def decode(bot, dataset, teacher_mode=True):
"""Runs a chat session between the given chatbot and user."""
# We decode one sentence at a time.
bot.batch_size = 1
# Decode from standard input.
print("Type \"exit\" to exit.")
print(
"Write stuff after the \">\" below and I, your robot friend, will respond."
)
sentence = io_utils.get_sentence()
while sentence:
# Convert input sentence to token-ids.
token_ids = sentence_to_token_ids(tf.compat.as_bytes(sentence),
dataset.word_to_idx)
# Get output sentence from the chatbot.
outputs = decode_inputs(token_ids, dataset.idx_to_word, bot)
# Print the chatbot's response.
print(outputs)
if teacher_mode:
print("What should I have said?")
feedback = io_utils.get_sentence()
feedback_ids = sentence_to_token_ids(tf.compat.as_bytes(feedback),
dataset.inputs_to_word)
outputs = train_on_feedback(bot, token_ids, feedback_ids,
dataset.idx_to_word)
print("Okay. Let me try again:\n", outputs)
# Wait for next input.
sentence = io_utils.get_sentence()
# Stop program if sentence == 'exit\n'.
if sentence == 'exit':
print("Fine, bye :(")
break
def __call__(self, sentence):
"""This is how we talk to the bot interactively.
While decode(self) above sets up/manages the chat session,
users can also use this directly to get responses from the bot,
given an input sentence.
For example, one could do:
sentence = 'Hi, bot!'
response = bot(sentence)
for a single input-to-response with the bot.
Args:
sentence: (str) Input sentence from user.
Returns:
response string from bot.
"""
# Convert input sentence to token-ids.
encoder_inputs = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(sentence), self.dataset.word_to_idx)
encoder_inputs = np.array([encoder_inputs[::-1]])
self.pipeline.feed_user_input(encoder_inputs)
# Get output sentence from the chatbot.
_, _, response = self.step(forward_only=True)
# response has shape [1, response_length].
# Its last element is the EOS_ID, which we don't show user.
response = self.dataset.as_words(response[0][:-1])
if 'UNK' in response:
response = "I don't know."
return response
def __call__(self, sentence):
"""This is how we talk to the bot interactively.
While decode(self) above sets up/manages the chat session,
users can also use this directly to get responses from the bot,
given an input sentence.
For example, one could do:
sentence = 'Hi, bot!'
response = bot(sentence)
for a single input-to-response with the bot.
Args:
sentence: (str) Input sentence from user.
Returns:
response string from bot.
"""
# Convert input sentence to token-ids.
encoder_inputs = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(sentence), self.dataset.word_to_idx)
encoder_inputs = np.array([encoder_inputs[::-1]])
self.pipeline.feed_user_input(encoder_inputs)
# Get output sentence from the chatbot.
_, _, response = self.step(forward_only=True)
# response has shape [1, response_length].
# Its last element is the EOS_ID, which we don't show user.
response = self.dataset.as_words(response[0][:-1])
if 'UNK' in response:
response = "I don't know."
return response
def __call__(self, sentence):
encoder_inputs = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(sentence), self.dataset.word_to_idx)
encoder_inputs = np.array([encoder_inputs[::-1]])
self.pipeline.feed_user_input(encoder_inputs)
# Get output sentence from the chatbot.
response = self.step(forward_only=True)
return self.dataset.as_words(response[0])
def __call__(self, sentence):
encoder_inputs = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(sentence), self.dataset.word_to_idx)
encoder_inputs = np.array([encoder_inputs[::-1]])
self.pipeline.feed_user_input(encoder_inputs)
# Get output sentence from the chatbot.
response = self.step(forward_only=True)
return self.dataset.as_words(response[0])
def __call__(self, sentence):
"""Outputs response sentence (string) given input (string)."""
# Convert input sentence to token-ids.
sentence_tokens = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(sentence), self.word_to_idx)
sentence_tokens = np.array([sentence_tokens[::-1]])
# Get output sentence from the chatbot.
fetches = self.tensor_dict['outputs']
feed_dict = {self.tensor_dict['inputs']: sentence_tokens}
response = self.sess.run(fetches=fetches, feed_dict=feed_dict)
return self.as_words(response[0][:-1])
def __call__(self, sentence):
"""Outputs response sentence (string) given input (string)."""
# Convert input sentence to token-ids.
sentence_tokens = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(sentence), self.word_to_idx)
sentence_tokens = np.array([sentence_tokens[::-1]])
# Get output sentence from the chatbot.
fetches = self.tensor_dict['outputs']
feed_dict={self.tensor_dict['inputs']: sentence_tokens}
response = self.sess.run(fetches=fetches, feed_dict=feed_dict)
return self.as_words(response[0][:-1])
def test_manual_freeze(self):
"""Make sure we can freeze the bot, unfreeze, and still chat."""
# ================================================
# 1. Create & train bot.
# ================================================
flags = TEST_FLAGS
flags = flags._replace(
model_params=dict(ckpt_dir=os.path.join(TEST_DIR, 'out'),
reset_model=True,
steps_per_ckpt=20,
max_steps=40))
bot = create_bot(flags)
self.assertEqual(bot.reset_model, True)
# Simulate small train sesh on bot.
bot.train()
# ================================================
# 2. Recreate a chattable bot.
# ================================================
# Recreate bot from scratch with decode set to true.
logging.info("Resetting default graph . . . ")
tf.reset_default_graph()
flags = flags._replace(
model_params={
**flags.model_params, 'reset_model': False,
'decode': True,
'max_steps': 100,
'steps_per_ckpt': 50
})
self.assertTrue(flags.model_params.get('decode'))
bot = create_bot(flags)
self.assertTrue(bot.is_chatting)
self.assertTrue(bot.decode)
print("Testing quick chat sesh . . . ")
config = io_utils.parse_config(flags=flags)
dataset_class = pydoc.locate(config['dataset']) \
or getattr(data, config['dataset'])
dataset = dataset_class(config['dataset_params'])
test_input = "How's it going?"
encoder_inputs = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(test_input), dataset.word_to_idx)
encoder_inputs = np.array([encoder_inputs[::-1]])
bot.pipeline._feed_dict = {bot.pipeline.user_input: encoder_inputs}
# Get output sentence from the chatbot.
_, _, response = bot.step(forward_only=True)
print("Robot:", dataset.as_words(response[0][:-1]))
# ================================================
# 3. Freeze the chattable bot.
# ================================================
logging.info("Calling bot.freeze() . . . ")
bot.freeze()
# ================================================
# 4. Try to unfreeze and use it.
# ================================================
logging.info("Resetting default graph . . . ")
tf.reset_default_graph()
logging.info("Importing frozen graph into default . . . ")
frozen_graph = bot_freezer.load_graph(bot.ckpt_dir)
logging.info("Extracting input/output tensors.")
tensors, frozen_graph = bot_freezer.unfreeze_bot(bot.ckpt_dir)
self.assertIsNotNone(tensors['inputs'])
self.assertIsNotNone(tensors['outputs'])
with tf.Session(graph=frozen_graph) as sess:
raw_input = "How's it going?"
encoder_inputs = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(raw_input), dataset.word_to_idx)
encoder_inputs = np.array([encoder_inputs[::-1]])
feed_dict = {tensors['inputs'].name: encoder_inputs}
response = sess.run(tensors['outputs'], feed_dict=feed_dict)
logging.info('Reponse: %s', response)
def test_manual_freeze(self):
"""Make sure we can freeze the bot, unfreeze, and still chat."""
# ================================================
# 1. Create & train bot.
# ================================================
flags = TEST_FLAGS
flags = flags._replace(model_params=dict(
ckpt_dir=os.path.join(TEST_DIR, 'out'),
reset_model=True,
steps_per_ckpt=20,
max_steps=40))
bot = create_bot(flags)
self.assertEqual(bot.reset_model, True)
# Simulate small train sesh on bot.
bot.train()
# ================================================
# 2. Recreate a chattable bot.
# ================================================
# Recreate bot from scratch with decode set to true.
logging.info("Resetting default graph . . . ")
tf.reset_default_graph()
flags = flags._replace(model_params={
**flags.model_params,
'reset_model': False,
'decode': True,
'max_steps': 100,
'steps_per_ckpt': 50})
self.assertTrue(flags.model_params.get('decode'))
bot = create_bot(flags)
self.assertTrue(bot.is_chatting)
self.assertTrue(bot.decode)
print("Testing quick chat sesh . . . ")
config = io_utils.parse_config(flags=flags)
dataset_class = pydoc.locate(config['dataset']) \
or getattr(data, config['dataset'])
dataset = dataset_class(config['dataset_params'])
test_input = "How's it going?"
encoder_inputs = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(test_input),
dataset.word_to_idx)
encoder_inputs = np.array([encoder_inputs[::-1]])
bot.pipeline._feed_dict = {
bot.pipeline.user_input: encoder_inputs}
# Get output sentence from the chatbot.
_, _, response = bot.step(forward_only=True)
print("Robot:", dataset.as_words(response[0][:-1]))
# ================================================
# 3. Freeze the chattable bot.
# ================================================
logging.info("Calling bot.freeze() . . . ")
bot.freeze()
# ================================================
# 4. Try to unfreeze and use it.
# ================================================
logging.info("Resetting default graph . . . ")
tf.reset_default_graph()
logging.info("Importing frozen graph into default . . . ")
frozen_graph = bot_freezer.load_graph(bot.ckpt_dir)
logging.info("Extracting input/output tensors.")
tensors, frozen_graph = bot_freezer.unfreeze_bot(bot.ckpt_dir)
self.assertIsNotNone(tensors['inputs'])
self.assertIsNotNone(tensors['outputs'])
with tf.Session(graph=frozen_graph) as sess:
raw_input = "How's it going?"
encoder_inputs = io_utils.sentence_to_token_ids(
tf.compat.as_bytes(raw_input),
dataset.word_to_idx)
encoder_inputs = np.array([encoder_inputs[::-1]])
feed_dict = {tensors['inputs'].name: encoder_inputs}
response = sess.run(tensors['outputs'], feed_dict=feed_dict)
logging.info('Reponse: %s', response)