def create_demo_batch(sentences, dataset_type, vocab,
attribute_name, target_columns, tmp_path='/tmp'):
'''
Args:
sentences: List of string.
'''
# Create a temporary file.
tmp_path = os.path.join(tmp_path, common.random_string(5))
index = [i for i in xrange(len(sentences))]
dic = {
'index': index,
'sentence': sentences,
}
for col in target_columns:
dic[col] = [EMPTY]
df = pd.DataFrame(dic).ix[:, ['index', 'sentence'] + target_columns].set_index('index')
sys.stdout = sys.stderr
with open(tmp_path, 'w') as f:
f.write(df.to_csv() + '\n')
pathes = common.dotDict({'train': tmp_path, 'valid':tmp_path, 'test':tmp_path})
num_training_sentences = 0 # Fake value.
dataset = getattr(self_module, dataset_type)(dataset_type, pathes, num_training_sentences, vocab, attribute_name, target_columns)
dataset.test.load_data()
os.system('rm %s' % tmp_path)
return dataset.test
def main(args):
if args.mode == 'train':
sys.stderr.write('Saving config...\n')
config = common.dotDict(args.__dict__)
save_config(args)
else:
sys.stderr.write('Loading config...\n')
config = load_config(args)
model = getattr(myself, config.model_type)(args, config)
if args.mode == 'train':
model.train()
elif args.mode == 'test':
tests, origins = read_human_annotations(args.test_file)
lines = [line for idx, line, anno in tests]
predictions, cluster_ids = model.test(lines,
test_filepath=args.test_file)
model.evaluate(tests, origins, predictions, cluster_ids=cluster_ids)
elif args.mode == 'evaluate':
tests, origins = read_human_annotations(args.test_file)
predictions = read_dplabels()
model.evaluate(tests, origins, predictions)
else:
raise ValueError('args.mode must be \'train\' or \'test\'.')
def setup_embeddings(self, config, vocab):
self.embeddings = dotDict()
n_start_vocab = len(vocab.e_word.start_vocab)
special_tokens_emb = self.initialize_embeddings(
'SpecialTokens',
vocab.e_word.embeddings[:n_start_vocab].shape,
initializer=tf.constant_initializer(
vocab.e_word.embeddings[:n_start_vocab]),
trainable=True)
# e_words_emb = tf.constant(vocab.e_word.embeddings[n_start_vocab:],
# dtype=tf.float32)
# j_words_emb = tf.constant(vocab.j_word.embeddings[n_start_vocab:],
# dtype=tf.float32)
e_words_emb = self.initialize_embeddings(
'EnWords',
vocab.e_word.embeddings[n_start_vocab:].shape,
initializer=tf.constant_initializer(
vocab.e_word.embeddings[n_start_vocab:]),
trainable=config.train_embedding)
j_words_emb = self.initialize_embeddings(
'JPWords',
vocab.j_word.embeddings[n_start_vocab:].shape,
initializer=tf.constant_initializer(
vocab.j_word.embeddings[n_start_vocab:]),
trainable=config.train_embedding)
self.embeddings.e_word = tf.concat([special_tokens_emb, e_words_emb],
axis=0)
self.embeddings.j_word = tf.concat([special_tokens_emb, j_words_emb],
axis=0)
def __init__(self, args, sess, vocab=None):
self.sess = sess
self.config = self.load_config(args)
self.logger = common.logManager(handler=FileHandler(args.log_file)) if args.log_file else common.logManager()
sys.stderr.write(str(self.config) + '\n')
data_class = getattr(datasets, self.config.dataset_type)
self.vocab = common.dotDict()
if self.config.embeddings:
emb_conf = self.config.embeddings
self.vocab.e_word = vocabularies.WordVocabularyWithEmbedding(
emb_conf.en.path,
vocab_size=self.config.w_vocab_size,
lowercase=self.config.lowercase,
normalize_digits=self.config.normalize_digits,
skip_first=emb_conf.en.skip_first)
self.vocab.j_word = vocabularies.WordVocabularyWithEmbedding(
emb_conf.ja.path,
vocab_size=self.config.w_vocab_size,
lowercase=self.config.lowercase,
normalize_digits=self.config.normalize_digits,
skip_first=emb_conf.ja.skip_first)
self.c_vocab = None
#self.w_vocab, self.c_vocab = data_class.create_vocab_from_data(self.config)
self.dataset = data_class(self.config.dataset_info,
self.vocab.e_word, self.c_vocab)
def yield_batch(self, batch_by_column):
b_sources, b_targets, b_ori_sources, b_pos= batch_by_column
b_targets = list(zip(*b_targets)) # to column-major.
return common.dotDict({
'sources': np.array(b_sources),
# Include only the labels in 'target_columns' to batch.
'targets': [np.array(t) for t, col in zip(b_targets, self.all_columns)
if col in self.target_columns],
'original_sources': b_ori_sources,
'pos': b_pos,
})
def load_data(self):
self.load = True
sys.stderr.write('Loading dataset from %s ...\n' % (self.path))
df = pd.read_csv(self.path, nrows=self.max_lines)
sys.stderr.write('Preprocessing ...\n')
contexts, responses, speaker_changes = self.preprocess(df)
if not self.wbase and not self.cbase:
raise ValueError('Either \'wbase\' or \'cbase\' must be True.')
self.speaker_changes = [self.sc_vocab.sent2id(sc) for sc in speaker_changes]
# Separate contexts and responses into words (or chars), and convert them into their IDs.
self.original = common.dotDict({})
self.symbolized = common.dotDict({})
if self.wbase:
self.original.w_contexts = [[self.w_vocab.tokenizer(u) for u in context]
for context in contexts]
self.symbolized.w_contexts = [[self.w_vocab.sent2id(u) for u in context]
for context in self.original.w_contexts]
else:
self.original.w_contexts = [None for context in contexts]
self.symbolized.w_contexts = [None for context in contexts]
if self.cbase:
self.original.c_contexts = [[self.c_vocab.tokenizer(u) for u in context]
for context in contexts]
self.symbolized.c_contexts = [[self.c_vocab.sent2id(u) for u in context]
for context in self.original.c_contexts]
else:
self.original.c_contexts = [None for context in contexts]
self.symbolized.c_contexts = [None for context in contexts]
self.original.responses = [self.w_vocab.tokenizer(r) for r in responses]
self.symbolized.responses = [self.w_vocab.sent2id(r) for r in responses]
responses = self.symbolized.responses
w_contexts = self.symbolized.w_contexts
self.texts = common.flatten(w_contexts) + list(responses)
def yield_batch(self, batch):
'''
Args
- batch: A list of a list containing 'batch_size' examples (specified as an argument to get_batch()), batch[i] contains each of the return values of get_batch_data(). (i.e. the shape of 'batch' = [len(self.get_batch_data(...)), batch_size]).
Return : A batch as a dictionary.
'''
b_sources, b_targets, b_ori_sources = batch
b_targets = list(zip(*b_targets)) # to column-major.
return common.dotDict({
'sources': np.array(b_sources),
# Include only the labels in 'target_columns' to batch.
'targets': [np.array(t) for t, col in zip(b_targets, self.all_columns) if col in self.target_columns],
'original_sources': b_ori_sources,
})
def get_batch(self,
batch_size,
input_max_len=None,
output_max_len=None,
shuffle=False):
sources, targets = self.symbolized
if input_max_len:
paired = [(s, t) for s, t in zip(sources, targets)
if not len(s) > input_max_len]
sources, targets = list(zip(*paired))
sources = tf.keras.preprocessing.sequence.pad_sequences(
sources,
maxlen=input_max_len,
padding='post',
truncating='post',
value=PAD_ID)
targets = list(zip(*targets)) # to column-major. (for padding)
targets = [
tf.keras.preprocessing.sequence.pad_sequences(
targets_by_column,
maxlen=output_max_len,
padding='post',
truncating='post',
value=PAD_ID) for targets_by_column in targets
]
targets = list(zip(*targets)) # to idx-major. (for shuffling)
data = [
tuple(x)
for x in zip(sources, targets, self.original_sources, self.targets)
]
if shuffle:
random.shuffle(data)
for i, b in itertools.groupby(enumerate(data), lambda x: x[0] //
(batch_size)):
batch = [x[1] for x in b]
b_sources, b_targets, b_ori_sources, b_ori_targets = zip(*batch)
b_targets = list(zip(*b_targets)) # to column-major.
yield common.dotDict({
'sources': np.array(b_sources),
'targets': [np.array(t) for t in b_targets],
'original_sources': b_ori_sources,
'original_targets': b_ori_targets,
})
def load_config(args):
if os.path.exists(os.path.join(args.output_dir, CONFIG_NAME + '.txt')):
config = collections.defaultdict()
for l in open(os.path.join(args.output_dir, CONFIG_NAME + '.txt')):
k, v, type_name = l.replace('\n', '').split('\t')
if type_name == 'tuple':
config[k] = common.str2tuple(v)
elif type_name == 'int':
config[k] = int(v)
elif type_name == 'float':
config[k] = float(v)
else:
config[k] = v
config = common.dotDict(config)
else:
raise ValueError('No config file is found.')
sys.stderr.write(str(config)+'\n')
return config
def create_demo_batch(self, text, output_max_len):
source = [self.vocab.tokens2ids(text)]
targets = [[[0] for _ in self.targets_name]]
targets = list(zip(*targets)) # to column-major. (for padding)
source = tf.keras.preprocessing.sequence.pad_sequences(
source, padding='post', truncating='post', value=PAD_ID)
targets = [
tf.keras.preprocessing.sequence.pad_sequences(
targets_by_column,
maxlen=output_max_len,
padding='post',
truncating='post',
value=PAD_ID) for targets_by_column in targets
]
yield common.dotDict({
'sources': np.array(source),
'targets': [np.array(t) for t in targets],
})
def __init__(self, args, sess, vocab=None):
self.sess = sess
self.config = config = self.get_config(args)
self.mode = args.mode
self.logger = common.logManager(handler=FileHandler(
args.log_file)) if args.log_file else common.logManager()
sys.stderr.write(str(self.config) + '\n')
# Lazy loading.
self.dataset = common.dotDict({
'train': None,
'valid': None,
'test': None
})
self.dataset_type = getattr(datasets, config.dataset_type)
if not args.interactive: # For saving time when running in jupyter.
self.vocab = WordVocabularyWithEmbedding(
config.embeddings,
vocab_size=config.vocab_size,
lowercase=config.lowercase) if vocab is None else vocab
def __init__(self, args, sess, vocab=None):
self.sess = sess
self.config = self.load_config(args)
self.mode = args.mode
self.logger = common.logManager(handler=FileHandler(
args.log_file)) if args.log_file else common.logManager()
sys.stderr.write(str(self.config) + '\n')
if True or not args.interactive:
self.vocab = common.dotDict()
self.vocab.word = WordVocabularyWithEmbedding(
self.config.embeddings,
vocab_size=self.config.vocab_size,
lowercase=self.config.lowercase,
normalize_digits=self.config.normalize_digits,
) if vocab is None else vocab
self.dataset = getattr(datasets, self.config.dataset_type)(
self.config.dataset_type, self.config.dataset_path,
self.config.num_train_data, self.vocab,
self.config.target_attribute, self.config.target_columns)
def get_batch(self, batch_size, word_max_len=0,
utterance_max_len=0, shuffle=False):
if not self.load:
self.load_data() # lazy loading.
data = self.texts
if shuffle: # For training.
random.shuffle(data)
for i, b in itertools.groupby(enumerate(data),
lambda x: x[0] // batch_size):
batch = [x[1] for x in b]
texts = batch
_utterance_max_len_data = max([len(u) for u in texts])
if not utterance_max_len or _utterance_max_len_data < utterance_max_len:
_utterance_max_len = _utterance_max_len_data
else:
_utterance_max_len = utterance_max_len
texts = np.array(texts)
texts = tf.keras.preprocessing.sequence.pad_sequences(
texts, maxlen=_utterance_max_len,
padding='post', truncating='post', value=PAD_ID)
yield common.dotDict({
'texts': texts
})
