUT_CHANNELS}_batch-${BATCH_SIZE}_filt-${CNN_FILTER_SIZE}_fold-${FOLD_NUM}.txt
    cp ${COPY_FROM_PATH} ${COPY_TO_PATH}
    rm ${TEMP_DIR}/*
  done
done