-
Notifications
You must be signed in to change notification settings - Fork 5
/
main.py
269 lines (257 loc) · 15.1 KB
/
main.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
import argparse
import random
import os
os.environ['CURL_CA_BUNDLE'] = ''
import torch
import numpy as np
import json
import nltk
import time
import pickle
from tqdm import tqdm
from datasets import load_metric
from transformers import AutoTokenizer,GPTJForCausalLM
from MetaICL.metaicl.data import MetaICLData
from MetaICL.metaicl.model import MetaICLModel
from get_task import get_task
from utils import calculate_sentence_transformer_embedding,codex_execution,expand_to_aliases
from core_method import selective_annotation,prompt_retrieval
parser = argparse.ArgumentParser()
parser.add_argument('--task_name', required=True,type=str)
parser.add_argument('--selective_annotation_method', required=True,type=str)
parser.add_argument('--model_cache_dir', required=True,type=str)
parser.add_argument('--data_cache_dir', required=True,type=str)
parser.add_argument('--model_key', type=str)
parser.add_argument('--prompt_retrieval_method', default='similar',type=str)
parser.add_argument('--model_name', default='EleutherAI/gpt-j-6B',type=str)
parser.add_argument('--embedding_model', default='sentence-transformers/paraphrase-mpnet-base-v2',type=str)
parser.add_argument('--annotation_size', default=100,type=int)
parser.add_argument('--seed', default=0,type=int)
parser.add_argument('--batch_size', default=10,type=int)
parser.add_argument('--debug', action='store_true')
parser.add_argument('--cuda_id', type=int, default = 0)
args = parser.parse_args()
torch.cuda.set_device(args.cuda_id)
def set_seed(seed: int):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
if __name__=='__main__':
set_seed(args.seed)
args.output_dir = os.path.join("result", args.task_name, args.selective_annotation_method + "_" + str(args.seed))
if not os.path.isdir(args.output_dir):
os.makedirs(args.output_dir,exist_ok=True)
train_examples,eval_examples,train_text_to_encode,eval_text_to_encode,format_example,label_map = get_task(args=args)
if os.path.isfile(os.path.join(args.output_dir, "train_embeds.pickle")) and os.path.join(args.output_dir, "eval_embeds.pickle"):
with open(os.path.join(args.output_dir, "train_embeds.pickle"), 'rb') as myfile:
total_train_embeds = pickle.load(myfile)
print("load_train_embed")
with open(os.path.join(args.output_dir, "eval_embeds.pickle"), 'rb') as myfile:
total_eval_embeds = pickle.load(myfile)
print("load_eval_embed")
else:
total_train_embeds = calculate_sentence_transformer_embedding(text_to_encode=train_text_to_encode,
args=args)
total_eval_embeds = calculate_sentence_transformer_embedding(text_to_encode=eval_text_to_encode,
args=args)
with open(os.path.join(args.output_dir, "train_embeds.pickle"), 'wb') as myfile:
pickle.dump(total_train_embeds, myfile)
with open(os.path.join(args.output_dir, "eval_embeds.pickle"), 'wb') as myfile:
pickle.dump(total_eval_embeds, myfile)
if args.task_name in ['mnli','rte','sst5','mrpc','dbpedia_14','hellaswag','xsum','nq']:
if args.task_name=='xsum':
tokenizer_gpt = AutoTokenizer.from_pretrained(args.model_name,cache_dir=args.model_cache_dir)
inference_model = GPTJForCausalLM.from_pretrained("EleutherAI/gpt-j-6B",cache_dir=args.model_cache_dir)
inference_model.cuda()
inference_model.eval()
data_module = None
return_string = True
device = torch.device('cuda')
single_input_len = None
maximum_input_len = 1900
elif args.task_name=='nq':
maximum_input_len = 3800
return_string = True
single_input_len = None
inference_model = None
data_module = None
tokenizer_gpt = None
model_keys = args.model_key.split('##')
else:
data_module = MetaICLData(method="direct", max_length=1024, max_length_per_example=256)
inference_model = MetaICLModel(args=args)
inference_model.load()
inference_model.cuda()
inference_model.eval()
tokenizer_gpt = None
return_string = False
single_input_len = 250
maximum_input_len = 1000
select_start_time = time.time()
if os.path.isfile(os.path.join(args.output_dir,'first_phase_selected_indices.json')):
with open(os.path.join(args.output_dir,'first_phase_selected_indices.json')) as f:
first_phase_selected_indices = json.load(f)
else:
first_phase_selected_indices = selective_annotation(embeddings=total_train_embeds,
train_examples=train_examples,
return_string=return_string,
format_example=format_example,
maximum_input_len=maximum_input_len,
label_map=label_map,
single_context_example_len=single_input_len,
inference_model=inference_model,
inference_data_module=data_module,
tokenizer_gpt=tokenizer_gpt,
args=args)
print(type(first_phase_selected_indices))
print(first_phase_selected_indices)
with open(os.path.join(args.output_dir,'first_phase_selected_indices.json'),'w') as f:
json.dump(first_phase_selected_indices,f,indent=4)
select_time = time.time()-select_start_time
processed_train_examples = [train_examples[idx] for idx in first_phase_selected_indices]
processed_eval_examples = eval_examples
prompt_retrieval(train_embs=total_train_embeds[first_phase_selected_indices],test_embs=total_eval_embeds,train_examples=processed_train_examples,
eval_examples=eval_examples,return_string=return_string,format_example=format_example,
maximum_input_len=maximum_input_len,single_context_example_len=single_input_len,label_map=label_map,args=args)
prompt_cache_dir = os.path.join(args.output_dir, 'prompts')
candidate_prompt_files = os.listdir(prompt_cache_dir)
prompt_files = [f for f in candidate_prompt_files if f.endswith('.json')]
assert len(prompt_files) == len(processed_eval_examples), f"len(prompt_files)={len(prompt_files)}," \
f"len(processed_eval_examples)={len(processed_eval_examples)}"
output_dir = os.path.join(args.output_dir,'results')
if not os.path.isdir(output_dir):
os.makedirs(output_dir, exist_ok=True)
count = 0
running_flag = True
golds = []
preds = []
if not args.task_name in ['hellaswag','xsum','nq']:
all_labels = []
label_to_digit = {}
for k, v in label_map.items():
all_labels.append(v)
label_to_digit[v] = k
execution_count = 0
while running_flag:
running_flag = False
count += 1
bar = tqdm(range(len(prompt_files)), desc=f" LLM inference")
for file in prompt_files:
bar.update(1)
if not os.path.isfile(os.path.join(output_dir,file)):
running_flag = True
if args.task_name == 'hellaswag':
with open(os.path.join(prompt_cache_dir, file)) as f:
one_test_example = json.load(f)
cur_train_data = one_test_example[1]
cur_input = {'input': format_example(one_test_example[2], label_map=label_map, args=args)[0],
'options': one_test_example[2]['endings']}
data_module.k = len(cur_train_data)
data_module.tensorize(cur_train_data, [cur_input])
prediction = inference_model.do_predict(data_module)[0]
assert prediction in one_test_example[2]['endings']
with open(f"{output_dir}/{file}", 'w') as f:
json.dump([prediction, one_test_example[2]['endings'][one_test_example[2]['label']]], f)
preds.append(prediction)
golds.append(one_test_example[2]['endings'][one_test_example[2]['label']])
elif args.task_name == 'xsum':
with open(os.path.join(prompt_cache_dir, file)) as f:
one_test_example = json.load(f)
context = one_test_example[1]
input_ids = tokenizer_gpt(context, return_tensors="pt").input_ids
input_ids = input_ids[:, :1900]
input_len = input_ids.shape[1]
input_ids = input_ids.to(device)
gen_tokens = inference_model.generate(input_ids, do_sample=False, temperature=0.7,
max_length=input_len + 64,
output_scores=True, return_dict_in_generate=True)
generated_text = tokenizer_gpt.batch_decode(gen_tokens.sequences.view(-1, 1))
stop = ['--', '\n', ';', '#']
stop_index = len(generated_text)
for i, c in enumerate(generated_text):
if i > input_len and c.strip(' ') in stop:
stop_index = i
break
prediction = ' '.join(generated_text[input_len:stop_index])
golds.append(one_test_example[2]['summary'])
preds.append(prediction)
with open(f"{output_dir}/{file}", 'w') as f:
json.dump(
[' '.join(generated_text[input_len:]), ' '.join(generated_text[input_len:stop_index]),
one_test_example[2]['summary'], input_len, stop_index], f, indent=4)
elif args.task_name == 'nq':
cur_key = model_keys[execution_count % len(model_keys)]
execution_count += 1
try:
codex_execution(key=cur_key, output_path=os.path.join(output_dir, file),
prompt_path=os.path.join(prompt_cache_dir, file))
except Exception as e:
print(e)
time.sleep(3)
else:
with open(os.path.join(prompt_cache_dir, file)) as f:
one_test_example = json.load(f)
cur_train_data = one_test_example[1]
for idx in range(len(cur_train_data)):
cur_train_data[idx]['options'] = all_labels
for idx in range(len(cur_train_data)):
cur_train_data[idx]['options'] = all_labels
cur_input = format_example(one_test_example[2], label_map=label_map, args=args)[0]
data_module.k = len(cur_train_data)
data_module.tensorize(cur_train_data, [cur_input], options=all_labels)
prediction = inference_model.do_predict(data_module)[0]
with open(os.path.join(output_dir, file), 'w') as f:
json.dump([prediction, one_test_example[2]['label']], f)
preds.append(label_to_digit[prediction])
golds.append(one_test_example[2]['label'])
if args.task_name=='xsum':
assert len(golds) == len(preds), f"len(golds)={len(golds)}, len(preds)={len(preds)}"
preds, golds = postprocess_text(preds, golds)
metric = load_metric("rouge")
result = metric.compute(predictions=preds, references=golds, use_stemmer=True)
result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
result = {k: round(v, 4) for k, v in result.items()}
result["time_cost"] = select_time
print(os.path.join(args.output_dir,'result_summary.json'))
with open(os.path.join(args.output_dir,'result_summary.json'), 'w') as f:
json.dump(result, f)
print(result)
elif args.task_name=='nq':
correct = 0
total = 0
for file in prompt_files:
with open(os.path.join(prompt_cache_dir, file)) as f:
one_test_example = json.load(f)
answers = expand_to_aliases(one_test_example[2]["long"] + one_test_example[2]["short_targets"],
make_sub_answers=True)
with open(os.path.join(output_dir, file)) as f:
pred_dict = json.load(f)
prediction = pred_dict['choices'][0]['text'].replace('\n', ' ')
prediction = ' '.join(prediction.split(' ')[1:])
predictions = expand_to_aliases([prediction])
if len(list(answers & predictions)) > 0:
correct += 1
total += 1
with open(os.path.join(args.output_dir,'result_summary.txt'), 'w') as f:
f.write(f"{total} examples, accuracy is: {correct / total}\n")
f.write(f"select_time is: {select_time}\n")
print(f"{total} examples, accuracy is: {correct / total}\n")
else:
assert len(golds) == len(preds), f"len(golds)={len(golds)}, len(preds)={len(preds)}"
total = len(golds)
correct = 0
for p, g in zip(golds, preds):
if p == g:
correct += 1
with open(os.path.join(args.output_dir,'result_summary.txt'), 'w') as f:
f.write(f"{len(golds)} examples, accuracy is: {correct / total}\n")
f.write(f"select_time is: {select_time}\n")
print(f'The accuracy is {correct / total}\n')
# nohup python main.py --model_cache_dir models --data_cache_dir datasets --task_name mnli --selective_annotation_method ideal --annotation_size 18 --seed 3 --cuda_id 0 --model_name EleutherAI/gpt-neo-2.7B &