-
Notifications
You must be signed in to change notification settings - Fork 0
/
engine.py
167 lines (130 loc) · 5.97 KB
/
engine.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
import time
import numpy as np
import torch
import torch.cuda.amp as amp
from metrics import quadratic_weighted_kappa
def train(model, optimizer, data_loader):
st = time.time()
losses = []
model.train()
scaler = amp.GradScaler()
for input_ids, attention_mask, features, labels in data_loader:
input_ids = input_ids.cuda(non_blocking=True)
attention_mask = attention_mask.cuda(non_blocking=True)
features = features.cuda(non_blocking=True)
batch_size, num_features = features.shape
optimizer.zero_grad()
with amp.autocast():
pred_score, weight_memory = model(input_ids, attention_mask)
total_mask = torch.ones((batch_size, batch_size))
idx_pairs = torch.nonzero(total_mask).cuda()
features_a = features[idx_pairs[:, 0], :] # [num_pairs, num_features]
features_b = features[idx_pairs[:, 1], :] # [num_pairs, num_features]
ge_mask = torch.where(features_a >= features_b, weight_memory[None, :], 1. - weight_memory[None, :])
pred_score_a = pred_score[idx_pairs[:, 0], :]
pred_score_b = pred_score[idx_pairs[:, 1], :]
term = pred_score_a.exp() / (pred_score_a.exp() + pred_score_b.exp())
loss = -torch.log(ge_mask * term + (1 - ge_mask) * (1 - term)).mean()
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
losses.append(loss.detach())
return torch.stack(losses).mean(), time.time() - st
@torch.no_grad()
def evaluate(model, data_loader, num_classes):
st = time.time()
model.eval()
pred_scores_list, gt_labels_list = [], []
for input_ids, attention_mask, _, cls_labels in data_loader:
input_ids = input_ids.cuda(non_blocking=True)
attention_mask = attention_mask.cuda(non_blocking=True)
cls_labels = cls_labels.cuda(non_blocking=True)
with amp.autocast():
pred_scores, weight_memory = model(input_ids, attention_mask)
pred_scores_list.append(pred_scores)
gt_labels_list.append(cls_labels)
pred_scores_list = torch.cat(pred_scores_list, dim=0).cpu().squeeze()
gt_labels_list = torch.cat(gt_labels_list, dim=0).cpu()
reg_scores = reg_scoring(pred_scores_list, num_classes)
reg_qwk = quadratic_weighted_kappa(gt_labels_list, reg_scores)
gt_scores = gt_scoring(pred_scores_list, gt_labels_list)
gt_qwk = quadratic_weighted_kappa(gt_labels_list, gt_scores)
u_scores = uniform_scoring(pred_scores_list, num_classes)
u_qwk = quadratic_weighted_kappa(gt_labels_list, u_scores)
t_scores = tri_scoring(pred_scores_list, num_classes)
t_qwk = quadratic_weighted_kappa(gt_labels_list, t_scores)
n_scores = normal_scoring(pred_scores_list, num_classes)
n_qwk = quadratic_weighted_kappa(gt_labels_list, n_scores)
return gt_qwk, u_qwk, reg_qwk, t_qwk, n_qwk, time.time() - st
def gt_scoring(pred_scores_list, gt_labels_list):
gt_labels_sorted, gt_indices_sorted = torch.sort(gt_labels_list)
gt_labels_set = sorted(set(gt_labels_sorted.numpy()))
pred_scores_sorted, pred_indices_sorted = torch.sort(pred_scores_list)
scores = torch.zeros(len(pred_scores_sorted))
for label in gt_labels_set:
ids = torch.nonzero(gt_labels_sorted == label).squeeze()
scores[pred_indices_sorted[ids]] = int(label)
return scores
def normal_scoring(pred_scores_list, num_classes):
pred_scores_sorted, pred_indices_sorted = torch.sort(pred_scores_list)
k = 1 / np.sqrt(2 * np.pi) * torch.exp(-(torch.arange(num_classes) - (num_classes - 1) / 2) ** 2 / 2)
k = k / k.sum()
num_list = [int(np.floor(len(pred_scores_list) * i)) for i in k]
count = len(pred_scores_list) - sum(num_list)
for i in range(num_classes):
num_list[i] += 1
count -= 1
if count == 0:
break
scores = torch.zeros(len(pred_scores_list))
sorted_labels = []
for i, a in enumerate(num_list):
sorted_labels.extend([i for _ in range(a)])
sorted_labels = torch.tensor(sorted_labels, dtype=torch.float)
scores[pred_indices_sorted] = sorted_labels
return scores
def tri_scoring(pred_scores_list, num_classes):
pred_scores_sorted, pred_indices_sorted = torch.sort(pred_scores_list)
num_samples = len(pred_scores_list)
# print(num_samples)
k = -torch.abs(torch.arange(num_classes) - 1 - (num_classes - 1) / 2) + (num_classes + 1) / 2
k /= torch.sum(k)
# print(k)
# k = np.floor(num_classes / 2)
# if num_classes % 2 == 0:
# x = 1 / (k * (k + 1))
# num_list = [-np.abs(i + 1 - (k + 0.5)) + k + 0.5 for i in range(num_classes)]
# else:
# x = 1 / ((k + 1) * (k + 1))
# num_list = [-np.abs(i + 1 - (k + 1)) + k + 1 for i in range(num_classes)]
num_list = [int(np.floor(a * num_samples)) for a in k]
# print(len(num_list))
# print(num_list)
count = num_samples - sum(num_list)
for i in range(num_classes):
if count <= 0:
break
num_list[i] += 1
count -= 1
# print(num_list)
scores = torch.zeros(num_samples)
sorted_labels = []
for i, a in enumerate(num_list):
sorted_labels.extend([i for _ in range(a)])
sorted_labels = torch.tensor(sorted_labels, dtype=torch.float)
# print(sorted_labels.shape)
scores[pred_indices_sorted] = sorted_labels
return scores
def uniform_scoring(pred_scores_list, num_classes):
pred_scores_sorted, pred_indices_sorted = torch.sort(pred_scores_list)
scores = torch.zeros(len(pred_scores_sorted))
scores[pred_indices_sorted] = torch.arange(len(pred_scores_sorted)).float()
scores = scores / (len(pred_scores_sorted)) * num_classes
scores = torch.round(scores).long()
return scores
def reg_scoring(pred_scores_list, num_classes):
pred_scores_list -= min(pred_scores_list.clone())
pred_scores_list /= max(pred_scores_list.clone())
pred_scores_list *= (num_classes - 1)
scores = torch.round(pred_scores_list.float()).long()
return scores