model.py

'''
Main application script for tagging parts-of-speech and morphosyntactic tags. Run with --help for command line arguments.
'''
from __future__ import division
from collections import Counter
from _collections import defaultdict
from evaluate_morphotags import Evaluator
from sys import maxint

import collections
import argparse
import random
import cPickle
import logging
import progressbar
import os
import dynet as dy
import numpy as np

import utils

__author__ = "Yuval Pinter and Robert Guthrie, 2017"

Instance = collections.namedtuple("Instance", ["sentence", "tags"])

NONE_TAG = "<NONE>"
START_TAG = "<START>"
END_TAG = "<STOP>"
POS_KEY = "POS"
PADDING_CHAR = "<*>"

DEFAULT_WORD_EMBEDDING_SIZE = 64
DEFAULT_CHAR_EMBEDDING_SIZE = 20

class LSTMTagger:
    '''
    Joint POS/morphosyntactic attribute tagger based on LSTM.
    Embeddings are fed into Bi-LSTM layers, then hidden phases are fed into an MLP for each attribute type (including POS tags).
    Class "inspired" by Dynet's BiLSTM tagger tutorial script available at:
    https://github.com/clab/dynet_tutorial_examples/blob/master/tutorial_bilstm_tagger.py
    '''

    def __init__(self, tagset_sizes, num_lstm_layers, hidden_dim, word_embeddings, no_we_update, use_char_rnn, charset_size, char_embedding_dim, att_props=None, vocab_size=None, word_embedding_dim=None):
        '''
        :param tagset_sizes: dictionary of attribute_name:number_of_possible_tags
        :param num_lstm_layers: number of desired LSTM layers
        :param hidden_dim: size of hidden dimension (same for all LSTM layers, including character-level)
        :param word_embeddings: pre-trained list of embeddings, assumes order by word ID (optional)
        :param no_we_update: if toggled, don't update embeddings
        :param use_char_rnn: use "char->tag" option, i.e. concatenate character-level LSTM outputs to word representations (and train underlying LSTM). Only 1-layer is supported.
        :param charset_size: number of characters expected in dataset (needed for character embedding initialization)
        :param char_embedding_dim: desired character embedding dimension
        :param att_props: proportion of loss to assign each attribute for back-propagation weighting (optional)
        :param vocab_size: number of words in model (ignored if pre-trained embeddings are given)
        :param word_embedding_dim: desired word embedding dimension (ignored if pre-trained embeddings are given)
        '''
        self.model = dy.Model()
        self.tagset_sizes = tagset_sizes
        self.attributes = tagset_sizes.keys()
        self.we_update = not no_we_update
        if att_props is not None:
            self.att_props = defaultdict(float, {att:(1.0-p) for att,p in att_props.iteritems()})
        else:
            self.att_props = None

        if word_embeddings is not None: # Use pretrained embeddings
            vocab_size = word_embeddings.shape[0]
            word_embedding_dim = word_embeddings.shape[1]

        self.words_lookup = self.model.add_lookup_parameters((vocab_size, word_embedding_dim), name="we")

        if word_embeddings is not None:
            self.words_lookup.init_from_array(word_embeddings)

        # Char LSTM Parameters
        self.use_char_rnn = use_char_rnn
        if use_char_rnn:
            self.char_lookup = self.model.add_lookup_parameters((charset_size, char_embedding_dim), name="ce")
            self.char_bi_lstm = dy.BiRNNBuilder(1, char_embedding_dim, hidden_dim, self.model, dy.LSTMBuilder)

        # Word LSTM parameters
        if use_char_rnn:
            input_dim = word_embedding_dim + hidden_dim
        else:
            input_dim = word_embedding_dim
        self.word_bi_lstm = dy.BiRNNBuilder(num_lstm_layers, input_dim, hidden_dim, self.model, dy.LSTMBuilder)

        # Matrix that maps from Bi-LSTM output to num tags
        self.lstm_to_tags_params = {}
        self.lstm_to_tags_bias = {}
        self.mlp_out = {}
        self.mlp_out_bias = {}
        for att, set_size in tagset_sizes.items():
            self.lstm_to_tags_params[att] = self.model.add_parameters((set_size, hidden_dim), name=att+"H")
            self.lstm_to_tags_bias[att] = self.model.add_parameters(set_size, name=att+"Hb")
            self.mlp_out[att] = self.model.add_parameters((set_size, set_size), name=att+"O")
            self.mlp_out_bias[att] = self.model.add_parameters(set_size, name=att+"Ob")

    def word_rep(self, word, char_ids):
        '''
        :param word: index of word in lookup table
        '''
        wemb = dy.lookup(self.words_lookup, word, update=self.we_update)
        if char_ids is None:
            return wemb

        # add character representation
        char_embs = [self.char_lookup[cid] for cid in char_ids]
        char_exprs = self.char_bi_lstm.transduce(char_embs)
        return dy.concatenate([ wemb, char_exprs[-1] ])

    def build_tagging_graph(self, sentence, word_chars):
        dy.renew_cg()

        if word_chars == None:
            embeddings = [self.word_rep(w, None) for w in sentence]
        else:
            embeddings = [self.word_rep(w, chars) for w, chars in zip(sentence, word_chars)]

        lstm_out = self.word_bi_lstm.transduce(embeddings)

        H = {}
        Hb = {}
        O = {}
        Ob = {}
        scores = {}
        for att in self.attributes:
            H[att] = dy.parameter(self.lstm_to_tags_params[att])
            Hb[att] = dy.parameter(self.lstm_to_tags_bias[att])
            O[att] = dy.parameter(self.mlp_out[att])
            Ob[att] = dy.parameter(self.mlp_out_bias[att])
            scores[att] = []
            for rep in lstm_out:
                score_t = O[att] * dy.tanh(H[att] * rep + Hb[att]) + Ob[att]
                scores[att].append(score_t)

        return scores

    def loss(self, sentence, word_chars, tags_set):
        '''
        For use in training phase.
        Tag sentence (all attributes) and compute loss based on probability of expected tags.
        '''
        observations_set = self.build_tagging_graph(sentence, word_chars)
        errors = {}
        for att, tags in tags_set.iteritems():
            err = []
            for obs, tag in zip(observations_set[att], tags):
                err_t = dy.pickneglogsoftmax(obs, tag)
                err.append(err_t)
            errors[att] = dy.esum(err)
        if self.att_props is not None:
            for att, err in errors.iteritems():
                prop_vec = dy.inputVector([self.att_props[att]] * err.dim()[0])
                err = dy.cmult(err, prop_vec)
        return errors

    def tag_sentence(self, sentence, word_chars):
        '''
        For use in testing phase.
        Tag sentence and return tags for each attribute, without caluclating loss.
        '''
        observations_set = self.build_tagging_graph(sentence, word_chars)
        tag_seqs = {}
        for att, observations in observations_set.iteritems():
            observations = [ dy.softmax(obs) for obs in observations ]
            probs = [ obs.npvalue() for obs in observations ]
            tag_seq = []
            for prob in probs:
                tag_t = np.argmax(prob)
                tag_seq.append(tag_t)
            tag_seqs[att] = tag_seq
        return tag_seqs

    def set_dropout(self, p):
        self.word_bi_lstm.set_dropout(p)

    def disable_dropout(self):
        self.word_bi_lstm.disable_dropout()
        
    def save(self, file_name):
        '''
        Serialize model parameters for future loading and use.
        TODO change reading in scripts/test_model.py
        '''
        self.model.save(file_name)
        
        with open(file_name + "-atts", 'w') as attdict:
            attdict.write("\t".join(sorted(self.attributes)))

    def old_save(self, file_name):
        '''
        Serialize model parameters for future loading and use.
        Old version (pre dynet 2.0) loaded using initializer in scripts/test_model.py
        '''
        members_to_save = []
        members_to_save.append(self.words_lookup)
        if (self.use_char_rnn):
            members_to_save.append(self.char_lookup)
            members_to_save.append(self.char_bi_lstm)
        members_to_save.append(self.word_bi_lstm)
        members_to_save.extend(utils.sortvals(self.lstm_to_tags_params))
        members_to_save.extend(utils.sortvals(self.lstm_to_tags_bias))
        members_to_save.extend(utils.sortvals(self.mlp_out))
        members_to_save.extend(utils.sortvals(self.mlp_out_bias))
        self.model.save(file_name, members_to_save)

        with open(file_name + "-atts", 'w') as attdict:
            attdict.write("\t".join(sorted(self.attributes)))

    @property
    def model(self):
        return self.model

### END OF CLASSES ###

def get_att_prop(instances):
    logging.info("Calculating attribute proportions for proportional loss margin or proportional loss magnitude")
    total_tokens = 0
    att_counts = Counter()
    for instance in instances:
        total_tokens += len(instance.sentence)
        for att, tags in instance.tags.items():
            t2i = t2is[att]
            att_counts[att] += len([t for t in tags if t != t2i.get(NONE_TAG, -1)])
    return {att:(1.0 - (att_counts[att] / total_tokens)) for att in att_counts}

def get_word_chars(sentence, i2w, c2i):
    pad_char = c2i[PADDING_CHAR]
    return [[pad_char] + [c2i[c] for c in i2w[word]] + [pad_char] for word in sentence]

if __name__ == "__main__":

    # ===-----------------------------------------------------------------------===
    # Argument parsing
    # ===-----------------------------------------------------------------------===
    parser = argparse.ArgumentParser()
    parser.add_argument("--dataset", required=True, dest="dataset", help=".pkl file to use")
    parser.add_argument("--word-embeddings", dest="word_embeddings", help="File from which to read in pretrained embeds (if not supplied, will be random)")
    parser.add_argument("--num-epochs", default=20, dest="num_epochs", type=int, help="Number of full passes through training set (default - 20)")
    parser.add_argument("--num-lstm-layers", default=2, dest="lstm_layers", type=int, help="Number of LSTM layers (default - 2)")
    parser.add_argument("--hidden-dim", default=128, dest="hidden_dim", type=int, help="Size of LSTM hidden layers (default - 128)")
    parser.add_argument("--training-sentence-size", default=maxint, dest="training_sentence_size", type=int, help="Instance count of training set (default - unlimited)")
    parser.add_argument("--token-size", default=maxint, dest="token_size", type=int, help="Token count of training set (default - unlimited)")
    parser.add_argument("--learning-rate", default=0.01, dest="learning_rate", type=float, help="Initial learning rate (default - 0.01)")
    parser.add_argument("--dropout", default=-1, dest="dropout", type=float, help="Amount of dropout to apply to LSTM part of graph (default - off)")
    parser.add_argument("--no-we-update", dest="no_we_update", action="store_true", help="Word Embeddings aren't updated")
    parser.add_argument("--loss-prop", dest="loss_prop", action="store_true", help="Proportional loss magnitudes")
    parser.add_argument("--use-char-rnn", dest="use_char_rnn", action="store_true", help="Use character RNN (default - off)")
    parser.add_argument("--log-dir", default="log", dest="log_dir", help="Directory where to write logs / serialized models")
    parser.add_argument("--no-model", dest="no_model", action="store_true", help="Don't serialize models")
    parser.add_argument("--dynet-mem", help="Ignore this external argument")
    parser.add_argument("--debug", dest="debug", action="store_true", help="Debug mode")
    options = parser.parse_args()


    # ===-----------------------------------------------------------------------===
    # Set up logging
    # ===-----------------------------------------------------------------------===
    if not os.path.exists(options.log_dir):
        os.mkdir(options.log_dir)
    logging.basicConfig(filename=options.log_dir + "/log.txt", filemode="w", format="%(message)s", level=logging.INFO)
    train_dev_cost = utils.CSVLogger(options.log_dir + "/train_dev.log", ["Train.cost", "Dev.cost"])


    # ===-----------------------------------------------------------------------===
    # Log run parameters
    # ===-----------------------------------------------------------------------===

    logging.info(
    """
    Dataset: {}
    Pretrained Embeddings: {}
    Num Epochs: {}
    LSTM: {} layers, {} hidden dim
    Concatenating character LSTM: {}
    Training set size limit: {} sentences or {} tokens
    Initial Learning Rate: {}
    Dropout: {}
    LSTM loss weights proportional to attribute frequency: {}

    """.format(options.dataset, options.word_embeddings, options.num_epochs, options.lstm_layers, options.hidden_dim, options.use_char_rnn, \
               options.training_sentence_size, options.token_size, options.learning_rate, options.dropout, options.loss_prop))

    if options.debug:
        print "DEBUG MODE"

    # ===-----------------------------------------------------------------------===
    # Read in dataset
    # ===-----------------------------------------------------------------------===
    dataset = cPickle.load(open(options.dataset, "r"))
    w2i = dataset["w2i"]
    t2is = dataset["t2is"]
    c2i = dataset["c2i"]
    i2w = { i: w for w, i in w2i.items() } # Inverse mapping
    i2ts = { att: {i: t for t, i in t2i.items()} for att, t2i in t2is.items() }
    i2c = { i: c for c, i in c2i.items() }

    training_instances = dataset["training_instances"]
    training_vocab = dataset["training_vocab"]
    dev_instances = dataset["dev_instances"]
    dev_vocab = dataset["dev_vocab"]
    test_instances = dataset["test_instances"]

    # trim training set for size evaluation (sentence based)
    if len(training_instances) > options.training_sentence_size:
        random.shuffle(training_instances)
        training_instances = training_instances[:options.training_sentence_size]

    # trim training set for size evaluation (token based)
    training_corpus_size = sum(training_vocab.values())
    if training_corpus_size > options.token_size:
        random.shuffle(training_instances)
        cumulative_tokens = 0
        cutoff_index = -1
        for i,inst in enumerate(training_instances):
            cumulative_tokens += len(inst.sentence)
            if cumulative_tokens >= options.token_size:
                training_instances = training_instances[:i+1]
                break

    # ===-----------------------------------------------------------------------===
    # Build model and trainer
    # ===-----------------------------------------------------------------------===
    if options.word_embeddings is not None:
        word_embeddings = utils.read_pretrained_embeddings(options.word_embeddings, w2i)
    else:
        word_embeddings = None

    tag_set_sizes = { att: len(t2i) for att, t2i in t2is.items() }

    if options.loss_prop:
        att_props = get_att_prop(training_instances)
    else:
        att_props = None

    model = LSTMTagger(tagset_sizes=tag_set_sizes,
                       num_lstm_layers=options.lstm_layers,
                       hidden_dim=options.hidden_dim,
                       word_embeddings=word_embeddings,
                       no_we_update = options.no_we_update,
                       use_char_rnn=options.use_char_rnn,
                       charset_size=len(c2i),
                       char_embedding_dim=DEFAULT_CHAR_EMBEDDING_SIZE,
                       att_props=att_props,
                       vocab_size=len(w2i),
                       word_embedding_dim=DEFAULT_WORD_EMBEDDING_SIZE)

    trainer = dy.MomentumSGDTrainer(model.model, options.learning_rate, 0.9)
    logging.info("Training Algorithm: {}".format(type(trainer)))

    logging.info("Number training instances: {}".format(len(training_instances)))
    logging.info("Number dev instances: {}".format(len(dev_instances)))

    for epoch in xrange(int(options.num_epochs)):
        bar = progressbar.ProgressBar()

        # set up epoch
        random.shuffle(training_instances)
        train_loss = 0.0

        if options.dropout > 0:
            model.set_dropout(options.dropout)

        # debug samples small set for faster full loop
        if options.debug:
            train_instances = training_instances[0:int(len(training_instances)/20)]
        else:
            train_instances = training_instances

        # main training loop
        for idx,instance in enumerate(bar(train_instances)):
            if len(instance.sentence) == 0: continue

            gold_tags = instance.tags
            for att in model.attributes:
                if att not in instance.tags:
                    # 'pad' entire sentence with none tags
                    gold_tags[att] = [t2is[att][NONE_TAG]] * len(instance.sentence)

            word_chars = None if not options.use_char_rnn\
                                else get_word_chars(instance.sentence, i2w, c2i)

            # calculate all losses for sentence
            loss_exprs = model.loss(instance.sentence, word_chars, gold_tags)
            loss_expr = dy.esum(loss_exprs.values())
            loss = loss_expr.scalar_value()

            # bail if loss is NaN
            if np.isnan(loss):
                assert False, "NaN occured"

            train_loss += (loss / len(instance.sentence))

            # backward pass and parameter update
            loss_expr.backward()
            trainer.update()

        # log epoch's train phase
        logging.info("\n")
        logging.info("Epoch {} complete".format(epoch + 1))
        # here used to be a learning rate update, no longer supported in dynet 2.0
        print trainer.status()

        train_loss = train_loss / len(train_instances)

        # evaluate dev data
        model.disable_dropout()
        dev_loss = 0.0
        dev_correct = Counter()
        dev_total = Counter()
        dev_oov_total = Counter()
        bar = progressbar.ProgressBar()
        total_wrong = Counter()
        total_wrong_oov = Counter()
        f1_eval = Evaluator(m = 'att')
        if options.debug:
            d_instances = dev_instances[0:int(len(dev_instances)/10)]
        else:
            d_instances = dev_instances
        with open("{}/devout_epoch-{:02d}.txt".format(options.log_dir, epoch + 1), 'w') as dev_writer:
            for instance in bar(d_instances):
                if len(instance.sentence) == 0: continue
                gold_tags = instance.tags
                word_chars = None if not options.use_char_rnn else get_word_chars(instance.sentence, i2w, c2i)
                for att in model.attributes:
                    if att not in instance.tags:
                        gold_tags[att] = [t2is[att][NONE_TAG]] * len(instance.sentence)
                losses = model.loss(instance.sentence, word_chars, gold_tags)
                total_loss = sum([l.scalar_value() for l in losses.values()])
                out_tags_set = model.tag_sentence(instance.sentence, word_chars)

                gold_strings = utils.morphotag_strings(i2ts, gold_tags)
                obs_strings = utils.morphotag_strings(i2ts, out_tags_set)
                for g, o in zip(gold_strings, obs_strings):
                    f1_eval.add_instance(utils.split_tagstring(g, has_pos=True), utils.split_tagstring(o, has_pos=True))
                for att, tags in gold_tags.items():
                    out_tags = out_tags_set[att]
                    correct_sent = True

                    oov_strings = []
                    for word, gold, out in zip(instance.sentence, tags, out_tags):
                        if gold == out:
                            dev_correct[att] += 1
                        else:
                            # Got the wrong tag
                            total_wrong[att] += 1
                            correct_sent = False
                            if i2w[word] not in training_vocab:
                                total_wrong_oov[att] += 1

                        if i2w[word] not in training_vocab:
                            dev_oov_total[att] += 1
                            oov_strings.append("OOV")
                        else:
                            oov_strings.append("")

                    dev_total[att] += len(tags)

                dev_loss += (total_loss / len(instance.sentence))

                dev_writer.write(("\n"
                                 + "\n".join(["\t".join(z) for z in zip([i2w[w] for w in instance.sentence],
                                                                             gold_strings, obs_strings, oov_strings)])
                                 + "\n").encode('utf8'))


        dev_loss = dev_loss / len(d_instances)

        # log epoch results
        logging.info("POS Dev Accuracy: {}".format(dev_correct[POS_KEY] / dev_total[POS_KEY]))
        logging.info("POS % OOV accuracy: {}".format((dev_oov_total[POS_KEY] - total_wrong_oov[POS_KEY]) / dev_oov_total[POS_KEY]))
        if total_wrong[POS_KEY] > 0:
            logging.info("POS % Wrong that are OOV: {}".format(total_wrong_oov[POS_KEY] / total_wrong[POS_KEY]))
        for attr in t2is.keys():
            if attr != POS_KEY:
                logging.info("{} F1: {}".format(attr, f1_eval.mic_f1(att = attr)))
        logging.info("Total attribute F1s: {} micro, {} macro, POS included = {}".format(f1_eval.mic_f1(), f1_eval.mac_f1(), False))

        logging.info("Total dev tokens: {}, Total dev OOV: {}, % OOV: {}".format(dev_total[POS_KEY], dev_oov_total[POS_KEY], dev_oov_total[POS_KEY] / dev_total[POS_KEY]))

        logging.info("Train Loss: {}".format(train_loss))
        logging.info("Dev Loss: {}".format(dev_loss))
        train_dev_cost.add_column([train_loss, dev_loss])

        if epoch > 1 and epoch % 10 != 0: # leave outputs from epochs 1,10,20, etc.
            old_devout_file_name = "{}/devout_epoch-{:02d}.txt".format(options.log_dir, epoch)
            os.remove(old_devout_file_name)

        # serialize model
        if not options.no_model:
            new_model_file_name = "{}/model_epoch-{:02d}.bin".format(options.log_dir, epoch + 1)
            logging.info("Saving model to {}".format(new_model_file_name))
            model.save(new_model_file_name)
            if epoch > 1 and epoch % 10 != 0: # leave models from epochs 1,10,20, etc.
                logging.info("Removing files from previous epoch.")
                old_model_file_name = "{}/model_epoch-{:02d}.bin".format(options.log_dir, epoch)
                os.remove(old_model_file_name)
                os.remove(old_model_file_name + "-atts")

        # epoch loop ends

    # evaluate test data (once)
    logging.info("\n")
    logging.info("Number test instances: {}".format(len(test_instances)))
    model.disable_dropout()
    test_correct = Counter()
    test_total = Counter()
    test_oov_total = Counter()
    bar = progressbar.ProgressBar()
    total_wrong = Counter()
    total_wrong_oov = Counter()
    f1_eval = Evaluator(m = 'att')
    if options.debug:
        t_instances = test_instances[0:int(len(test_instances)/10)]
    else:
        t_instances = test_instances
    with open("{}/testout.txt".format(options.log_dir), 'w') as test_writer:
        for instance in bar(t_instances):
            if len(instance.sentence) == 0: continue
            gold_tags = instance.tags
            for att in model.attributes:
                if att not in instance.tags:
                    gold_tags[att] = [t2is[att][NONE_TAG]] * len(instance.sentence)
            word_chars = word_chars = None if not options.use_char_rnn else get_word_chars(instance.sentence, i2w, c2i)
            out_tags_set = model.tag_sentence(instance.sentence, word_chars)

            gold_strings = utils.morphotag_strings(i2ts, gold_tags)
            obs_strings = utils.morphotag_strings(i2ts, out_tags_set)
            for g, o in zip(gold_strings, obs_strings):
                f1_eval.add_instance(utils.split_tagstring(g, has_pos=True), utils.split_tagstring(o, has_pos=True))
            for att, tags in gold_tags.items():
                out_tags = out_tags_set[att]

                oov_strings = []
                for word, gold, out in zip(instance.sentence, tags, out_tags):
                    if gold == out:
                        test_correct[att] += 1
                    else:
                        # Got the wrong tag
                        total_wrong[att] += 1
                        if i2w[word] not in training_vocab:
                            total_wrong_oov[att] += 1

                    if i2w[word] not in training_vocab:
                        test_oov_total[att] += 1
                        oov_strings.append("OOV")
                    else:
                        oov_strings.append("")

                test_total[att] += len(tags)
            test_writer.write(("\n"
                             + "\n".join(["\t".join(z) for z in zip([i2w[w] for w in instance.sentence],
                                                                         gold_strings, obs_strings, oov_strings)])
                             + "\n").encode('utf8'))


    # log test results
    logging.info("POS Test Accuracy: {}".format(test_correct[POS_KEY] / test_total[POS_KEY]))
    logging.info("POS % Test OOV accuracy: {}".format((test_oov_total[POS_KEY] - total_wrong_oov[POS_KEY]) / test_oov_total[POS_KEY]))
    if total_wrong[POS_KEY] > 0:
        logging.info("POS % Test Wrong that are OOV: {}".format(total_wrong_oov[POS_KEY] / total_wrong[POS_KEY]))
    for attr in t2is.keys():
        if attr != POS_KEY:
            logging.info("{} F1: {}".format(attr, f1_eval.mic_f1(att = attr)))
    logging.info("Total attribute F1s: {} micro, {} macro, POS included = {}".format(f1_eval.mic_f1(), f1_eval.mac_f1(), False))

    logging.info("Total test tokens: {}, Total test OOV: {}, % OOV: {}".format(test_total[POS_KEY], test_oov_total[POS_KEY], test_oov_total[POS_KEY] / test_total[POS_KEY]))