Skip to content
This repository was archived by the owner on Jan 1, 2021. It is now read-only.

Commit 34555e0

Browse files
chiphuyenchiphuyen
committed
starter code for the chatbot assignment
1 parent b8a3cbb commit 34555e0

5 files changed

Lines changed: 766 additions & 1 deletion

File tree

‎assignments/chatbot/README.md‎

Lines changed: 45 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1 +1,45 @@
1-
You can see sample conversations in output_convo.txt
1+
A neural chatbot using sequence to sequence model with
2+
attentional decoder.
3+
4+
This is based on Google Translate Tensorflow model
5+
https://github.com/tensorflow/models/blob/master/tutorials/rnn/translate/
6+
7+
Sequence to sequence model by Cho et al.(2014)
8+
9+
Created by Chip Huyen as the starter code for assignment 3,
10+
class CS 20SI: "TensorFlow for Deep Learning Research"
11+
cs20si.stanford.edu
12+
13+
The detailed assignment handout can be found at: (not updated)
14+
15+
See output_convo.txt for sample conversations.
16+
17+
Usage:
18+
Step 1: create a data folder in your project directory, download
19+
the Cornell Movie-Dialogs Corpus from
20+
https://www.cs.cornell.edu/~cristian/Cornell_Movie-Dialogs_Corpus.html
21+
Unzip it
22+
23+
Step 2: python data.py
24+
This will do all the pre-processing for the Cornell dataset.
25+
26+
Step 3:
27+
python chatbot.py --mode [train/chat]
28+
If mode is train, then you train the chatbot. By default, the model will
29+
restore the previously trained weights (if there is any) and continue
30+
training up on that.
31+
32+
If you want to start training from scratch, please delete all the checkpoints
33+
in the checkpoints folder.
34+
35+
If the mode is chat, you'll go into the interaction mode with the bot.
36+
37+
By default, all the conversations you have with the chatbot will be written
38+
into the file output_convo.txt in the processed folder. If you run this chatbot,
39+
I kindly ask you to send me the output_convo.txt so that I can improve
40+
the chatbot. My email is huyenn@stanford.edu
41+
42+
If you find the tutorial helpful, please head over to Anonymous Chatlog Donation
43+
to see how you can help us create the first realistic dialogue dataset.
44+
45+
Thank you very much!

‎assignments/chatbot/chatbot.py‎

Lines changed: 262 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,262 @@
1+
""" A neural chatbot using sequence to sequence model with
2+
attentional decoder.
3+
4+
This is based on Google Translate Tensorflow model
5+
https://github.com/tensorflow/models/blob/master/tutorials/rnn/translate/
6+
7+
Sequence to sequence model by Cho et al.(2014)
8+
9+
Created by Chip Huyen as the starter code for assignment 3,
10+
class CS 20SI: "TensorFlow for Deep Learning Research"
11+
cs20si.stanford.edu
12+
13+
This file contains the code to run the model.
14+
15+
See readme.md for instruction on how to run the starter code.
16+
"""
17+
from __future__ import division
18+
from __future__ import print_function
19+
20+
import argparse
21+
import os
22+
import random
23+
import sys
24+
import time
25+
26+
import numpy as np
27+
import tensorflow as tf
28+
29+
from model import ChatBotModel
30+
import config
31+
import data
32+
33+
def _get_random_bucket(train_buckets_scale):
34+
""" Get a random bucket from which to choose a training sample """
35+
rand = random.random()
36+
return min([i for i in xrange(len(train_buckets_scale))
37+
if train_buckets_scale[i] > rand])
38+
39+
def _assert_lengths(encoder_size, decoder_size, encoder_inputs, decoder_inputs, decoder_masks):
40+
""" Assert that the encoder inputs, decoder inputs, and decoder masks are
41+
of the expected lengths """
42+
if len(encoder_inputs) != encoder_size:
43+
raise ValueError("Encoder length must be equal to the one in bucket,"
44+
" %d != %d." % (len(encoder_inputs), encoder_size))
45+
if len(decoder_inputs) != decoder_size:
46+
raise ValueError("Decoder length must be equal to the one in bucket,"
47+
" %d != %d." % (len(decoder_inputs), decoder_size))
48+
if len(decoder_masks) != decoder_size:
49+
raise ValueError("Weights length must be equal to the one in bucket,"
50+
" %d != %d." % (len(decoder_masks), decoder_size))
51+
52+
def run_step(sess, model, encoder_inputs, decoder_inputs, decoder_masks, bucket_id, forward_only):
53+
""" Run one step in training.
54+
@forward_only: boolean value to decide whether a backward path should be created
55+
forward_only is set to True when you just want to evaluate on the test set,
56+
or when you want to the bot to be in chat mode. """
57+
encoder_size, decoder_size = config.BUCKETS[bucket_id]
58+
_assert_lengths(encoder_size, decoder_size, encoder_inputs, decoder_inputs, decoder_masks)
59+
60+
# input feed: encoder inputs, decoder inputs, target_weights, as provided.
61+
input_feed = {}
62+
for step in xrange(encoder_size):
63+
input_feed[model.encoder_inputs[step].name] = encoder_inputs[step]
64+
for step in xrange(decoder_size):
65+
input_feed[model.decoder_inputs[step].name] = decoder_inputs[step]
66+
input_feed[model.decoder_masks[step].name] = decoder_masks[step]
67+
68+
last_target = model.decoder_inputs[decoder_size].name
69+
input_feed[last_target] = np.zeros([model.batch_size], dtype=np.int32)
70+
71+
# output feed: depends on whether we do a backward step or not.
72+
if not forward_only:
73+
output_feed = [model.train_ops[bucket_id], # update op that does SGD.
74+
model.gradient_norms[bucket_id], # gradient norm.
75+
model.losses[bucket_id]] # loss for this batch.
76+
else:
77+
output_feed = [model.losses[bucket_id]] # loss for this batch.
78+
for step in xrange(decoder_size): # output logits.
79+
output_feed.append(model.outputs[bucket_id][step])
80+
81+
outputs = sess.run(output_feed, input_feed)
82+
if not forward_only:
83+
return outputs[1], outputs[2], None # Gradient norm, loss, no outputs.
84+
else:
85+
return None, outputs[0], outputs[1:] # No gradient norm, loss, outputs.
86+
87+
def _get_buckets():
88+
""" Load the dataset into buckets based on their lengths.
89+
train_buckets_scale is the inverval that'll help us
90+
choose a random bucket later on.
91+
"""
92+
test_buckets = data.load_data('test_ids.enc', 'test_ids.dec')
93+
data_buckets = data.load_data('train_ids.enc', 'train_ids.dec')
94+
train_bucket_sizes = [len(data_buckets[b]) for b in xrange(len(config.BUCKETS))]
95+
print("Number of samples in each bucket:\n", train_bucket_sizes)
96+
train_total_size = sum(train_bucket_sizes)
97+
# list of increasing numbers from 0 to 1 that we'll use to select a bucket.
98+
train_buckets_scale = [sum(train_bucket_sizes[:i + 1]) / train_total_size
99+
for i in xrange(len(train_bucket_sizes))]
100+
print("Bucket scale:\n", train_buckets_scale)
101+
return test_buckets, data_buckets, train_buckets_scale
102+
103+
def _get_skip_step(iteration):
104+
""" How many steps should the model train before it saves all the weights. """
105+
if iteration < 100:
106+
return 30
107+
return 100
108+
109+
def _check_restore_parameters(sess, saver):
110+
""" Restore the previously trained parameters if there are any. """
111+
ckpt = tf.train.get_checkpoint_state(os.path.dirname(config.CPT_PATH + '/checkpoint'))
112+
if ckpt and ckpt.model_checkpoint_path:
113+
print("Loading parameters for the Chatbot")
114+
saver.restore(sess, ckpt.model_checkpoint_path)
115+
else:
116+
print("Initializing fresh parameters for the Chatbot")
117+
118+
def _eval_test_set(sess, model, test_buckets):
119+
""" Evaluate on the test set. """
120+
for bucket_id in xrange(len(config.BUCKETS)):
121+
if len(test_buckets[bucket_id]) == 0:
122+
print(" Test: empty bucket %d" % (bucket_id))
123+
continue
124+
start = time.time()
125+
encoder_inputs, decoder_inputs, decoder_masks = data.get_batch(test_buckets[bucket_id],
126+
bucket_id,
127+
batch_size=config.BATCH_SIZE)
128+
_, step_loss, _ = run_step(sess, model, encoder_inputs, decoder_inputs,
129+
decoder_masks, bucket_id, True)
130+
print('Test bucket {}: loss {}, time {}'.format(bucket_id, step_loss, time.time() - start))
131+
132+
def train():
133+
""" Train the bot """
134+
test_buckets, data_buckets, train_buckets_scale = _get_buckets()
135+
# in train mode, we need to create the backward path, so forwrad_only is False
136+
model = ChatBotModel(False, config.BATCH_SIZE)
137+
model.build_graph()
138+
139+
saver = tf.train.Saver()
140+
initial_step = 0
141+
142+
with tf.Session() as sess:
143+
print('Running session')
144+
sess.run(tf.global_variables_initializer())
145+
_check_restore_parameters(sess, saver)
146+
147+
iteration = model.global_step.eval()
148+
total_loss = 0
149+
while True:
150+
skip_step = _get_skip_step(iteration)
151+
bucket_id = _get_random_bucket(train_buckets_scale)
152+
encoder_inputs, decoder_inputs, decoder_masks = data.get_batch(data_buckets[bucket_id],
153+
bucket_id,
154+
batch_size=config.BATCH_SIZE)
155+
start = time.time()
156+
_, step_loss, _ = run_step(sess, model, encoder_inputs, decoder_inputs, decoder_masks, bucket_id, False)
157+
total_loss += step_loss
158+
iteration += 1
159+
160+
if iteration % skip_step == 0:
161+
print('Iter {}: loss {}, time {}'.format(iteration, total_loss/skip_step, time.time() - start))
162+
start = time.time()
163+
total_loss = 0
164+
saver.save(sess, os.path.join(config.CPT_PATH, 'chatbot'), global_step=model.global_step)
165+
if iteration % (10 * skip_step) == 0:
166+
# Run evals on development set and print their loss
167+
_eval_test_set(sess, model, test_buckets)
168+
start = time.time()
169+
sys.stdout.flush()
170+
171+
def _get_user_input():
172+
""" Get user's input, which will be transformed into encoder input later """
173+
print("> ", end="")
174+
sys.stdout.flush()
175+
return sys.stdin.readline()
176+
177+
def _find_right_bucket(length):
178+
""" Find the proper bucket for an encoder input based on its length """
179+
return min([b for b in xrange(len(config.BUCKETS))
180+
if config.BUCKETS[b][0] >= length])
181+
182+
def _construct_response(output_logits, inv_dec_vocab):
183+
""" Construct a response to the user's encoder input.
184+
@output_logits: the outputs from sequence to sequence wrapper.
185+
output_logits is decoder_size np array, each of dim 1 x DEC_VOCAB
186+
187+
This is a greedy decoder - outputs are just argmaxes of output_logits.
188+
"""
189+
print(output_logits[0])
190+
outputs = [int(np.argmax(logit, axis=1)) for logit in output_logits]
191+
# If there is an EOS symbol in outputs, cut them at that point.
192+
if config.EOS_ID in outputs:
193+
outputs = outputs[:outputs.index(config.EOS_ID)]
194+
# Print out sentence corresponding to outputs.
195+
return " ".join([tf.compat.as_str(inv_dec_vocab[output]) for output in outputs])
196+
197+
def chat():
198+
""" in test mode, we don't to create the backward path
199+
"""
200+
_, enc_vocab = data.load_vocab(os.path.join(config.PROCESSED_PATH, 'vocab.enc'))
201+
inv_dec_vocab, _ = data.load_vocab(os.path.join(config.PROCESSED_PATH, 'vocab.dec'))
202+
203+
model = ChatBotModel(True, batch_size=1)
204+
model.build_graph()
205+
206+
saver = tf.train.Saver()
207+
208+
with tf.Session() as sess:
209+
sess.run(tf.global_variables_initializer())
210+
_check_restore_parameters(sess, saver)
211+
output_file = open(os.path.join(config.PROCESSED_PATH, config.OUTPUT_FILE), 'a+')
212+
# Decode from standard input.
213+
max_length = config.BUCKETS[-1][0]
214+
print('Welcome to TensorBro. Say something. Enter to exit. Max length is', max_length)
215+
while True:
216+
line = _get_user_input()
217+
if len(line) > 0 and line[-1] == '\n':
218+
line = line[:-1]
219+
if line == '':
220+
break
221+
output_file.write('HUMAN ++++ ' + line + '\n')
222+
# Get token-ids for the input sentence.
223+
token_ids = data.sentence2id(enc_vocab, str(line))
224+
if (len(token_ids) > max_length):
225+
print('Max length I can handle is:', max_length)
226+
line = _get_user_input()
227+
continue
228+
# Which bucket does it belong to?
229+
bucket_id = _find_right_bucket(len(token_ids))
230+
# Get a 1-element batch to feed the sentence to the model.
231+
encoder_inputs, decoder_inputs, decoder_masks = data.get_batch([(token_ids, [])],
232+
bucket_id,
233+
batch_size=1)
234+
# Get output logits for the sentence.
235+
_, _, output_logits = run_step(sess, model, encoder_inputs, decoder_inputs,
236+
decoder_masks, bucket_id, True)
237+
response = _construct_response(output_logits, inv_dec_vocab)
238+
print(response)
239+
output_file.write('BOT ++++ ' + response + '\n')
240+
output_file.write('=============================================\n')
241+
output_file.close()
242+
243+
def main():
244+
parser = argparse.ArgumentParser()
245+
parser.add_argument('--mode', choices={'train', 'chat'},
246+
default='train', help="mode. if not specified, it's in the train mode")
247+
args = parser.parse_args()
248+
249+
if not os.path.isdir(config.PROCESSED_PATH):
250+
data.prepare_raw_data()
251+
data.process_data()
252+
print('Data ready!')
253+
# create checkpoints folder if there isn't one already
254+
data.make_dir(config.CPT_PATH)
255+
256+
if args.mode == 'train':
257+
train()
258+
elif args.mode == 'chat':
259+
chat()
260+
261+
if __name__ == '__main__':
262+
main()

‎assignments/chatbot/config.py‎

Lines changed: 60 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,60 @@
1+
""" A neural chatbot using sequence to sequence model with
2+
attentional decoder.
3+
4+
This is based on Google Translate Tensorflow model
5+
https://github.com/tensorflow/models/blob/master/tutorials/rnn/translate/
6+
7+
Sequence to sequence model by Cho et al.(2014)
8+
9+
Created by Chip Huyen as the starter code for assignment 3,
10+
class CS 20SI: "TensorFlow for Deep Learning Research"
11+
cs20si.stanford.edu
12+
13+
This file contains the hyperparameters for the model.
14+
15+
See readme.md for instruction on how to run the starter code.
16+
"""
17+
18+
# parameters for processing the dataset
19+
DATA_PATH = 'data/cornell movie-dialogs corpus'
20+
CONVO_FILE = 'movie_conversations.txt'
21+
LINE_FILE = 'movie_lines.txt'
22+
OUTPUT_FILE = 'output_convo.txt'
23+
PROCESSED_PATH = 'processed'
24+
CPT_PATH = 'checkpoints'
25+
26+
THRESHOLD = 2
27+
28+
PAD_ID = 0
29+
UNK_ID = 1
30+
START_ID = 2
31+
EOS_ID = 3
32+
33+
TESTSET_SIZE = 25000
34+
35+
# model parameters
36+
""" Train encoder length distribution:
37+
[175, 92, 11883, 8387, 10656, 13613, 13480, 12850, 11802, 10165,
38+
8973, 7731, 7005, 6073, 5521, 5020, 4530, 4421, 3746, 3474, 3192,
39+
2724, 2587, 2413, 2252, 2015, 1816, 1728, 1555, 1392, 1327, 1248,
40+
1128, 1084, 1010, 884, 843, 755, 705, 660, 649, 594, 558, 517, 475,
41+
426, 444, 388, 349, 337]
42+
These buckets size seem to work the best
43+
"""
44+
# [19530, 17449, 17585, 23444, 22884, 16435, 17085, 18291, 18931]
45+
# BUCKETS = [(6, 8), (8, 10), (10, 12), (13, 15), (16, 19), (19, 22), (23, 26), (29, 32), (39, 44)]
46+
47+
# [37049, 33519, 30223, 33513, 37371]
48+
# BUCKETS = [(8, 10), (12, 14), (16, 19), (23, 26), (39, 43)]
49+
50+
BUCKETS = [(8, 10), (12, 14), (16, 19)]
51+
52+
NUM_LAYERS = 3
53+
HIDDEN_SIZE = 256
54+
BATCH_SIZE = 64
55+
56+
LR = 0.5
57+
LR_DECAY_FACTOR = 0.99
58+
MAX_GRAD_NORM = 5.0
59+
60+
NUM_SAMPLES = 512

0 commit comments

Comments
 (0)