-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
151 lines (128 loc) · 5.59 KB
/
model.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from torch.nn import CrossEntropyLoss
from transformers import LayoutLMPreTrainedModel, LayoutLMModel
from sklearn.covariance import EmpiricalCovariance
class LayoutLMForSequenceClassification(LayoutLMPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.layoutlm = LayoutLMModel(config)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
def forward(
self,
input_ids=None,
bbox=None,
attention_mask=None,
token_type_ids=None,
label=None,
):
outputs = self.layoutlm(
input_ids=input_ids,
bbox=bbox,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
)
pooled_output = pooled = outputs[1]
pooled_output = self.dropout(pooled_output)
logits = self.classifier(pooled_output)
loss = None
if label is not None:
if self.config.loss == 'margin':
dist = ((pooled.unsqueeze(1) - pooled.unsqueeze(0)) ** 2).mean(-1)
mask = (label.unsqueeze(1) == label.unsqueeze(0)).float()
mask = mask - torch.diag(torch.diag(mask))
neg_mask = (label.unsqueeze(1) != label.unsqueeze(0)).float()
max_dist = (dist * mask).max()
cos_loss = (dist * mask).sum(-1) / (mask.sum(-1) + 1e-3) + (F.relu(max_dist - dist) * neg_mask).sum(-1) / (neg_mask.sum(-1) + 1e-3)
cos_loss = cos_loss.mean()
else:
norm_pooled = F.normalize(pooled, dim=-1)
cosine_score = torch.exp(norm_pooled @ norm_pooled.t() / 0.3)
mask = (label.unsqueeze(1) == label.unsqueeze(0)).float()
cosine_score = cosine_score - torch.diag(torch.diag(cosine_score))
mask = mask - torch.diag(torch.diag(mask))
cos_loss = cosine_score / cosine_score.sum(dim=-1, keepdim=True)
cos_loss = -torch.log(cos_loss + 1e-5)
cos_loss = (mask * cos_loss).sum(-1) / (mask.sum(-1) + 1e-3)
cos_loss = cos_loss.mean()
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), label.view(-1))
loss = loss + self.config.alpha * cos_loss
output = (logits,) + outputs[2:]
output = output + (pooled,)
return ((loss, cos_loss) + output) if loss is not None else output
def compute_ood(
self,
input_ids=None,
bbox=None,
attention_mask=None,
token_type_ids=None,
label=None,
):
outputs = self.layoutlm(
input_ids=input_ids,
bbox=bbox,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
)
pooled_output = pooled = outputs[1]
pooled_output = self.dropout(pooled_output)
logits = self.classifier(pooled_output)
ood_keys = None
softmax_score = F.softmax(logits, dim=-1).max(-1)[0]
maha_score = []
for c in self.all_classes:
centered_pooled = pooled - self.class_mean[c].unsqueeze(0)
ms = torch.diag(centered_pooled @ self.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 @ self.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(),
}
return ood_keys
def prepare_ood(self, dataloader=None):
self.bank = None
self.label_bank = None
# device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
for batch in dataloader:
self.eval()
batch = {key: value.cuda() for key, value in batch.items()}
label = batch['label']
outputs = self.layoutlm(
input_ids=batch['input_ids'],
bbox=batch['bbox'],
attention_mask=batch['attention_mask'],
token_type_ids=batch['token_type_ids'],
)
pooled = outputs[1]
if self.bank is None:
self.bank = pooled.clone().detach()
self.label_bank = label.clone().detach()
else:
bank = pooled.clone().detach()
label_bank = label.clone().detach()
self.bank = torch.cat([bank, self.bank], dim=0)
self.label_bank = torch.cat([label_bank, self.label_bank], dim=0)
self.norm_bank = F.normalize(self.bank, dim=-1)
N, d = self.bank.size()
self.all_classes = list(set(self.label_bank.tolist()))
self.class_mean = torch.zeros(max(self.all_classes) + 1, d).cuda()
for c in self.all_classes:
self.class_mean[c] = (self.bank[self.label_bank == c].mean(0))
centered_bank = (self.bank - self.class_mean[self.label_bank]).detach().cpu().numpy()
precision = EmpiricalCovariance().fit(centered_bank).precision_.astype(np.float32)
self.class_var = torch.from_numpy(precision).float().cuda()