-
Notifications
You must be signed in to change notification settings - Fork 7
/
default_pi.py
305 lines (249 loc) · 10.6 KB
/
default_pi.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
import time
import torch
import numpy as np
from value_func import ValueFunc
class E2EHeuristic:
def __init__(self,
equation_env,
rl_env,
model,
k,
num_beams,
horizon,
device,
use_seq_cache,
use_prefix_cache,
length_penalty,
train_value_mode=False,
value_func=None,
debug=False):
self.model = model
self.rl_env = rl_env
self.equation_env = equation_env
self.k = k
self.num_beams = num_beams
self.horizon = horizon
self.device = device
self.length_penalty = length_penalty
self.debug = debug
self.use_seq_cache = use_seq_cache
self.use_prefix_cache = use_prefix_cache
self.train_value_mode = train_value_mode
if self.train_value_mode:
# fixme hardcoded state dimension
self.value_func = ValueFunc(state_size=1600, device=self.device)
if self.use_seq_cache:
self.use_seq_cache= False
print("need to turn off use_seq_cache, otherwise some training data are not collected.")
if value_func is not None:
self.value_func = value_func
self.use_value_mode = True
else:
self.use_value_mode = False
self.output_hash = []
self.top_k_hash = {}
self.sample_times = 0
self.candidate_programs = []
self.terminal_token = self.equation_env.equation_word2id['<EOS>']
@property
def is_train_value_mode(self):
return self.train_value_mode
@property
def is_use_value_mode(self):
return self.use_value_mode
def get_predict_sequence(self, state, ret_states=False):
"""
Args:
ret_states: Return the hidden states of the Transformer in the generation process.
Only used to train a value function so far.
Returns:
Get the most likely sequence starting from state.
"""
with torch.no_grad():
encoded_ids = state
input_ids = torch.LongTensor(encoded_ids).unsqueeze(0).to(self.device)
if self.use_seq_cache and self.num_beams == 1:
# If no beam search is used, if the prefix of a previously generated sequences generated state matches
# state, Transformer will generate the exact sequence. So use cache.
for cached_ids in self.output_hash:
if encoded_ids == cached_ids[:len(encoded_ids)]:
if self.debug: print('sequence cache hit')
return cached_ids
start_time = time.time()
generated_hyps, top_k_hash_updated = self.model.generate_beams(
input_ids,
top_k=self.k,
num_beams=self.num_beams,
length_penalty = self.length_penalty,
early_stopping=True,
max_length=self.horizon,
top_k_hash = self.top_k_hash,
use_prefix_cache = self.use_prefix_cache
)
self.top_k_hash = top_k_hash_updated
output_ids_list = []
for b in range(self.num_beams):
output_ids_list.append(generated_hyps[0].hyp[b][1])
if len(output_ids_list) > 1:
# if got multiple output_ids using beam search, pick the one that has the highest reward
cand_rewards = [self.rl_env.get_reward(output_ids) for output_ids in output_ids_list]
output_ids = output_ids_list[np.argmax(cand_rewards)]
else:
output_ids = output_ids_list[0]
if self.use_seq_cache:
self.output_hash.append(output_ids.tolist())
self.sample_times += 1
self.candidate_programs.append(output_ids)
if self.train_value_mode and ret_states:
return output_ids, last_layers
else:
return output_ids
def get_top_k_predict(self, state):
"""
Returns:
A list of k most likely tokens generate in state (descending in their scores)
"""
with torch.no_grad():
if self.use_prefix_cache:
if tuple(state) in self.top_k_hash:
if self.debug: print('top-k cache hit')
return self.top_k_hash[tuple(state)]
encoded_ids = state
input_ids = torch.LongTensor(encoded_ids).unsqueeze(0).to(self.device)
start_time = time.time()
top_k_tokens = self.model.top_k(input_ids,top_k = self.k)
top_k_tokens = top_k_tokens.tolist()[0]
if self.use_prefix_cache:
self.top_k_hash[tuple(state)] = top_k_tokens
return top_k_tokens
def train_value_func(self, states, value):
self.value_func.train(states, value)
def update_cache(self, new_state):
if self.use_seq_cache:
# clear hashed sequences that are not consistent with new_state
self.output_hash = list(filter(lambda x: new_state == x[:len(new_state)], self.output_hash))
if self.use_prefix_cache:
new_state = tuple(new_state)
keys_to_remove = []
for cached_key in self.top_k_hash:
if cached_key[:len(new_state)] != new_state:
keys_to_remove.append(cached_key)
for k in keys_to_remove: del self.top_k_hash[k]
class NesymresHeuristic:
def __init__(self,
rl_env,
model,
k,
num_beams,
horizon,
device,
use_seq_cache,
use_prefix_cache,
length_penalty,
cfg_params,
train_value_mode=False,
value_func=None,
debug=False):
self.model = model
self.rl_env = rl_env
self.k = k
self.num_beams = num_beams
self.horizon = horizon
self.device = device
self.length_penalty = length_penalty
self.cfg_params = cfg_params
self.use_seq_cache = use_seq_cache
self.use_prefix_cache = use_prefix_cache
self.train_value_mode = train_value_mode
self.debug = debug
if self.train_value_mode:
# fixme hardcoded state dimension
self.value_func = ValueFunc(state_size=1600, device=self.device)
if self.use_seq_cache:
self.use_seq_cache= False
print("need to turn off use_seq_cache, otherwise some training data are not collected.")
if value_func is not None:
self.value_func = value_func
self.use_value_mode = True
else:
self.use_value_mode = False
self.output_hash = []
self.sample_times = 0
self.candidate_programs = []
self.terminal_token = cfg_params.word2id["F"]
@property
def is_train_value_mode(self):
return self.train_value_mode
@property
def is_use_value_mode(self):
return self.use_value_mode
def get_predict_sequence(self, state, ret_states=False):
"""
Args:
ret_states: Return the hidden states of the Transformer in the generation process.
Only used to train a value function so far.
Returns:
Get the most likely sequence starting from state.
"""
with torch.no_grad():
encoded_ids = state
input_ids = torch.LongTensor(encoded_ids).unsqueeze(0).to(self.device)
if self.use_seq_cache and self.num_beams == 1:
# If no beam search is used, if the prefix of a previously generated sequences generated state matches
# state, Transformer will generate the exact sequence. So use cache.
for cached_ids in self.output_hash:
if encoded_ids == cached_ids[:len(encoded_ids)]:
if self.debug: print('sequence cache hit')
return cached_ids
start_time = time.time()
generated_hyps = self.model.generate_beam_from_state(
input_ids,
self.num_beams,
self.cfg_params
)
# print('generate sequence time: ' + str(time.time() - start_time))
output_ids_list = []
for b in range(self.num_beams):
output_ids_list.append(generated_hyps.hyp[b][1])
if len(output_ids_list) > 1:
# if got multiple output_ids using beam search, pick the one that has the highest reward
cand_rewards = [self.rl_env.get_reward(output_ids) for output_ids in output_ids_list]
output_ids = output_ids_list[np.argmax(cand_rewards)]
else:
output_ids = output_ids_list[0]
if self.use_seq_cache:
self.output_hash.append(output_ids.tolist())
# breakpoint()
self.sample_times += 1
self.candidate_programs.append(output_ids)
if self.train_value_mode and ret_states:
return output_ids, last_layers
else:
return output_ids
def get_top_k_predict(self, state):
"""
Returns:
A list of k most likely tokens generate in state (descending in their scores)
"""
with torch.no_grad():
if self.use_prefix_cache:
top_k_tokens = self.model.top_k_hash.get(state)
if top_k_tokens is not None:
if self.debug: print('top-k cache hit')
return top_k_tokens
encoded_ids = state
input_ids = torch.LongTensor(encoded_ids).unsqueeze(0).to(self.device)
start_time = time.time()
top_k_tokens = self.model.extract_top_k(input_ids,top_k = self.k)
top_k_tokens = top_k_tokens.tolist()[0]
return top_k_tokens
def train_value_func(self, states, value):
self.value_func.train(states, value)
def update_cache(self, new_state):
if self.use_seq_cache:
# clear hashed sequences that are not consistent with new_state
self.output_hash = list(filter(lambda x: new_state == x[:len(new_state)], self.output_hash))
if self.use_prefix_cache:
# clear hashed key, value pairs that are not consistent with new_state
self.model.prefix_key_values.clear(new_state)
self.model.top_k_hash.clear(new_state)