-
Notifications
You must be signed in to change notification settings - Fork 1
/
models.py
291 lines (232 loc) · 8.12 KB
/
models.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
from constants import (
CODEX_MODEL_NAME,
CP_MODEL_NAME,
CURIE_MODEL_NAME,
GPT2_MODEL_NAME,
GPT3_MODEL_NAME,
HF_CACHE_DIR_NAME,
INSTRUCT_MODEL_NAME,
JURASSIC_MODEL_NAME,
JURASSIC_SPACE,
RETRY_SLEEP_TIME
)
from dataclasses import dataclass
from transformers import AutoTokenizer, AutoModelForCausalLM
from utils import idx_to_ltr, prep_openai_obj_for_save
import numpy as np
import openai
import requests
import time
@dataclass
class ModelResponseNatural:
logprobs: dict
response_list: list
@dataclass
class ModelResponseBrown:
logprobs: dict
unconditional_logprobs: dict
lens: dict
response_list: list
class Test:
def _get_uniform_response(self, n_choices):
return {idx_to_ltr(i): np.log(1/n_choices) for i in range(n_choices)}
def process_question_natural(self, question):
n_choices = question.get_n_choices()
logprobs = self._get_uniform_response(n_choices=n_choices)
return ModelResponseNatural(
logprobs=logprobs,
response_list=list()
)
def process_question_brown(self, question):
n_choices = question.get_n_choices()
logprobs = self._get_uniform_response(n_choices=n_choices)
lens = {idx_to_ltr(i): 1 for i in range(n_choices)}
return ModelResponseBrown(
logprobs=logprobs,
unconditional_logprobs=logprobs,
lens=lens,
response_list=list()
)
class GPT2Model:
def __init__(self, model_name):
self.tokenizer = AutoTokenizer.from_pretrained(
model_name,
cache_dir=HF_CACHE_DIR_NAME
)
self.model = AutoModelForCausalLM.from_pretrained(
model_name,
cache_dir=HF_CACHE_DIR_NAME
)
self.lbls_map = {v: k for k, v in self.tokenizer.vocab.items()}
def process_question_natural(self, question):
prompt_text = question.get_natural_prompt()
inputs = self.tokenizer(prompt_text, return_tensors="pt")
outputs = self.model(**inputs)
logits = outputs.logits[0, -1]
probs = logits.softmax(dim=-1)
logprobs_dict = {
self.lbls_map[i]:
np.log(probs[i].item()) for i in range(len(self.lbls_map))
}
# Reduce logprobs_dict to only keys with top 50 largest values
logprobs_dict = {
k: v for k, v in sorted(
logprobs_dict.items(),
key=lambda item: item[1],
reverse=True
)[:200]
}
return ModelResponseNatural(
logprobs=logprobs_dict,
response_list=list()
)
def process_question_brown(self):
pass
class CodeParrot(GPT2Model):
def __init__(self):
super().__init__(model_name=CP_MODEL_NAME)
class GPT2(GPT2Model):
def __init__(self):
super().__init__(model_name=GPT2_MODEL_NAME)
class Jurassic:
def __init__(self, api_key):
self.key = api_key
def process_question_natural(self, question):
prompt_text = question.get_natural_prompt()
response = requests.post(
f"https://api.ai21.com/studio/v1/{JURASSIC_MODEL_NAME}/complete",
headers={"Authorization": f"Bearer {self.key}"},
json={
"prompt": prompt_text,
"numResults": 1,
"maxTokens": 1,
"topKReturn": 64,
"temperature": 1.0,
}
)
while True:
resp_json = response.json()
try:
completion_tokens = resp_json["completions"][0]["data"]["tokens"][0]["topTokens"]
break
except Exception:
print(resp_json)
print(f"Will retry API call in {RETRY_SLEEP_TIME} seconds...")
time.sleep(RETRY_SLEEP_TIME)
log_probs = {t["token"]: t["logprob"] for t in completion_tokens}
completion_tokens = {k.replace(JURASSIC_SPACE, " "): v
for k, v in log_probs.items()}
return ModelResponseNatural(
logprobs=completion_tokens,
response_list=[resp_json]
)
def process_question_brown(self):
pass
class OpenAIModel:
def __init__(self, api_key, model_name, add_space=False):
openai.api_key = api_key
self.add_space = add_space
self.model_name = model_name
def process_question_natural(self, question):
prompt_text = question.get_natural_prompt()
response = self._get_response(text=prompt_text, echo=False)
logprobs = dict(response["choices"][0]["logprobs"]["top_logprobs"][0])
return ModelResponseNatural(
logprobs=logprobs,
response_list=[
prep_openai_obj_for_save(
obj=response,
prompt_text=prompt_text
)
]
)
def _get_response(self, text, echo):
while True:
try:
response = openai.Completion.create(
model=self.model_name,
prompt=text+(" " if self.add_space else ""),
temperature=0, # Doesn't actually matter here
max_tokens=1, # Just need to get letter
logprobs=5, # Get max number of logprobs
echo=echo
)
return response
except Exception as e:
print(e)
print("Will wait and retry...")
time.sleep(RETRY_SLEEP_TIME)
def process_question_brown(self, question):
prompt_text = question.get_brown_prompt()
response_list = list()
logprobs = dict()
unconditional_logprobs = dict()
lens = dict()
for idx, choice in enumerate(question.choices):
ltr = idx_to_ltr(idx)
# Get unconditional logprobs
response = self._get_response(text=f"Answer: {choice}", echo=True)
choice_logprobs = (
response["choices"][0]["logprobs"]["token_logprobs"][2:-1]
)
choice_n_tokens = len(choice_logprobs)
unconditional_logprobs[ltr] = sum(choice_logprobs)
lens[ltr] = choice_n_tokens
response_list.append(
prep_openai_obj_for_save(
obj=response,
prompt_text=f"Answer: {choice}"
)
)
# Get conditional logprobs
response = self._get_response(
text=f"{prompt_text} {choice}", echo=True
)
token_logprobs = (
response["choices"][0]["logprobs"]["token_logprobs"]
)
choice_logprobs = token_logprobs[-(choice_n_tokens+1):-1]
logprobs[ltr] = sum(choice_logprobs)
response_list.append(
prep_openai_obj_for_save(
obj=response,
prompt_text=f"{prompt_text} {choice}"
)
)
return ModelResponseBrown(
logprobs=logprobs,
unconditional_logprobs=unconditional_logprobs,
lens=lens,
response_list=response_list
)
class Codex(OpenAIModel):
def __init__(self, api_key):
super().__init__(
api_key=api_key,
model_name=CODEX_MODEL_NAME,
add_space=True
)
class GPT3(OpenAIModel):
def __init__(self, api_key):
super().__init__(api_key=api_key, model_name=GPT3_MODEL_NAME)
class Instruct(OpenAIModel):
def __init__(self, api_key):
super().__init__(api_key=api_key, model_name=INSTRUCT_MODEL_NAME)
class Curie(OpenAIModel):
def __init__(self, api_key):
super().__init__(api_key=api_key, model_name=CURIE_MODEL_NAME)
def get_model_by_name(name, api_key):
try:
return {
"codex": Codex,
"gpt3": GPT3,
"instruct": Instruct,
"curie": Curie,
"jurassic": Jurassic
}[name](api_key=api_key)
except KeyError:
return {
"test": Test,
"cp": CodeParrot,
"gpt2": GPT2
}[name]()