-
Notifications
You must be signed in to change notification settings - Fork 0
/
compute_ood.py
310 lines (268 loc) · 11.4 KB
/
compute_ood.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
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
from peft import AutoPeftModelForSequenceClassification
from transformers import AutoTokenizer, DataCollatorWithPadding
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from torch.nn import CrossEntropyLoss, MSELoss
from data import load
from sklearn.covariance import EmpiricalCovariance
from evaluation import evaluate_ood
from torch.utils.data import DataLoader
import os
import pandas as pd
from utils import find_subdir_with_smallest_number
from config import parse_args
import evaluate
# os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
task_to_labels = {
'sst2': 2,
'imdb': 2,
'20ng': 20,
'trec': 6,
'clinc150': 150,
"bank": round(77 * 0.5),
'rostd': 3
# "clincMix": 150
}
def merge_keys(l, keys):
new_dict = {}
for key in keys:
new_dict[key] = []
for i in l:
new_dict[key].append(i[key])
new_dict[key] = np.mean(np.array(new_dict[key]))
return new_dict
# outputs["auroc_IN"] = auroc_in
# outputs["fpr95_IN"] = fpr_95_in
# outputs["aupr_IN"] = aupr_in
def detect_ood(model, dev_dataloader, test_dataset, benchmarks, data_collator):
class_var, class_mean, norm_bank, all_classes = prepare_ood(model, dev_dataloader)
print("finshed prepare")
res = []
keys = ["softmax_auroc_IN", "softmax_fpr95_IN", "softmax_aupr_IN",
"maha_auroc_IN", "maha_fpr95_IN", "maha_aupr_IN",
"cosine_auroc_IN", "cosine_fpr95_IN", "cosine_aupr_IN",
"energy_auroc_IN", "energy_fpr95_IN","energy_aupr_IN"]
in_scores = compute_ood(test_dataset, model, class_var, class_mean, norm_bank, all_classes)
# print("in_scores",len(in_scores))
for tag, ood_features in benchmarks:
dataloader = DataLoader(ood_features, batch_size=32, collate_fn=data_collator)
out_scores = compute_ood(dataloader, model, class_var, class_mean, norm_bank, all_classes)
print(tag, "out score finshed")
results = evaluate_ood(in_scores, out_scores)
# print("ood result", results)
res.append(results)
del dataloader, ood_features
# wandb.log(results, step=num_steps)
res = merge_keys(res, keys)
return res
def save_results(args, test_results):
if not os.path.exists(args.save_results_path):
os.makedirs(args.save_results_path)
var = [args.task_name, args.seed, args.ib]
names = ['dataset', 'seed', 'is_ib']
vars_dict = {k: v for k, v in zip(names, var)}
results = dict(test_results, **vars_dict)
keys = list(results.keys())
values = list(results.values())
file_name = args.task_name + '_results.csv'
results_path = os.path.join(args.save_results_path, file_name)
if not os.path.exists(results_path):
ori = []
ori.append(values)
df1 = pd.DataFrame(ori, columns=keys)
df1.to_csv(results_path, index=False)
else:
df1 = pd.read_csv(results_path)
new = pd.DataFrame(results, index=[1])
# df1 = df1.append(new, ignore_index=True)
df1 = pd.concat([df1, new], ignore_index=True)
df1.to_csv(results_path, index=False)
data_diagram = pd.read_csv(results_path)
print('test_results')
print(data_diagram)
def compute_ood(dataloader, model, class_var, class_mean, norm_bank, all_classes):
model.eval()
in_scores = []
# dataloader = DataLoader(dev_dataset, batch_size=128, collate_fn=data_collator)
for batch in dataloader:
batch.pop("idx", None)
batch = {key: value.cuda() for key, value in batch.items()}
input_ids = batch['input_ids']
batch['labels'] = None
with torch.no_grad():
outputs = model(**batch)
logits = outputs.get("logits")
pooled = outputs.get("hidden_states")[-1]
if input_ids is not None:
batch_size = input_ids.shape[0]
# if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
sequence_lengths = torch.eq(input_ids, model.config.pad_token_id).int().argmax(-1) - 1
sequence_lengths = sequence_lengths % input_ids.shape[-1]
# sequence_lengths = sequence_lengths
else:
sequence_lengths = -1
# pooled = pooled[torch.arange(args.val_batch_size), sequence_lengths]
pooled = pooled[torch.arange(batch_size, device=pooled.device), sequence_lengths]
ood_keys = None
softmax_score = F.softmax(logits, dim=-1).max(-1)[0]
maha_score = []
for c in all_classes:
centered_pooled = pooled - class_mean[c].unsqueeze(0)
ms = torch.diag(centered_pooled @ class_var @ centered_pooled.t())
maha_score.append(ms)
maha_score = torch.stack(maha_score, dim=-1)
maha_score = maha_score.min(-1)[0]
maha_score = -maha_score
norm_pooled = F.normalize(pooled, dim=-1)
cosine_score = norm_pooled @ norm_bank.t()
cosine_score = cosine_score.max(-1)[0]
energy_score = torch.logsumexp(logits, dim=-1)
ood_keys = {
'softmax': softmax_score.tolist(),
'maha': maha_score.tolist(),
'cosine': cosine_score.tolist(),
'energy': energy_score.tolist(),
}
in_scores.append(ood_keys)
return in_scores
# from torch.utils.data import DataLoader
#
# train_dataloader = DataLoader(small_train_dataset, shuffle=True, batch_size=8)
# eval_dataloader = DataLoader(small_eval_dataset, batch_size=8)
# metric = evaluate.load("accuracy")
# model.eval()
# for batch in eval_dataloader:
# batch = {k: v.to(device) for k, v in batch.items()}
# with torch.no_grad():
# outputs = model(**batch)
#
# logits = outputs.logits
# predictions = torch.argmax(logits, dim=-1)
# metric.add_batch(predictions=predictions, references=batch["labels"])
#
# metric.compute()
def prepare_ood(model, dataloader=None):
bank = None
label_bank = None
model.eval()
for batch in dataloader:
batch.pop("idx", None)
batch = {key: value.cuda() for key, value in batch.items()}
labels = batch['labels']
input_ids = batch['input_ids']
batch['labels'] = None
with torch.no_grad():
outputs = model(**batch)
# logits = outputs.get("logits")
pooled = outputs.get("hidden_states")[-1]
if input_ids is not None:
batch_size = input_ids.shape[0]
# if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
sequence_lengths = torch.eq(input_ids, model.config.pad_token_id).int().argmax(-1) - 1
sequence_lengths = sequence_lengths % input_ids.shape[-1]
# sequence_lengths = sequence_lengths
else:
sequence_lengths = -1
pooled = pooled[torch.arange(batch_size), sequence_lengths]
# pooled = pooled[:, sequence_lengths]
# print("poled:", pooled.shape)
if bank is None:
bank = pooled.clone().detach()
label_bank = labels.clone().detach()
else:
bank = pooled.clone().detach()
label_bank = labels.clone().detach()
bank = torch.cat([bank, bank], dim=0)
label_bank = torch.cat([label_bank, label_bank], dim=0)
norm_bank = F.normalize(bank, dim=-1).cuda()
N, d = bank.size()
all_classes = list(set(label_bank.tolist()))
class_mean = torch.zeros(max(all_classes) + 1, d).cuda()
for c in all_classes:
class_mean[c] = (bank[label_bank == c].mean(0))
centered_bank = (bank - class_mean[label_bank]).detach().cpu().numpy()
precision = EmpiricalCovariance().fit(centered_bank).precision_.astype(np.float32)
class_var = torch.from_numpy(precision).float().cuda()
return class_var, class_mean, norm_bank, all_classes
if __name__ == '__main__':
args = parse_args("test")
args.ib = True
out_dir = os.path.join(args.save_results_path, args.task_name, str(args.seed), str(args.ib))
print(out_dir)
out_dir = find_subdir_with_smallest_number(out_dir)
print(out_dir)
assert out_dir is not None
num_labels = task_to_labels[args.task_name]
model = AutoPeftModelForSequenceClassification.from_pretrained(out_dir, num_labels=num_labels)
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, )
tokenizer.pad_token_id = tokenizer.eos_token_id
tokenizer.pad_token = tokenizer.eos_token
# tokenizer.padding_side="right"
model = model.to("cuda")
model.config.pad_token_id = model.config.eos_token_id
model.config.output_hidden_states = True
# print(model.config.output_hidden_states)
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
print(tokenizer.padding_side)
_, dev_dataset, test_dataset = load(args.task_name, tokenizer, max_seq_length=args.max_seq_length,
is_id=True)
dev_dataset.to_pandas().info()
test_dataset.to_pandas().info()
ood_datasets = ['rte', 'sst2', 'mnli', '20ng', 'trec', 'imdb', 'wmt16', 'multi30k']
# ood_datasets = ['rte', 'sst2', ]
benchmarks = ()
if args.task_name in ["sst2", "imdb"]:
ood_datasets = list(set(ood_datasets) - set(["sst2", "imdb"]))
else:
ood_datasets = list(set(ood_datasets) - set([args.task_name]))
for dataset in ood_datasets:
_, _, ood_dataset = load(dataset, tokenizer, max_seq_length=128)
benchmarks = (('ood_' + dataset, ood_dataset),) + benchmarks
ood_dataset.to_pandas().info()
# train_dataloader = DataLoader(small_train_dataset, shuffle=True, batch_size=8)
# outputs.hidden_states[-1]
# to
# match
# outputs.last_hidden_states
# exactly
##test acc
test_dataloader = DataLoader(test_dataset, batch_size=args.val_batch_size, collate_fn=data_collator)
# print("",len(test_dataloader))
eval_dataloader = DataLoader(dev_dataset, batch_size=args.val_batch_size, collate_fn=data_collator)
metric = evaluate.load("accuracy")
model.eval()
for batch in test_dataloader:
batch.pop("idx", None)
batch = {k: v.cuda() for k, v in batch.items()}
# print(batch["input_ids"][0])
with torch.no_grad():
outputs = model(**batch)
logits = outputs.logits
# # hs = outputs.hidden_states # 33 128, 66, 4096
predictions = torch.argmax(logits, dim=-1)
metric.add_batch(predictions=predictions, references=batch["labels"])
# print("test acc:", metric.compute())
test_acc = metric.compute()
metric = evaluate.load("accuracy")
for batch in eval_dataloader:
batch.pop("idx", None)
batch = {k: v.cuda() for k, v in batch.items()}
# print(batch["input_ids"][0])
with torch.no_grad():
outputs = model(**batch)
logits = outputs.logits
# # hs = outputs.hidden_states # 33 128, 66, 4096
predictions = torch.argmax(logits, dim=-1)
metric.add_batch(predictions=predictions, references=batch["labels"])
eval_acc = metric.compute()
# del test_dataset, test_dataloader, dev_dataset, eval_dataloader
ood_res = detect_ood(model, eval_dataloader, test_dataloader, benchmarks, data_collator)
final_res = dict({"test_acc": test_acc['accuracy'], 'eval_acc': eval_acc["accuracy"]}, **ood_res)
# final_res = ood_res
save_results(args, final_res)
# print(res)
## comput OOD