-
Notifications
You must be signed in to change notification settings - Fork 1
/
play_withSent.py
320 lines (249 loc) · 15.4 KB
/
play_withSent.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
import tensorflow as tf
#import matplotlib.pyplot as plt
tf.enable_eager_execution()
import pdb
import numpy as np
from itertools import count
from seq2seq import Encoder, Decoder
from environment_CoreNLP import Environment, char_tokenizer, BEGIN_TAG, END_TAG, CONVO_LEN
from agent import Baseline
import data
import random
from encoder_decoder import Seq2Seq
from utils import load_trained_model, max_length
from sklearn.metrics.pairwise import cosine_similarity
from corpus_utils import tokenize_sentence
# https://github.com/gabrielgarza/openai-gym-policy-gradient/blob/master/policy_gradient.py
# https://github.com/yaserkl/RLSeq2Seq/blob/7e019e8e8c006f464fdc09e77169680425e83ad1/src/model.py#L348
EPISODES = 10000
BATCH_SIZE = 32
GAMMA = 1
USE_GLOVE = True
if USE_GLOVE:
# 1024 if using glove
EMBEDDING_DIM = 100
else:
# 256 if without pretrained embedding
EMBEDDING_DIM = 100
UNITS = 128
MAX_TARGET_LEN = 20 # TODO: hack
def initialize_hidden_state(batch_sz, num_enc_units):
return tf.zeros((batch_sz, num_enc_units))
def get_returns(r: float, seq_len: int):
return list(reversed([
r * (GAMMA ** t) for t in range(seq_len)
]))
def sentence_to_idxs(sentence: str, lang):
return [lang.word2idx[token]
for token in tokenize_sentence(sentence)]
def maybe_pad_sentence(s):
return tf.keras.preprocessing.sequence.pad_sequences(
s,
maxlen=MAX_TARGET_LEN,
padding='post'
)
def main():
env = Environment()
# print(env.lang.word2idx)
SAY_HI = "hello"
targ_lang = env.lang
vocab_inp_size = len(env.lang.word2idx)
vocab_tar_size = len(targ_lang.word2idx)
# GET WORD SCORES
# sentimental_words = ["absolutely","abundant","accept","acclaimed","accomplishment","achievement","action","active","activist","acumen","adjust","admire","adopt","adorable","adored","adventure","affirmation","affirmative","affluent","agree","airy","alive","alliance","ally","alter","amaze","amity","animated","answer","appreciation","approve","aptitude","artistic","assertive","astonish","astounding","astute","attractive","authentic","basic","beaming","beautiful","believe","benefactor","benefit","bighearted","blessed","bliss","bloom","bountiful","bounty","brave","bright","brilliant","bubbly","bunch","burgeon","calm","care","celebrate","certain","change","character","charitable","charming","cheer","cherish","clarity","classy","clean","clever","closeness","commend","companionship","complete","comradeship","confident","connect","connected","constant","content","conviction","copious","core","coupled","courageous","creative","cuddle","cultivate","cure","curious","cute","dazzling","delight","direct","discover","distinguished","divine","donate","each","day","eager","earnest","easy","ecstasy","effervescent","efficient","effortless","electrifying","elegance","embrace","encompassing","encourage","endorse","energized","energy","enjoy","enormous","enthuse","enthusiastic","entirely","essence","established","esteem","everyday","everyone","excited","exciting","exhilarating","expand","explore","express","exquisite","exultant","faith","familiar","family","famous","feat","fit","flourish","fortunate","fortune","freedom","fresh","friendship","full","funny","gather","generous","genius","genuine","give","glad","glow","good","gorgeous","grace","graceful","gratitude","green","grin","group","grow","handsome","happy","harmony","healed","healing","healthful","healthy","heart","hearty","heavenly","helpful","here","highest","good","hold","holy","honest","honor","hug","i","affirm","i","allow","i","am","willing","i","am.","i","can","i","choose","i","create","i","follow","i","know","i","know,","without","a","doubt","i","make","i","realize","i","take","action","i","trust","idea","ideal","imaginative","increase","incredible","independent","ingenious","innate","innovate","inspire","instantaneous","instinct","intellectual","intelligence","intuitive","inventive","joined","jovial","joy","jubilation","keen","key","kind","kiss","knowledge","laugh","leader","learn","legendary","let","go","light","lively","love","loveliness","lucidity","lucrative","luminous","maintain","marvelous","master","meaningful","meditate","mend","metamorphosis","mind-blowing","miracle","mission","modify","motivate","moving","natural","nature","nourish","nourished","novel","now","nurture","nutritious","one","open","openhanded","optimistic","paradise","party","peace","perfect","phenomenon","pleasure","plenteous","plentiful","plenty","plethora","poise","polish","popular","positive","powerful","prepared","pretty","principle","productive","project","prominent","prosperous","protect","proud","purpose","quest","quick","quiet","ready","recognize","refinement","refresh","rejoice","rejuvenate","relax","reliance","rely","remarkable","renew","renowned","replenish","resolution","resound","resources","respect","restore","revere","revolutionize","rewarding","rich","robust","rousing","safe","secure","see","sensation","serenity","shift","shine","show","silence","simple","sincerity","smart","smile","smooth","solution","soul","sparkling","spirit","spirited","spiritual","splendid","spontaneous","still","stir","strong","style","success","sunny","support","sure","surprise","sustain","synchronized","team","thankful","therapeutic","thorough","thrilled","thrive","today","together","tranquil","transform","triumph","trust","truth","unity","unusual","unwavering","upbeat","value","vary","venerate","venture","very","vibrant","victory","vigorous","vision","visualize","vital","vivacious","voyage","wealthy","welcome","well","whole","wholesome","willing","wonder","wonderful","wondrous","xanadu","yes","yippee","young","youth","youthful","zeal","zest","zing","zip"]
# sentimental_words = ["good", "excellent", "well"]
# targ_lang_embd = get_GloVe_embeddings(targ_lang.vocab, EMBEDDING_DIM)
# sentimental_words_embd = get_GloVe_embeddings(
# sentimental_words, EMBEDDING_DIM)
# sim_scores = np.dot(sentimental_words_embd, np.transpose(targ_lang_embd))
# print(sim_scores.shape)
l_optimizer = tf.train.RMSPropOptimizer(0.001)
bl_optimizer = tf.train.RMSPropOptimizer(0.001)
# LOAD PRETRAINED MODEL HERE
# For now...
# model = load_trained_model(
# BATCH_SIZE, EMBEDDING_DIM, UNITS, tf.train.AdamOptimizer())
'''
encoder = Encoder(vocab_inp_size, EMBEDDING_DIM,
UNITS, batch_sz=BATCH_SIZE, inp_lang=env.lang.vocab)
decoder = Decoder(vocab_tar_size, EMBEDDING_DIM,
UNITS, batch_sz=BATCH_SIZE, targ_lang=targ_lang.vocab)
'''
model = Seq2Seq(
vocab_inp_size, vocab_tar_size, EMBEDDING_DIM, UNITS, BATCH_SIZE,
inp_lang=env.lang, targ_lang=targ_lang,
max_length_tar=MAX_TARGET_LEN,
use_GloVe=USE_GLOVE,
display_result=True,
use_beam_search=False
)
import os
checkpoint_dir = './training_checkpoints'
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt-2")
checkpoint = tf.train.Checkpoint(optimizer=l_optimizer, seq2seq=model)
checkpoint.restore(checkpoint_prefix)
encoder = model.encoder
decoder = model.decoder
# LOAD PRETRAINED MODEL HERE
# For now...
# encoder = Encoder(vocab_inp_size, EMBEDDING_DIM,
# UNITS, batch_sz=BATCH_SIZE, inp_lang=env.lang.vocab)
# decoder = Decoder(vocab_tar_size, EMBEDDING_DIM,
# UNITS, batch_sz=BATCH_SIZE, targ_lang=targ_lang.vocab)
baseline = Baseline(UNITS)
history = []
l_optimizer = tf.train.AdamOptimizer()
bl_optimizer = tf.train.RMSPropOptimizer(0.01)
batch = None
avg_rewards = []
avg_losses = []
for episode in range(EPISODES):
# Start of Episode
env.reset()
# get first state from the env
state, _, done = env.step(SAY_HI)
while not done: # NOT REALLY USING DONE (Conv_length=1)
# Run an episode using the TRAINED ENCODER-DECODER model #TODO: test this!!
init_hidden = initialize_hidden_state(1, UNITS)
state_inp = [env.lang.word2idx[token]
for token in tokenize_sentence(state)]
enc_hidden = encoder(
tf.convert_to_tensor([state_inp]), init_hidden)
dec_hidden = enc_hidden
w = BEGIN_TAG
curr_w_enc = tf.expand_dims(
[targ_lang.word2idx[tokenize_sentence(w)[0]]], 0
)
#pdb.set_trace() ######################################################################################
outputs = []
actions = []
# words_score = 0
while w != END_TAG and len(outputs) < MAX_TARGET_LEN:
w_probs_b, dec_hidden = decoder(curr_w_enc, dec_hidden)
w_dist = tf.distributions.Categorical(probs=w_probs_b[0])
w_idx = w_dist.sample(1)
#pdb.set_trace() ######################################################################################
actions.append(w_idx)
# w_idx = tf.argmax(w_probs[0]).numpy()
w = targ_lang.idx2word[w_idx.numpy()[0]]
#pdb.set_trace() ######################################################################################
# NEW: accumulate score of words in full response
# words_score += np.max(sim_scores[1:, w_idx.numpy()[0]])
curr_w_enc = tf.expand_dims(
[targ_lang.word2idx[w]] * 1, 1)
outputs.append(w)
# action is a sentence (string)
action_str = ' '.join(outputs)
next_state, reward, done = env.step(action_str)
#pdb.set_trace() ######################################################################################
# Reward is sentence score + words score. For now, words score is NOT USED
history.append((state, actions, action_str, reward))
state = next_state
#pdb.set_trace() ######################################################################################
# record history (to be used for gradient updating after the episode is done)
# End of Episode
# Update policy
while len(history) >= BATCH_SIZE:
batch = history[:BATCH_SIZE]
state_inp_b, action_encs_b, reward_b, ret_seq_b = zip(*[
[
sentence_to_idxs(state, env.lang),
actions_enc_b,
reward,
get_returns(reward, MAX_TARGET_LEN)
]
for state, actions_enc_b, _, reward in batch
])
#pdb.set_trace() ######################################################################################
action_encs_b = list(action_encs_b)
action_encs_b = maybe_pad_sentence(action_encs_b)
action_encs_b = tf.expand_dims(
tf.convert_to_tensor(action_encs_b), -1)
ret_mean = np.mean(ret_seq_b)
ret_std = np.std(ret_seq_b)
if ret_std == 0:
ret_seq_b = ret_seq_b - ret_mean
else:
ret_seq_b = (ret_seq_b - ret_mean) / ret_std
ret_seq_b = tf.cast(tf.convert_to_tensor(ret_seq_b), 'float32')
loss = 0
loss_bl = 0
with tf.GradientTape() as l_tape, tf.GradientTape() as bl_tape:
# accumulate gradient with GradientTape
init_hidden_b = initialize_hidden_state(BATCH_SIZE, UNITS)
state_inp_b = maybe_pad_sentence(state_inp_b)
state_inp_b = tf.convert_to_tensor(state_inp_b)
enc_hidden_b = encoder(state_inp_b, init_hidden_b)
dec_hidden_b = enc_hidden_b
max_sentence_len = action_encs_b.numpy().shape[1]
prev_w_idx_b = tf.expand_dims(
tf.cast(
tf.convert_to_tensor(
[env.lang.word2idx[tokenize_sentence(BEGIN_TAG)[0]]] * BATCH_SIZE),
'float32'
), -1
)
#pdb.set_trace() ######################################################################################
for t in range(max_sentence_len):
bl_val_b = baseline(tf.cast(dec_hidden_b, 'float32'))
ret_b = tf.reshape(ret_seq_b[:, t], (BATCH_SIZE, 1))
delta_b = ret_b - bl_val_b
# print(prev_w_idx_b.shape)
w_probs_b, dec_hidden_b = decoder(
tf.cast(prev_w_idx_b, dtype='int32'),
dec_hidden_b
)
curr_w_idx_b = action_encs_b[:, t]
# w_probs_b = tf.nn.softmax(w_logits_b)
dist = tf.distributions.Categorical(probs=w_probs_b)
loss_bl += - \
tf.multiply(delta_b, bl_val_b)
# cost_b = -tf.multiply(
# tf.transpose(dist.log_prob(
# tf.transpose(curr_w_idx_b))), delta_b
# )
#pdb.set_trace() ######################################################################################
cost_b = -tf.multiply(
tf.transpose(dist.log_prob(
tf.transpose(curr_w_idx_b))), ret_b
)
# print(cost_b.shape)
loss += cost_b
prev_w_idx_b = curr_w_idx_b
#pdb.set_trace() ######################################################################################
# calculate cumulative gradients
#pdb.set_trace() ######################################################################################
model_vars = encoder.variables + decoder.variables
grads = l_tape.gradient(loss, model_vars)
grads_bl = bl_tape.gradient(loss_bl, baseline.variables)
# finally, apply gradient
l_optimizer.apply_gradients(zip(grads, model_vars))
bl_optimizer.apply_gradients(zip(grads_bl, baseline.variables))
# Reset everything for the next episode
history = history[BATCH_SIZE:]
if episode % 20 == 0 and batch != None:
print(">>>>>>>>>>>>>>>>>>>>>>>>>>")
print("Episode # ", episode)
print("Samples from episode with rewards > 0: ")
good_rewards = [(s, a_str, r) for s, _, a_str, r in batch]
for s, a, r in random.sample(good_rewards, min(len(good_rewards), 3)):
print("prev_state: ", s)
print("action: ", a)
print("reward: ", r)
# print("return: ", get_returns(r, MAX_TARGET_LEN))
print(
"all returns: min=%f, max=%f, median=%f" %
(np.min(ret_seq_b), np.max(ret_seq_b), np.median(ret_seq_b))
)
avg_reward = sum(reward_b) / len(reward_b)
avg_rewards.append(avg_reward)
print("avg reward: ", avg_reward)
avg_loss = tf.reduce_mean(loss).numpy()
print("avg loss: ", avg_loss)
avg_losses.append(avg_loss)
print("avg grad: ", np.mean(grads[1].numpy()))
# print("<<<<<<<<<<<<<<<<<<<<<<<<<<")
if episode % 200 == 0 and batch != None:
print("Avg rewards: ", avg_rewards)
print("Avg losses:", avg_losses)
main()