def main(training_file, training_dir, load_model, skip_train): logging.debug('Initializing random seed to 0.') random.seed(0) np.random.seed(0) if load_model: tagger = Tagger.load(load_model) data = TaggingDataset.load_from_file(training_file, vocab=tagger.vocab, tags=tagger.tags) else: assert not skip_train, 'Cannot --skip_train without a saved model.' logging.debug('Loading dataset from: %s' % training_file) data = TaggingDataset.load_from_file(training_file) logging.debug('Initializing model.') tagger = Tagger(data.vocab, data.tags) if not skip_train: train_data, dev_data = data.split(0.7) batches_train = train_data.prepare_batches(n_seqs_per_batch=10) batches_dev = dev_data.prepare_batches(n_seqs_per_batch=100) train_mgr = TrainingManager( avg_n_losses=len(batches_train), training_dir=training_dir, tagger_taste_fn=lambda: taste_tagger(tagger, batches_train), tagger_dev_eval_fn=lambda: eval_tagger(tagger, batches_dev), tagger_save_fn=lambda fname: tagger.save(fname)) logging.debug('Starting training.') while train_mgr.should_continue(): mb_x, mb_y = random.choice(batches_train) mb_loss = tagger.learn(mb_x, mb_y) train_mgr.tick(mb_loss=mb_loss) evaluate_tagger_and_writeout(tagger)
def main(training_file, training_dir, load_model, skip_train): logging.debug('Initializing random seed to 0.') random.seed(0) np.random.seed(0) if load_model: tagger = Tagger.load(load_model) data = TaggingDataset.load_from_file(training_file, vocab=tagger.vocab, tags=tagger.tags) else: assert not skip_train, 'Cannot --skip_train without a saved model.' logging.debug('Loading dataset from: %s' % training_file) data = TaggingDataset.load_from_file(training_file) logging.debug('Initializing model.') tagger = Tagger(data.vocab, data.tags) if not skip_train: train_data, dev_data = data.split(0.7) batches_train = train_data.prepare_batches(n_seqs_per_batch=10) batches_dev = dev_data.prepare_batches(n_seqs_per_batch=100) train_mgr = TrainingManager( avg_n_losses=len(batches_train), training_dir=training_dir, tagger_taste_fn=lambda: taste_tagger(tagger, batches_train), tagger_dev_eval_fn=lambda: eval_tagger(tagger, batches_dev), tagger_save_fn=lambda fname: tagger.save(fname) ) logging.debug('Starting training.') while train_mgr.should_continue(): mb_x, mb_y = random.choice(batches_train) mb_loss = tagger.learn(mb_x, mb_y) train_mgr.tick(mb_loss=mb_loss) evaluate_tagger_and_writeout(tagger)
def main(args): logging.debug('Initializing random seed to 0.') random.seed(0) np.random.seed(0) tf.set_random_seed(0) logging.debug('Loading training dataset from: %s' % args.training_file) train_data = TaggingDataset.load_from_file(args.training_file) dev_data = TaggingDataset.load_from_file(None, vocab=train_data.vocab, alphabet=train_data.alphabet, tags=train_data.tags) logging.debug('Initializing model.') tagger = Tagger(train_data.vocab, train_data.tags, train_data.alphabet, word_embedding_size=args.word_embedding_size, char_embedding_size=args.char_embedding_size, num_chars=args.max_word_length, num_steps=args.max_sentence_length, optimizer_desc=args.optimizer, generate_lemmas=args.generate_lemmas, l2=args.l2, dropout_prob_values=[float(x) for x in args.dropout.split(",")], experiment_name=args.exp_name, supply_form_characters_to_lemma=args.supply_form_characters_to_lemma, threads=args.threads, use_attention=args.use_attention, scheduled_sampling=args.scheduled_sampling) batches_train = train_data.prepare_batches( args.batch_size, args.max_sentence_length, args.max_word_length) batches_dev = dev_data.prepare_batches( 2100, args.max_sentence_length, args.max_word_length) train_mgr = TrainingManager( len(batches_train), args.eval_interval, training_dir=args.training_dir, tagger_taste_fn=lambda: taste_tagger(tagger, batches_train), tagger_dev_eval_fn=lambda: eval_tagger(tagger, batches_dev), tagger_save_fn=lambda fname: tagger.save(fname) ) import signal force_eval = {"value": False} def handle_sigquit(signal, frame): logging.debug("Ctrl+\\ recieved, evaluation will be forced.") force_eval["value"] = True pass signal.signal(signal.SIGQUIT, handle_sigquit) logging.debug('Starting training.') try: permuted_batches = [] while train_mgr.should_continue(max_epochs=args.max_epochs): if not permuted_batches: permuted_batches = batches_train[:] random.shuffle(permuted_batches) words, chars, tags, lengths, lemma_chars, chars_lengths = permuted_batches.pop() oov_mask = np.vectorize(lambda x: train_data.vocab.count(x) == 1 and np.random.uniform() < args.oov_sampling_p)(words) words = np.where(oov_mask, np.zeros(words.shape), words) mb_loss = tagger.learn(words, chars, tags, lengths, lemma_chars, chars_lengths) train_mgr.tick(mb_loss=mb_loss, force_eval=force_eval["value"]) force_eval["value"] = False except KeyboardInterrupt: logging.debug("Ctrl+C recieved, stopping training.") run_tagger_and_writeout(tagger, dev_data)