def main(in_dataset_folder, in_noisy_dataset_folder, in_custom_vocab_file, in_model_folder, in_config):
with open(in_config, encoding='utf-8') as config_in:
config = json.load(config_in)
train_dialogs, train_indices = read_dialogs(os.path.join(in_dataset_folder, 'dialog-babi-task6-dstc2-trn.txt'),
with_indices=True)
dev_dialogs, dev_indices = read_dialogs(os.path.join(in_dataset_folder, 'dialog-babi-task6-dstc2-dev.txt'),
with_indices=True)
test_dialogs, test_indices = read_dialogs(os.path.join(in_noisy_dataset_folder, 'dialog-babi-task6-dstc2-tst.txt'),
with_indices=True)
kb = make_augmented_knowledge_base(os.path.join(in_dataset_folder, 'dialog-babi-task6-dstc2-kb.txt'),
os.path.join(in_dataset_folder, 'dialog-babi-task6-dstc2-candidates.txt'))
max_noisy_dialog_length = max([item['end'] - item['start'] + 1 for item in test_indices])
config['max_input_length'] = max_noisy_dialog_length
post_ood_turns_clean, post_ood_turns_noisy = mark_post_ood_turns(test_dialogs, BABI_CONFIG['backoff_utterance'].lower())
et = EntityTracker(kb)
at = ActionTracker(os.path.join(in_dataset_folder, 'dialog-babi-task6-dstc2-candidates.txt'), et)
if in_custom_vocab_file is not None:
with open(in_custom_vocab_file) as vocab_in:
rev_vocab = [line.rstrip() for line in vocab_in]
vocab = {word: idx for idx, word in enumerate(rev_vocab)}
else:
utterances_tokenized = []
for dialog in train_dialogs:
utterances_tokenized += list(map(lambda x: x.split(), dialog))
vocab, rev_vocab = make_vocabulary(utterances_tokenized,
config['max_vocabulary_size'],
special_tokens=[PAD, START, UNK, EOS] + list(kb.keys()))
config['vocabulary_size'] = len(vocab)
data_train = make_dataset_for_hierarchical_hcn(train_dialogs, train_indices, vocab, et, at, **config)
data_dev = make_dataset_for_hierarchical_hcn(dev_dialogs, dev_indices, vocab, et, at, **config)
data_test = make_dataset_for_hierarchical_hcn(test_dialogs, test_indices, vocab, et, at, **config)
random_input = generate_dropout_turns_for_hierarchical_hcn(10000,
config['max_sequence_length'],
[utterance[0] for utterance in train_dialogs],
vocab,
config['turn_word_dropout_prob'])
save_model(rev_vocab, config, kb, at.action_templates, in_model_folder)
trainer = Trainer(data_train,
data_dev,
data_test,
at.action_templates,
random_input,
post_ood_turns_noisy,
config,
vocab,
in_model_folder)
trainer.train()
def main(in_dataset_folder, in_custom_vocab_file, in_model_folder, in_config):
with open(in_config, encoding='utf-8') as config_in:
config = json.load(config_in)
train_dialogs, train_indices = read_dialogs(os.path.join(
in_dataset_folder, 'dialog-babi-task6-dstc2-trn.txt'),
with_indices=True)
dev_dialogs, dev_indices = read_dialogs(os.path.join(
in_dataset_folder, 'dialog-babi-task6-dstc2-dev.txt'),
with_indices=True)
test_dialogs, test_indices = read_dialogs(os.path.join(
in_dataset_folder, 'dialog-babi-task6-dstc2-tst.txt'),
with_indices=True)
kb = make_augmented_knowledge_base(
os.path.join(in_dataset_folder, 'dialog-babi-task6-dstc2-kb.txt'),
os.path.join(in_dataset_folder,
'dialog-babi-task6-dstc2-candidates.txt'))
et = EntityTracker(kb)
at = ActionTracker(
os.path.join(in_dataset_folder,
'dialog-babi-task6-dstc2-candidates.txt'), et)
if in_custom_vocab_file is not None:
with open(in_custom_vocab_file) as vocab_in:
rev_vocab = [line.rstrip() for line in vocab_in]
vocab = {word: idx for idx, word in enumerate(rev_vocab)}
else:
utterances_tokenized = []
for dialog in train_dialogs:
utterances_tokenized += list(map(lambda x: x.split(), dialog))
vocab, rev_vocab = make_vocabulary(
utterances_tokenized,
config['max_vocabulary_size'],
special_tokens=[PAD, START, UNK, EOS] + list(kb.keys()))
config['vocabulary_size'] = len(vocab)
data_train = make_dataset_for_variational_hcn(train_dialogs, train_indices,
vocab, et, at, **config)
data_dev = make_dataset_for_variational_hcn(dev_dialogs, dev_indices,
vocab, et, at, **config)
data_test = make_dataset_for_variational_hcn(test_dialogs, test_indices,
vocab, et, at, **config)
random_input = generate_random_input_for_variational_hcn(
10000, config['max_sequence_length'], vocab, rev_vocab)
save_model(rev_vocab, config, kb, at.action_templates, in_model_folder)
trainer = Trainer(data_train, data_dev, data_test, at.action_templates,
random_input, rev_vocab, config, in_model_folder)
trainer.train()
def main(in_dataset_folder, in_result_file, in_max_size, in_config):
with open(in_config, encoding='utf-8') as config_in:
config = json.load(config_in)
w2v_model = KeyedVectors.load_word2vec_format(
UtteranceEmbed.DEFAULT_W2V_FILE, binary=True)
kb = make_augmented_knowledge_base(
os.path.join(in_dataset_folder, 'dialog-babi-task6-dstc2-kb.txt'),
os.path.join(in_dataset_folder,
'dialog-babi-task6-dstc2-candidates.txt'))
vocab = [PAD, START, UNK, EOS] + list(kb.keys())
vocab += [
word for word, vocab_entry in sorted(
w2v_model.vocab.items(), key=lambda x: x[1].count, reverse=True)
[:in_max_size - len(vocab)]
]
with open(in_result_file, 'w') as vocab_out:
for word in vocab:
print(word, file=vocab_out)
def main(in_dataset_folder, in_noisy_dataset_folder, in_custom_vocab_file,
in_model_folder, in_config):
with open(in_config, encoding='utf-8') as config_in:
config = json.load(config_in)
train_json = load_hcn_json(os.path.join(in_dataset_folder, 'train.json'))
dev_json = load_hcn_json(os.path.join(in_dataset_folder, 'dev.json'))
# test_json = load_hcn_json(os.path.join(in_dataset_folder, 'test.json'))
test_ood_json = load_hcn_json(
os.path.join(in_noisy_dataset_folder, 'test_ood.json'))
kb = make_augmented_knowledge_base(
os.path.join(BABI_FOLDER, 'dialog-babi-task6-dstc2-kb.txt'),
os.path.join(BABI_FOLDER, 'dialog-babi-task6-dstc2-candidates.txt'))
action_templates = train_json['actions']
max_noisy_dialog_length = max(
[len(dialog['turns']) for dialog in test_ood_json['dialogs']])
config['max_input_length'] = max_noisy_dialog_length
et = EntityTracker(kb)
post_ood_turns_clean, post_ood_turns_noisy = mark_post_ood_turns(
test_ood_json)
if in_custom_vocab_file is not None:
with open(in_custom_vocab_file) as vocab_in:
rev_vocab = [line.rstrip() for line in vocab_in]
vocab = {word: idx for idx, word in enumerate(rev_vocab)}
else:
utterances_tokenized = []
for dialog in train_json['dialogs']:
for utterance in dialog['turns']:
utterances_tokenized.append(utterance['input'].split())
vocab, rev_vocab = make_vocabulary(
utterances_tokenized,
config['max_vocabulary_size'],
special_tokens=[PAD, START, UNK, EOS] + list(kb.keys()))
ctx_features = []
for dialog in train_json['dialogs']:
for utterance in dialog['turns']:
if 'context_features' in utterance:
ctx_features.append(utterance['context_features'])
ctx_features_vocab, ctx_features_rev_vocab = make_vocabulary(
ctx_features, config['max_vocabulary_size'], special_tokens=[])
config['vocabulary_size'] = len(vocab)
print('Training with config: {}'.format(json.dumps(config)))
data_preparation_function = getattr(utils.preprocessing,
config['data_preparation_function'])
data_train = data_preparation_function(train_json, vocab,
ctx_features_vocab, et, **config)
data_dev = data_preparation_function(dev_json, vocab, ctx_features_vocab,
et, **config)
# data_test = data_preparation_function(test_json, vocab, ctx_features_vocab, et, **config)
data_test_ood = data_preparation_function(test_ood_json, vocab,
ctx_features_vocab, et, **config)
dropout_turn_generation_function = getattr(
utils.preprocessing, config['dropout_turn_generation_function'])
random_input = dropout_turn_generation_function(
10000, 3, config['max_sequence_length'], train_json, vocab,
config['turn_word_dropout_prob'])
save_model(rev_vocab, config, kb, action_templates, in_model_folder)
net = getattr(modules, config['model_name'])(vocab, config,
len(ctx_features_vocab),
len(action_templates))
trainer = Trainer(data_train, data_dev, data_test_ood, action_templates,
random_input, post_ood_turns_noisy, config, net,
in_model_folder)
trainer.train()