-
Notifications
You must be signed in to change notification settings - Fork 4
/
node_transition.py
273 lines (247 loc) · 12.5 KB
/
node_transition.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
from transformers import AutoTokenizer, LlamaTokenizer
from typing import List, Optional, Union
from copy import deepcopy
from reward import batch_evaluate
from model import generator as generator_class
import math
"""
Class for Nodes
"""
class Node:
def __init__(self):
self.chains = []
self.parent = None
self.children = []
self.substep_num = -1
self.reward = 0
self.Q = 0
self.sum_reward = 0
self.step_num = 0
self.halt_status = False
self.visited = 1
self.root_distance = 0
self.terminal_distance = 0
def select_opt_id(opt: str, tokenizer: any) -> int:
idx = tokenizer.encode(opt)
flag = False
for id in idx:
if flag:
return id
else:
token = tokenizer.decode(id)
if token == "(":
flag = True
return idx[0]
class node_transition:
def __init__(self, generator: generator_class,
args: any,
max_length: int = 512,
set_edge_links: bool = False):
self.generator = generator
self.max_length = max_length
self.temperature = args.temperature
self.Aid = select_opt_id(opt="(A)", tokenizer=generator.tokenizer)
self.Bid = select_opt_id(opt="(B)", tokenizer=generator.tokenizer)
self.reward = 0
self.prompt_style = args.prompt_style
self.set_edge_links = set_edge_links
self.reward_types = args.reward_types
def prompt_format(self, node):
prompt_state = "".join(node.chains) # combine reasoning steps (includes initial question prompt)
if self.prompt_style in ["struct", "struct_min"]:
# experimental prompt instructed to decompose into a subproblem, then list facts, then solution
# each step have 3 substeps - 0) Subproblem creation 1) list relevant facts 2) Solution
# we treat the last two substeps as a single substep for reward -- thus, substep number varies in a modulo 2 fashion
# struct_min removes the "list relevant facts" substep.
substep_num = int((node.substep_num + 1) % 2)
if substep_num == 0:
step = node.step_num + 1
else:
step = node.step_num # if substep_num non-zero it means we are in the same step (so we keep the same step_num)
if substep_num == 0:
step_prompt = "STEP {} (Subproblem): ".format(step)
else:
if self.prompt_style == "struct":
step_prompt = "STEP {} (Facts): ".format(step)
else:
# if struct_mini
step_prompt = "STEP {} (Solution): ".format(step)
return {"prompt_state": prompt_state + step_prompt,
"step_prompt": step_prompt,
"substep_num": substep_num,
"step": step}
elif self.prompt_style == "cot_step":
step = node.step_num + 1
step_prompt = "Step {}: ".format(step)
return {"prompt_state": prompt_state + step_prompt,
"step_prompt": step_prompt,
"substep_num": None,
"step": step}
else:
step = node.step_num + 1
return {"prompt_state": prompt_state,
"step_prompt": "",
"substep_num": None,
"step": step}
def get_confidence(self, sample):
logprobs = sample.logprobs
confidence = 0
for logprob in logprobs:
for k in logprob:
confidence += logprob[k]
break
return confidence / len(logprobs)
def child_format(self, parent_node: Node, outputs: any,
step_prompt: str, step: int, substep_num: int, unique: bool) -> List[Node]:
if parent_node.halt_status:
parent_node.reward = 0
return [parent_node]
else:
new_reason_chains = [output.text for output in outputs]
# print("new_reason_chains: ", new_reason_chains)
confidences = [self.get_confidence(output) for output in outputs]
children = []
dict_chains = {}
for confidence, reason_chain in zip(confidences, new_reason_chains):
if reason_chain not in dict_chains:
if unique:
dict_chains[reason_chain] = 1
new_node = Node()
new_node.parent = parent_node
if self.prompt_style in ["struct", "struct-mini"]:
if substep_num == 1:
flag_str = "STEP {} ".format(step + 1)
if reason_chain[-len(flag_str):] == flag_str:
reason_chain = reason_chain[0:-len(flag_str)]
else:
new_node.halt_status = True
reason_chain = reason_chain.strip() + "\n\n"
else:
if reason_chain[-2:] == "ST":
reason_chain = reason_chain[0:-2]
reason_chain = reason_chain.rstrip() + "\n\n"
elif self.prompt_style == "cot_step":
flag_str = "St"
if reason_chain[-len(flag_str):] == flag_str:
reason_chain = reason_chain[0:-len(flag_str)]
else:
new_node.halt_status = True
reason_chain = reason_chain.rstrip() + "\n\n"
reason_chain = step_prompt + reason_chain
new_node.chains = parent_node.chains + [reason_chain]
new_node.step_num = step
new_node.substep_num = substep_num
if self.set_edge_links:
new_node.parent = parent_node
new_node.reward = confidence
children.append(new_node)
# print("\n\n\n")
# parent_node.children = children_
if not (self.prompt_style in ["struct", "struct_min", "cot_step"]):
batch_chains = ["".join(child.chains) for child in children]
batch_outputs = self.generator.generate(prompt=batch_chains,
max_length=10,
num_samples=1,
stop=[],
temperature=0)
chain2pred = {}
for output in batch_outputs:
chain2pred[output.prompt] = output.outputs[0].text
batch_outputs = [chain2pred[chain] for chain in batch_chains]
children_ = []
for node, text in zip(children, batch_outputs):
if text.strip() == "":
node.halt_status = True
node.chains = node.chains[0:-1] + [node.chains[-1] + "\n\n"]
else:
last_chain = node.chains[-1]
for c in text:
if c == "\n":
last_chain = last_chain + "\n"
else:
break
node.chains = node.chains[0:-1] + [last_chain]
children_.append(node)
children = children_
batch_chains = [child.chains for child in children]
if self.reward_types == "confidence":
rewards = [0] * len(children)
num_rewards = 0
else:
rewards, num_rewards = batch_evaluate(generator=self.generator,
prompt_style=self.prompt_style,
batch_chains=batch_chains,
positive_token_id=self.Aid,
negative_token_id=self.Bid,
step_nums=[step] * len(batch_chains),
substep_num=substep_num,
reward_types=self.reward_types)
children_ = []
for child, reward in zip(children, rewards):
if "confidence" in self.reward_types:
child.reward = (child.reward + reward) / (1 + num_rewards)
else:
child.reward = (reward) / (num_rewards)
child.sum_reward = child.reward
child.root_distance = parent_node.root_distance + 1
children_.append(child)
if self.set_edge_links:
child_dict = {}
if parent_node.children:
for child in parent_node.children:
chain = "".join(child.chains)
child_dict[chain] = child
for child in children_:
chain = "".join(child.chains)
child_dict[chain] = child
parent_node.children = [v for k, v in child_dict.items()]
return children_
def batch_transition(self, nodes: List[Node], child_num: int, unique: bool = True) -> List[List[Node]]:
prompt_state_dicts = [self.prompt_format(node) for node in nodes]
prompt_states = [x["prompt_state"] for x in prompt_state_dicts]
if (self.prompt_style == "struct") or (self.prompt_style == "struct_min"):
substep_num = prompt_state_dicts[0]["substep_num"]
if substep_num == 0:
batch_outputs = self.generator.generate(prompt=prompt_states,
max_length=self.max_length,
num_samples=child_num,
stop=["EP", "###"],
logprobs=1,
temperature=self.temperature)
else:
batch_outputs = self.generator.generate(prompt=prompt_states,
max_length=self.max_length,
num_samples=child_num,
stop=["(Subproblem):", "###"],
logprobs=1,
temperature=self.temperature)
elif self.prompt_style == "cot_step":
step_num = prompt_state_dicts[0]["step"]
batch_outputs = self.generator.generate(prompt=prompt_states,
max_length=self.max_length,
num_samples=child_num,
stop=["ep {}:".format(step_num + 1), "###"],
logprobs=1,
temperature=self.temperature)
else:
batch_outputs = self.generator.generate(prompt=prompt_states,
max_length=self.max_length,
num_samples=child_num,
stop=["\n", "###"],
logprobs=1,
temperature=self.temperature)
prompt2pred = {}
for output in batch_outputs:
prompt2pred[output.prompt] = output.outputs
batch_children = []
for prompt, node, prompt_state_dict in zip(prompt_states, nodes, prompt_state_dicts):
# print("parent prompt state: ", prompt_state_dict["prompt_state"])
# print("\n\n")
children = self.child_format(parent_node=node,
outputs=prompt2pred[prompt],
unique=unique,
step_prompt=prompt_state_dict["step_prompt"],
step=prompt_state_dict["step"],
substep_num=prompt_state_dict["substep_num"])
batch_children.append(children)
return batch_children