-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluate_sentence_embedding.py
262 lines (229 loc) · 9.6 KB
/
evaluate_sentence_embedding.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
# COPIED FROM https://github.com/voidism/DiffCSE/blob/master/evaluation.py
import sys
from collections import OrderedDict
import logging
import argparse
from prettytable import PrettyTable
import torch
from transformers import AutoModel, AutoTokenizer, AutoModelForSequenceClassification
# Set up logger
logging.basicConfig(format="%(asctime)s : %(message)s", level=logging.DEBUG)
# Set PATHs
PATH_TO_SENTEVAL = "./SentEval"
PATH_TO_DATA = "./SentEval/data"
# Import SentEval
sys.path.insert(0, PATH_TO_SENTEVAL)
import senteval
def print_table(task_names, scores):
tb = PrettyTable()
tb.field_names = task_names
tb.add_row(scores)
print(tb)
def reformat_parameters(weights: OrderedDict):
new_weights = OrderedDict()
# print("checkpoint weights: ==================\n", weights.keys())
for k, v in weights.items():
if "transformer_model.bert." in k and "pooler.dense" not in k:
new_weights[k.replace("transformer_model.bert.", "")] = v
elif "transformer_model.roberta." in k:
new_weights[k.replace("transformer_model.roberta.", "")] = v
return new_weights
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model_name_or_path", type=str, help="Transformers' model name or path")
parser.add_argument("--model_checkpoint", type=str, help="model checkpoint")
parser.add_argument(
"--pooler",
type=str,
choices=["cls", "cls_before_pooler", "avg", "avg_top2", "avg_first_last"],
default="cls",
help="Which pooler to use",
)
parser.add_argument(
"--mode",
type=str,
choices=["dev", "test", "fasttest"],
default="test",
help="What evaluation mode to use (dev: fast mode, dev results; test: full mode, test results); fasttest: fast mode, test results",
)
parser.add_argument(
"--task_set",
type=str,
choices=["sts", "transfer", "full", "na"],
default="sts",
help="What set of tasks to evaluate on. If not 'na', this will override '--tasks'",
)
parser.add_argument(
"--tasks",
type=str,
nargs="+",
default=[
"STS12",
"STS13",
"STS14",
"STS15",
"STS16",
"MR",
"CR",
"MPQA",
"SUBJ",
"SST2",
"TREC",
"MRPC",
"SICKRelatedness",
"STSBenchmark",
],
help="Tasks to evaluate on. If '--task_set' is specified, this will be overridden",
)
args = parser.parse_args()
# Load transformers' model checkpoint
model = AutoModel.from_pretrained(args.model_name_or_path, add_pooling_layer=False)
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
model.load_state_dict(reformat_parameters(torch.load(f"{args.model_checkpoint}/pytorch_model.bin")), strict=False)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = model.to(device)
# Set up the tasks
if args.task_set == "sts":
args.tasks = ["STS12", "STS13", "STS14", "STS15", "STS16", "STSBenchmark", "SICKRelatedness"]
elif args.task_set == "transfer":
args.tasks = ["MR", "CR", "MPQA", "SUBJ", "SST2", "TREC", "MRPC"]
elif args.task_set == "full":
args.tasks = ["STS12", "STS13", "STS14", "STS15", "STS16", "STSBenchmark", "SICKRelatedness"]
args.tasks += ["MR", "CR", "MPQA", "SUBJ", "SST2", "TREC", "MRPC"]
# Set params for SentEval
if args.mode == "dev" or args.mode == "fasttest":
# Fast mode
params = {"task_path": PATH_TO_DATA, "usepytorch": True, "kfold": 5}
params["classifier"] = {"nhid": 0, "optim": "rmsprop", "batch_size": 128, "tenacity": 3, "epoch_size": 2}
elif args.mode == "test":
# Full mode
params = {"task_path": PATH_TO_DATA, "usepytorch": True, "kfold": 10}
params["classifier"] = {"nhid": 0, "optim": "adam", "batch_size": 64, "tenacity": 5, "epoch_size": 4}
else:
raise NotImplementedError
# SentEval prepare and batcher
def prepare(params, samples):
return
def batcher(params, batch, max_length=None):
# Handle rare token encoding issues in the dataset
if len(batch) >= 1 and len(batch[0]) >= 1 and isinstance(batch[0][0], bytes):
batch = [[word.decode("utf-8") for word in s] for s in batch]
sentences = [" ".join(s) for s in batch]
# Tokenization
if max_length is not None:
batch = tokenizer.batch_encode_plus(
sentences, return_tensors="pt", padding=True, max_length=max_length, truncation=True
)
else:
batch = tokenizer.batch_encode_plus(
sentences,
return_tensors="pt",
padding=True,
)
# Move to the correct device
for k in batch:
batch[k] = batch[k].to(device)
# Get raw embeddings
with torch.no_grad():
outputs = model(**batch, output_hidden_states=True, return_dict=True)
last_hidden = outputs.last_hidden_state
pooler_output = outputs.pooler_output
hidden_states = outputs.hidden_states
# Apply different poolers
if args.pooler == "cls":
# There is a linear+activation layer after CLS representation
return pooler_output.cpu()
elif args.pooler == "cls_before_pooler":
return last_hidden[:, 0].cpu()
elif args.pooler == "avg":
return (
(last_hidden * batch["attention_mask"].unsqueeze(-1)).sum(1)
/ batch["attention_mask"].sum(-1).unsqueeze(-1)
).cpu()
elif args.pooler == "avg_first_last":
first_hidden = hidden_states[0]
last_hidden = hidden_states[-1]
pooled_result = ((first_hidden + last_hidden) / 2.0 * batch["attention_mask"].unsqueeze(-1)).sum(1) / batch[
"attention_mask"
].sum(-1).unsqueeze(-1)
return pooled_result.cpu()
elif args.pooler == "avg_top2":
second_last_hidden = hidden_states[-2]
last_hidden = hidden_states[-1]
pooled_result = ((last_hidden + second_last_hidden) / 2.0 * batch["attention_mask"].unsqueeze(-1)).sum(
1
) / batch["attention_mask"].sum(-1).unsqueeze(-1)
return pooled_result.cpu()
else:
raise NotImplementedError
results = {}
for task in args.tasks:
se = senteval.engine.SE(params, batcher, prepare)
result = se.eval(task)
results[task] = result
# Print evaluation results
if args.mode == "dev":
print("------ %s ------" % (args.mode))
task_names = []
scores = []
for task in ["STSBenchmark", "SICKRelatedness"]:
task_names.append(task)
if task in results:
scores.append("%.2f" % (results[task]["dev"]["spearman"][0] * 100))
else:
scores.append("0.00")
print_table(task_names, scores)
task_names = []
scores = []
for task in ["MR", "CR", "SUBJ", "MPQA", "SST2", "TREC", "MRPC"]:
task_names.append(task)
if task in results:
scores.append("%.2f" % (results[task]["devacc"]))
else:
scores.append("0.00")
task_names.append("Avg.")
scores.append("%.2f" % (sum([float(score) for score in scores]) / len(scores)))
print_table(task_names, scores)
elif args.mode == "test" or args.mode == "fasttest":
print("------ %s ------" % (args.mode))
task_names = []
scores = []
for task in ["STS12", "STS13", "STS14", "STS15", "STS16", "STSBenchmark", "SICKRelatedness"]:
task_names.append(task)
if task in results:
if task in ["STS12", "STS13", "STS14", "STS15", "STS16"]:
scores.append("%.2f" % (results[task]["all"]["spearman"]["all"] * 100))
else:
scores.append("%.2f" % (results[task]["test"]["spearman"].correlation * 100))
else:
scores.append("0.00")
task_names.append("Avg.")
scores.append("%.2f" % (sum([float(score) for score in scores]) / len(scores)))
print_table(task_names, scores)
if args.task_set == "sts" or args.task_set == "full":
sts_output = "{"
for name, score in zip(task_names, scores):
sts_output += '"' + name + '": ' + '"' + score + '"' + ", "
sts_output = sts_output[:-2] + "}"
task_names = []
scores = []
for task in ["MR", "CR", "SUBJ", "MPQA", "SST2", "TREC", "MRPC"]:
task_names.append(task)
if task in results:
scores.append("%.2f" % (results[task]["acc"]))
else:
scores.append("0.00")
task_names.append("Avg.")
scores.append("%.2f" % (sum([float(score) for score in scores]) / len(scores)))
print_table(task_names, scores)
if args.task_set == "transfer" or args.task_set == "full":
trans_output = "{"
for name, score in zip(task_names, scores):
trans_output += '"' + name + '": ' + '"' + score + '"' + ", "
trans_output = trans_output[:-2] + "}"
if args.task_set == "sts" or args.task_set == "full":
print(args.model_name_or_path + " " + sts_output)
if args.task_set == "transfer" or args.task_set == "full":
print(args.model_name_or_path + " " + trans_output)
if __name__ == "__main__":
main()