decode.py

# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
# Modifications Copyright 2017 Abigail See
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

"""This file contains code to run beam search decoding, including running ROUGE evaluation and producing JSON datafiles for the in-browser attention visualizer, which can be found here https://github.com/abisee/attn_vis"""

import os
import time
import tensorflow as tf
import beam_search
import data
import json
import pyrouge
import util
import logging
import numpy as np
import shutil
from evaluation import evaluation

FLAGS = tf.app.flags.FLAGS

SECS_UNTIL_NEW_CKPT = 60  # max number of seconds before loading new checkpoint


class BeamSearchDecoder(object):
    """Beam search decoder."""

    def __init__(self, model, batcher, vocab, epoch_num=None):
        """Initialize decoder.

        Args:
          model: a Seq2SeqAttentionModel object.
          batcher: a Batcher object.
          vocab: Vocabulary object
        """
        self._model = model
        self._model.build_graph()
        self._batcher = batcher
        self._vocab = vocab
        self._saver = tf.train.Saver()  # we use this to load checkpoints for decoding
        self._sess = tf.Session(config=util.get_config())

        # Load an initial checkpoint to use for decoding
        # ckpt_path = util.load_ckpt(self._saver, self._sess)
        if epoch_num:
            None
        else:
            while True:
                epoch_num = int(input('input epoch num (2~{100})'))
                if epoch_num >= 2 or epoch_num <= 100:
                    break
        tf.logging.info('Loading checkpoint')
        self._saver.restore(self._sess,
                            os.path.join(FLAGS.log_root,'epoch_ckpt', 'ep' + str(epoch_num)))

        if FLAGS.single_pass:
            # Make a descriptive decode directory name
            ckpt_name = str(epoch_num)  # this is something of the form "ckpt-123456"
            self._decode_dir = os.path.join(FLAGS.log_root, get_decode_dir_name(ckpt_name))
            if os.path.exists(self._decode_dir):
                shutil.rmtree(self._decode_dir)

        else:  # Generic decode dir name
            self._decode_dir = os.path.join(FLAGS.log_root, "decode")

        # Make the decode dir if necessary
        if not os.path.exists(self._decode_dir): os.mkdir(self._decode_dir)

        if FLAGS.single_pass:
            # Make the dirs to contain output written in the correct format for pyrouge
            self._rouge_ref_dir = os.path.join(self._decode_dir, "reference")
            if not os.path.exists(self._rouge_ref_dir): os.mkdir(self._rouge_ref_dir)
            self._rouge_dec_dir = os.path.join(self._decode_dir, "decoded")
            if not os.path.exists(self._rouge_dec_dir): os.mkdir(self._rouge_dec_dir)

    def decode(self):
        """Decode examples until data is exhausted (if FLAGS.single_pass) and return, or decode indefinitely, loading latest checkpoint at regular intervals"""
        t0 = time.time()
        counter = 0
        all_logits = []
        all_one_hot_lables = []
        while True:
            batch = self._batcher.next_batch()  # 1 example repeated across batch
            if batch is None:  # finished decoding dataset in single_pass mode
                assert FLAGS.single_pass, "Dataset exhausted, but we are not in single_pass mode"
                tf.logging.info("Decoder has finished reading dataset for single_pass.")
                tf.logging.info("Output has been saved in %s and %s. Now starting ROUGE eval...", self._rouge_ref_dir,
                                self._rouge_dec_dir)
                evaluation(all_logits, all_one_hot_lables, FLAGS.max_side_steps)
                results_dict = rouge_eval(self._rouge_ref_dir, self._rouge_dec_dir)
                rouge_log(results_dict, self._decode_dir)
                return

            original_article = batch.original_articles[0]  # string
            original_abstract = batch.original_abstracts[0]  # string
            original_abstract_sents = batch.original_abstracts_sents[0]  # list of strings

            article_withunks = data.show_art_oovs(original_article, self._vocab)  # string
            abstract_withunks = data.show_abs_oovs(original_abstract, self._vocab,
                                                   (batch.art_oovs[0] if FLAGS.pointer_gen else None))  # string

            # Run beam search to get best Hypothesis
            best_hyp, pic, sticker_logits = beam_search.run_beam_search(self._sess, self._model, self._vocab, batch)


            one_hot_targets = np.eye(FLAGS.max_side_steps)[batch.dec_pic_target]
            for i in range(FLAGS.batch_size):
                logits = sticker_logits[i].tolist()
                one_hot = one_hot_targets[i].tolist()
                for l, o in zip(logits, one_hot):
                    all_logits.append(l)
                    all_one_hot_lables.append(o)

            # Extract the output ids from the hypothesis and convert back to words
            output_ids = [int(t) for t in best_hyp.tokens[1:]]
            decoded_words = data.outputids2words(output_ids, self._vocab,
                                                 (batch.art_oovs[0] if FLAGS.pointer_gen else None))

            # Remove the [STOP] token from decoded_words, if necessary
            try:
                fst_stop_idx = decoded_words.index(data.STOP_DECODING)  # index of the (first) [STOP] symbol
                decoded_words = decoded_words[:fst_stop_idx]
            except ValueError:
                decoded_words = decoded_words
            decoded_output = ' '.join(decoded_words)  # single string

            if FLAGS.single_pass:
                self.write_for_rouge(original_abstract_sents, decoded_words,
                                     counter, pic, batch.dec_pic_target[0])  # write ref summary and decoded summary to file, to eval with pyrouge later
                counter += 1  # this is how many examples we've decoded
            else:
                print_results(article_withunks, abstract_withunks, decoded_output)  # log output to screen
                self.write_for_attnvis(article_withunks, abstract_withunks, decoded_words, best_hyp.attn_dists,
                                       best_hyp.p_gens)  # write info to .json file for visualization tool

                # Check if SECS_UNTIL_NEW_CKPT has elapsed; if so return so we can load a new checkpoint
                t1 = time.time()
                if t1 - t0 > SECS_UNTIL_NEW_CKPT:
                    tf.logging.info(
                        'We\'ve been decoding with same checkpoint for %i seconds. Time to load new checkpoint',
                        t1 - t0)
                    _ = util.load_ckpt(self._saver, self._sess)
                    t0 = time.time()


    def write_for_rouge(self, reference_sents, decoded_words, ex_index, pic, logit_pic):
        """Write output to file in correct format for eval with pyrouge. This is called in single_pass mode.

        Args:
          reference_sents: list of strings
          decoded_words: list of strings
          ex_index: int, the index with which to label the files
        """
        # First, divide decoded output into sentences
        decoded_sents = []
        while len(decoded_words) > 0:
            try:
                fst_period_idx = decoded_words.index(".")
            except ValueError:  # there is text remaining that doesn't end in "."
                fst_period_idx = len(decoded_words)
            sent = decoded_words[:fst_period_idx + 1]  # sentence up to and including the period
            decoded_words = decoded_words[fst_period_idx + 1:]  # everything else
            decoded_sents.append(' '.join(sent))

        # pyrouge calls a perl script that puts the data into HTML files.
        # Therefore we need to make our output HTML safe.
        decoded_sents = [make_html_safe(w) for w in decoded_sents]
        reference_sents = [make_html_safe(w) for w in reference_sents]

        # Write to file
        ref_file = os.path.join(self._rouge_ref_dir, "%06d_reference.txt" % ex_index)
        decoded_file = os.path.join(self._rouge_dec_dir, "%06d_decoded.txt" % ex_index)
        pic_file = os.path.join(self._rouge_dec_dir, "pic_decoded.txt")

        with open(ref_file, "w") as f:
            for idx, sent in enumerate(reference_sents):
                f.write(sent) if idx == len(reference_sents) - 1 else f.write(sent + "\n")
        with open(decoded_file, "w") as f:
            for idx, sent in enumerate(decoded_sents):
                f.write(sent) if idx == len(decoded_sents) - 1 else f.write(sent + "\n")
        with open(pic_file, 'a') as f:
            f.write(str(pic) + ' ' + str(logit_pic) + "\n")
        if str(pic) == str(logit_pic):
            open('/home1/lmz/video_data/new_test_multi', 'a').write(str(ex_index) + '\n')

        tf.logging.info("Wrote example %i to file" % ex_index)

    def write_for_attnvis(self, article, abstract, decoded_words, attn_dists, p_gens):
        """Write some data to json file, which can be read into the in-browser attention visualizer tool:
          https://github.com/abisee/attn_vis

        Args:
          article: The original article string.
          abstract: The human (correct) abstract string.
          attn_dists: List of arrays; the attention distributions.
          decoded_words: List of strings; the words of the generated summary.
          p_gens: List of scalars; the p_gen values. If not running in pointer-generator mode, list of None.
        """
        article_lst = article.split()  # list of words
        decoded_lst = decoded_words  # list of decoded words
        to_write = {
            'article_lst': [make_html_safe(t) for t in article_lst],
            'decoded_lst': [make_html_safe(t) for t in decoded_lst],
            'abstract_str': make_html_safe(abstract),
            'attn_dists': attn_dists
        }
        if FLAGS.pointer_gen:
            to_write['p_gens'] = p_gens
        output_fname = os.path.join(self._decode_dir, 'attn_vis_data.json')
        with open(output_fname, 'w') as output_file:
            json.dump(to_write, output_file)
        tf.logging.info('Wrote visualization data to %s', output_fname)


def print_results(article, abstract, decoded_output):
    """Prints the article, the reference summmary and the decoded summary to screen"""
    print("---------------------------------------------------------------------------")
    tf.logging.info('ARTICLE:  %s', article)
    tf.logging.info('REFERENCE SUMMARY: %s', abstract)
    tf.logging.info('GENERATED SUMMARY: %s', decoded_output)
    print("---------------------------------------------------------------------------")


def make_html_safe(s):
    """Replace any angled brackets in string s to avoid interfering with HTML attention visualizer."""
    s.replace("<", "&lt;")
    s.replace(">", "&gt;")
    return s


def rouge_eval(ref_dir, dec_dir):
    """Evaluate the files in ref_dir and dec_dir with pyrouge, returning results_dict"""
    r = pyrouge.Rouge155()
    r.model_filename_pattern = '#ID#_reference.txt'
    r.system_filename_pattern = '(\d+)_decoded.txt'
    r.model_dir = ref_dir
    r.system_dir = dec_dir
    logging.getLogger('global').setLevel(logging.WARNING)  # silence pyrouge logging
    rouge_results = r.convert_and_evaluate()
    return r.output_to_dict(rouge_results)


def rouge_log(results_dict, dir_to_write):
    """Log ROUGE results to screen and write to file.

    Args:
      results_dict: the dictionary returned by pyrouge
      dir_to_write: the directory where we will write the results to"""
    log_str = ""
    for x in ["1", "2", "l"]:
        log_str += "\nROUGE-%s:\n" % x
        for y in ["f_score", "recall", "precision"]:
            key = "rouge_%s_%s" % (x, y)
            key_cb = key + "_cb"
            key_ce = key + "_ce"
            val = results_dict[key]
            val_cb = results_dict[key_cb]
            val_ce = results_dict[key_ce]
            log_str += "%s: %.4f with confidence interval (%.4f, %.4f)\n" % (key, val, val_cb, val_ce)
    tf.logging.info(log_str)  # log to screen
    results_file = os.path.join(dir_to_write, "ROUGE_results.txt")
    tf.logging.info("Writing final ROUGE results to %s...", results_file)
    with open(results_file, "w") as f:
        f.write(log_str)


def get_decode_dir_name(ckpt_name):
    """Make a descriptive name for the decode dir, including the name of the checkpoint we use to decode. This is called in single_pass mode."""

    if "train" in FLAGS.data_path:
        dataset = "train"
    elif "val" in FLAGS.data_path:
        dataset = "val"
    elif "test" in FLAGS.data_path:
        dataset = "test"
    else:
        raise ValueError("FLAGS.data_path %s should contain one of train, val or test" % (FLAGS.data_path))
    dirname = "decode_%s_%imaxenc_%ibeam_%imindec_%imaxdec" % (
        dataset, FLAGS.max_enc_steps, FLAGS.beam_size, FLAGS.min_dec_steps, FLAGS.max_dec_steps)
    if ckpt_name is not None:
        dirname += "_%s" % ckpt_name
    return dirname