-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathue_estimator_decomposing.py
344 lines (259 loc) · 13.5 KB
/
ue_estimator_decomposing.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
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
from sklearn.metrics import accuracy_score
import numpy as np
import math
from torch.nn import functional as F
import torch
from torch import nn, optim
from torch.utils.data import TensorDataset, DataLoader
from torch.nn.utils import spectral_norm
from scipy.special import softmax
from utils.utils_inference import (
is_custom_head,
unpad_features,
pad_scores
)
from utils.utils_heads import (
ElectraClassificationHeadIdentityPooler,
BertClassificationHeadIdentityPooler,
ElectraNERHeadIdentityPooler,
)
import time
from tqdm import tqdm
import os
from ue4nlp.mahalanobis_distance import (
mahalanobis_distance,
mahalanobis_distance_relative,
mahalanobis_distance_marginal,
compute_centroids,
compute_covariance
)
def seed_everything(seed: int):
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = True
import logging
log = logging.getLogger()
def iCE(d_disc, d_nondisc, labels, num_classes):
y = F.one_hot(labels, num_classes=num_classes)
disc_loss = -F.log_softmax(d_disc, dim=-1) * y
nondisc_loss = F.log_softmax(d_nondisc, dim=-1) * y
loss = (disc_loss.sum(axis=1) + nondisc_loss.sum(axis=1)).mean()
return loss
class DecomposingModel(nn.Module):
def __init__(self, train_features, train_labels, use_spectral_norm=False):
super(DecomposingModel, self).__init__()
self.train_features = train_features
self.train_labels = train_labels
self.input_dim = self.train_features.shape[1]
self.output_dim = torch.unique(self.train_labels).shape[0]
self.F_linear = nn.Linear(self.input_dim, self.input_dim)
self.D_linear = nn.Linear(self.input_dim-self.output_dim, self.output_dim)
if use_spectral_norm:
self.F_linear = spectral_norm(self.F_linear)
self.D_linear = spectral_norm(self.D_linear)
def forward(self, features):
lin_mapping = self.F_linear(features)
d_disc = lin_mapping[:, :self.output_dim]
d_nondisc = self.D_linear(lin_mapping[:, self.output_dim:])
return d_disc, d_nondisc
def fit(self, lr=1e-5, batch_size=128, n_epochs=5, verbose=True):
train_dataset = TensorDataset(self.train_features, self.train_labels)
train_loader = DataLoader(train_dataset, batch_size=batch_size)
params = list(self.F_linear.parameters()) + list(self.D_linear.parameters())
optimizer = optim.Adam(params, lr=lr)
log.info('********Start Training Decomposing Representations********')
n_print = max(1, int(0.5*len(train_dataset)//batch_size)) if verbose else -1
for epoch in range(n_epochs):
running_loss = 0
for i, (inputs, labels) in enumerate(train_loader):
optimizer.zero_grad()
d_disc, d_nondisc = self.forward(inputs)
loss = iCE(d_disc, d_nondisc, labels, self.output_dim)
loss.backward()
optimizer.step()
running_loss += loss.item()
if i % n_print == (n_print-1):
log.info(f'epoch: {epoch + 1}, step: {i + 1:5d}, loss: {running_loss / n_print:.3f}')
running_loss = 0.0
log.info('********Finished Training********')
return self
class UeEstimatorDecomposing:
def __init__(self, cls, ue_args, config, train_dataset):
self.cls = cls
self.ue_args = ue_args
self.config = config
self.train_dataset = train_dataset
def __call__(self, X, y):
return self._predict_with_trained_decomposition(X, y)
def fit_ue(self, X, y=None, X_test=None):
cls = self.cls
model = self.cls._auto_model
log.info("****************Start fitting**************")
if y is None:
y = self._exctract_labels(X)
self._replace_model_head()
X_features = self._exctract_features(X)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
X_features = torch.Tensor(X_features).to(self.device)
y = torch.LongTensor(y).to(self.device)
seed_everything(self.config.training.seed)
self.dec_model = DecomposingModel(X_features, y, use_spectral_norm=False).to(self.device)
self.dec_model.fit(lr=self.ue_args.lr, batch_size=self.ue_args.batch_size, n_epochs=self.ue_args.n_epochs)
X_d_disc, X_d_nondisc = self.dec_model(X_features)
self.disc_class_cond_centroids = self._fit_centroids(X_d_disc.cpu().detach().numpy(), y.cpu().detach().numpy())
self.disc_class_cond_covariance = self._fit_covariance(X_d_disc.cpu().detach().numpy(), y.cpu().detach().numpy(),
centroids=self.disc_class_cond_centroids)
self.nondisc_class_cond_centroids = self._fit_centroids(X_d_nondisc.cpu().detach().numpy(), y.cpu().detach().numpy())
self.nondisc_class_cond_covariance = self._fit_covariance(X_d_nondisc.cpu().detach().numpy(), y.cpu().detach().numpy(),
centroids=self.nondisc_class_cond_centroids)
log.info("**************Done.**********************")
def _fit_covariance(self, X, y, class_cond=True, centroids=None):
if class_cond:
if centroids is None:
centroids = self.class_cond_centroids
return compute_covariance(centroids, X, y, class_cond)
if centroids is None:
centroids = self.train_centroid
return compute_covariance(centroids, X, y, class_cond)
def _fit_centroids(self, X, y, class_cond=True):
return compute_centroids(X, y, class_cond)
def _replace_model_head(self):
log.info("Change classifier to Identity Pooler")
cls = self.cls
model = self.cls._auto_model
if is_custom_head(model):
model.classifier = ElectraClassificationHeadIdentityPooler(model.classifier)
else:
model.classifier = BertClassificationHeadIdentityPooler(model.classifier)
def _exctract_labels(self, X):
return np.asarray([example["label"] for example in X])
def _exctract_features(self, X):
cls = self.cls
model = self.cls._auto_model
try:
X = X.remove_columns("label")
except:
X.dataset = X.dataset.remove_columns("label")
X_features = cls.predict(X, apply_softmax=False, return_preds=False)[0]
return X_features
def _predict_with_trained_decomposition(self, X, y):
cls = self.cls
model = self.cls._auto_model
log.info("****************Compute MD with fitted covariance and centroids **************")
start = time.time()
if y is None:
y = self._exctract_labels(X)
X_features = self._exctract_features(X)
X_features = torch.Tensor(X_features).to(self.device)
y = torch.LongTensor(y).to(self.device)
d_disc, d_nondisc = self.dec_model(X_features)
end = time.time()
d_disc_md, inf_time1 = mahalanobis_distance(None, None, d_disc.cpu().detach().numpy(),
self.disc_class_cond_centroids,
self.disc_class_cond_covariance)
d_nondisc_md, inf_time2 = mahalanobis_distance(None, None, d_nondisc.cpu().detach().numpy(),
self.nondisc_class_cond_centroids,
self.nondisc_class_cond_covariance)
eval_results = {}
sum_inf_time = (end - start) + inf_time1 + inf_time2
eval_results["disc_md"] = d_disc_md.tolist()
eval_results["nondisc_md"] = d_nondisc_md.tolist()
eval_results["ue_time"] = sum_inf_time
log.info(f"UE time: {sum_inf_time}")
log.info("**************Done.**********************")
return eval_results
class UeEstimatorDecomposingNer:
def __init__(self, cls, ue_args, config, train_dataset):
self.cls = cls
self.ue_args = ue_args
self.config = config
self.train_dataset = train_dataset
def __call__(self, X, y):
return self._predict_with_trained_decomposition(X, y)
def fit_ue(self, X, y=None, X_test=None):
cls = self.cls
model = self.cls._auto_model
log.info("****************Start fitting**************")
if y is None:
y, y_shape = self._exctract_labels(X)
self._replace_model_head()
X_features = self._exctract_features(X)
X_features, y = unpad_features(X_features, y)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
X_features = torch.Tensor(X_features).to(self.device)
y = torch.LongTensor(y).to(self.device)
seed_everything(self.config.training.seed)
self.dec_model = DecomposingModel(X_features, y, use_spectral_norm=False).to(self.device)
self.dec_model.fit(lr=self.ue_args.lr, batch_size=self.ue_args.batch_size, n_epochs=self.ue_args.n_epochs)
X_d_disc, X_d_nondisc = self.dec_model(X_features)
self.disc_class_cond_centroids = self._fit_centroids(X_d_disc.cpu().detach().numpy(), y.cpu().detach().numpy())
self.disc_class_cond_covariance = self._fit_covariance(X_d_disc.cpu().detach().numpy(), y.cpu().detach().numpy(),
centroids=self.disc_class_cond_centroids)
self.nondisc_class_cond_centroids = self._fit_centroids(X_d_nondisc.cpu().detach().numpy(), y.cpu().detach().numpy())
self.nondisc_class_cond_covariance = self._fit_covariance(X_d_nondisc.cpu().detach().numpy(), y.cpu().detach().numpy(),
centroids=self.nondisc_class_cond_centroids)
log.info("**************Done.**********************")
def _fit_covariance(self, X, y, class_cond=True, centroids=None):
if class_cond:
if centroids is None:
centroids = self.class_cond_centroids
return compute_covariance(centroids, X, y, class_cond)
if centroids is None:
centroids = self.train_centroid
return compute_covariance(centroids, X, y, class_cond)
def _fit_centroids(self, X, y, class_cond=True):
return compute_centroids(X, y, class_cond)
def _replace_model_head(self):
log.info("Change classifier to Identity Pooler")
cls = self.cls
model = self.cls._auto_model
if is_custom_head(model):
model.classifier = ElectraNERHeadIdentityPooler(model.classifier)
else:
model.classifier = BertClassificationHeadIdentityPooler(model.classifier)
def _exctract_labels(self, X):
y = np.asarray([example["labels"] for example in X])
y_shape = y.shape
return y.reshape(-1), y_shape
def _exctract_features(self, X, remove_col=True):
cls = self.cls
model = self.cls._auto_model
if remove_col:
try:
X = X.remove_columns("labels")
except:
X.dataset = X.dataset.remove_columns("labels")
X_features = cls.predict(X, apply_softmax=False, return_preds=False)[0]
X_features = X_features.reshape(-1, X_features.shape[-1])
return X_features
def _predict_with_trained_decomposition(self, X, y):
cls = self.cls
model = self.cls._auto_model
log.info("****************Compute MD with fitted covariance and centroids **************")
start = time.time()
y_pad, y_shape = self._exctract_labels(X)
X_features = self._exctract_features(X)
X_features, y = unpad_features(X_features, y_pad)
X_features = torch.Tensor(X_features).to(self.device)
y = torch.LongTensor(y).to(self.device)
d_disc, d_nondisc = self.dec_model(X_features)
end = time.time()
d_disc_md, inf_time1 = mahalanobis_distance(None, None, d_disc.cpu().detach().numpy(),
self.disc_class_cond_centroids,
self.disc_class_cond_covariance)
d_nondisc_md, inf_time2 = mahalanobis_distance(None, None, d_nondisc.cpu().detach().numpy(),
self.nondisc_class_cond_centroids,
self.nondisc_class_cond_covariance)
d_disc_md = pad_scores(d_disc_md, np.asarray(y_pad).reshape(y_shape), y_pad)
d_nondisc_md = pad_scores(d_nondisc_md, np.asarray(y_pad).reshape(y_shape), y_pad)
eval_results = {}
sum_inf_time = (end - start) + inf_time1 + inf_time2
eval_results["disc_md"] = d_disc_md.tolist()
eval_results["nondisc_md"] = d_nondisc_md.tolist()
eval_results["ue_time"] = sum_inf_time
log.info(f"UE time: {sum_inf_time}")
log.info("**************Done.**********************")
return eval_results