-
Notifications
You must be signed in to change notification settings - Fork 1
/
eval_utils.py
206 lines (170 loc) · 5.8 KB
/
eval_utils.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
from dataset import BBNLI, BBQ, Arithmetic, RealToxicityPrompts
from transformers import (
AutoModelForSeq2SeqLM,
AutoTokenizer,
)
from transformers import FlaxT5ForConditionalGeneration, AutoModelForCausalLM
from typing import List, Dict
from gptcache import GPTCache
from util import query_model, Cache, PROJECTP
from tqdm import tqdm
import ipdb
class_dict = {
"BBNLI": BBNLI,
"BBQ": BBQ,
"Arithmetic": Arithmetic,
"RealTox": RealToxicityPrompts,
}
def eval_api(
cc,
args,
model_name: str = "gpt-3.5-turbo",
with_edit: bool = False,
retriever: callable = None,
edits_all: List[str] = None,
max_new_tokens=20,
device="cuda:0",
) -> Dict:
if args.generations_cache is not None:
print("Recreating cache...")
# Create gpt cache.
if args.api == "openai":
cache = GPTCache(
cache_loc=f"{project_p}/cache/cache_{model_name}.json",
key_loc="openai_key.txt",
engine=model_name,
chat_prompt_dict_path=args.chat_prompt_dict_path,
)
elif args.api == "bard":
from bardcache import BardCache
cache = BardCache(
cache_loc=f"{project_p}/cache/cache_bard.json", key_loc="palm_api_key.txt"
)
else:
raise ValueError("api name not recognized")
# Get edits.
test_inputs = cc.get_test_inputs_only()
edits = None
# Get predictions.
if with_edit:
# Get edits.
edits = cc.edits if retriever is None else retriever(edits_all, test_inputs)
fwe = cc.form_with_edit
# Percent of edits matching gold edits.
pem = [e == g for e, g in zip(edits, cc.edits)]
print("Percent of edits matching gold edits: ", sum(pem) / len(pem))
else:
edits = [None for _ in test_inputs]
if hasattr(cc, "form_without_edit"):
fwoe = cc.form_without_edit
else:
fwoe = "{question}"
queries = []
for edit, t in zip(edits, test_inputs):
if edit is None:
queries.append(fwoe.format(question=t.strip()))
else:
queries.append(fwe.format(edit=edit, question=t.strip()))
preds = []
for q in tqdm(queries, desc="Generating..."):
preds.append(cache.generate(q, max_length=max_new_tokens))
score = cc.test_scores(preds, mean=True)
print("Mean score: ", score)
return {
"scores": score,
"preds": preds,
"edits": edits,
"test_inputs": test_inputs,
}
def load_model_tokenizer(args, model_name, device):
assert not args.llama or not args.from_flax
assert not args.from_flax or not args.peft
# Load model and tokenizer.
if args.from_flax:
model = FlaxT5ForConditionalGeneration.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
elif args.llama:
if args.peft:
from peft import PeftModel, PeftConfig
pconfig = PeftConfig.from_pretrained(model_name)
bmbp = pconfig.base_model_name_or_path
model = AutoModelForCausalLM.from_pretrained(bmbp).to(device)
model = PeftModel.from_pretrained(model, model_name)
tokenizer = AutoTokenizer.from_pretrained(pconfig.base_model_name_or_path)
else:
model = AutoModelForCausalLM.from_pretrained(model_name).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.add_special_tokens({"pad_token": "<pad>"})
model.resize_token_embeddings(len(tokenizer))
model.config.pad_token_id = tokenizer.pad_token_id
else:
model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_name)
return model, tokenizer
def eval_hf(
cc,
args,
model_name: str,
with_edit: bool = False,
retriever: callable = None,
edits_all: List[str] = None,
max_new_tokens=20,
device="cuda:0",
) -> Dict:
model, tokenizer = load_model_tokenizer(args, model_name, device)
test_inputs = cc.get_test_inputs_only()
edits = None
# Get predictions.
if with_edit:
# Get edits.
edits = cc.edits if retriever is None else retriever(edits_all, test_inputs)
fwe = cc.form_with_edit
# Percent of edits matching gold edits.
pem = [e == g for e, g in zip(edits, cc.edits)]
print("Percent of edits matching gold edits: ", sum(pem) / len(pem))
else:
edits = [None for _ in test_inputs]
# Identify format to use without edit.
if hasattr(cc, "form_without_edit"):
fwoe = cc.form_without_edit
else:
fwoe = "{question}"
# Create queries.
queries = []
for edit, t in zip(edits, test_inputs):
if edit is None:
qq = fwoe.format(question=t.strip())
else:
qq = fwe.format(edit=edit, question=t.strip())
# If the model is a chat model we need to add the chat prompt.
# e.g. <s>[INST] {query} [/INST]
if args.llama and "chat" in model_name:
qq = f"<s>[INST] {qq} [/INST]"
queries.append(qq)
# Create cache.
if args.generations_cache is not None:
cache = Cache(args.generations_cache)
else:
cache = None
preds = query_model(
queries,
model,
tokenizer,
device=device,
batch_size=args.batch_size,
do_sample=True if args.llama else False,
max_length=max_new_tokens,
flax=args.from_flax,
cache=cache,
)
# Get the part after [/INST]
if args.llama and "chat" in model_name:
preds = [p.split("[/INST]")[1].strip() for p in preds]
score = cc.test_scores(preds, mean=True)
print("Mean score: ", score)
return {
"scores": score,
"preds": preds,
"edits": edits,
"test_inputs": test_inputs,
}