-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrainer.py
201 lines (185 loc) · 9.07 KB
/
trainer.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
import torch
import torch.nn as nn
import numpy
import itertools as it
from dataprocessor import DataProcessor
from torch.nn.functional import cross_entropy, mse_loss
from utils.contrastive_loss import SupConLoss
def one_hot(x, class_count):
# 第一构造一个[class_count, class_count]的对角线为1的向量
# 第二保留label对应的行并返回
return torch.eye(class_count)[x,:]
class Trainer():
def __init__(self,
method: str = "baseline",
epoch: int = 50,
batch_size: int = 64,
class_num: int = 10,
instructor: nn.Module = None,
learner: nn.Module = None,
contrastive_temp: float = 0.5,
optimizer_learner: torch.optim = None,
optimizer_retrain_learner: torch.optim = None,
optimizer_instructor: torch.optim = None,
data_processor: DataProcessor = None,
writer: torch.utils.tensorboard.SummaryWriter = None,
device: str = "cpu"
):
self.method = method
self.epochs = epoch
self.batch_size = batch_size
self.class_num = class_num
self.instructor = instructor
self.learner = learner
self.optimizer_learner = optimizer_learner
self.optimizer_retrain_learner = optimizer_retrain_learner
self.optimizer_instructor = optimizer_instructor
self.data_processor = data_processor
self.criterion = torch.nn.CrossEntropyLoss(reduction='none')
self.contrastive_loss = SupConLoss(temperature=contrastive_temp)
self.device = device
self.writer = writer
def train(self):
data_loader_train, data_loader_val, data_loader_test = self.data_processor.load_data()
data_loader_error = self.data_processor.build_error_set()
data_loader_train_and_error = self.data_processor.build_train_and_error_set()
train_iteration = 0
retrain_iteration = 0
instructor_train_iteration = 0
train_losses = [0.]
retrain_losses = [0.]
for epoch in range(self.epochs):
# train learner model
self.learner.train()
for images, labels, indices in iter(data_loader_train):
images = images.to(self.device)
labels = labels.to(self.device)
pred_y, features = self.learner(images)
learner_loss = cross_entropy(pred_y, labels)
self.optimizer_learner.zero_grad()
learner_loss.backward()
self.optimizer_learner.step()
_, learner_id = torch.max(pred_y.data, 1)
learner_nums_correct = torch.sum(learner_id == labels.data)
self.writer.add_scalar("loss/train_learner", learner_loss.item(), train_iteration)
self.writer.add_scalar("accuracy/train_learner", learner_nums_correct / len(images), train_iteration)
train_iteration += 1
# build the error set
self.learner.eval()
error_indices = []
for images, labels, indices in iter(data_loader_val):
images = images.to(self.device)
labels = labels.to(self.device)
with torch.no_grad():
pred_y, features = self.learner(images)
_, id = torch.max(pred_y.data, 1)
error_indices += indices[torch.nonzero(id != labels.data)].squeeze(1).tolist()
print(f"The number of incorrect examples is {len(error_indices)}")
self.writer.add_scalar("numbers/error_set", len(error_indices), epoch)
self.data_processor.update_error_indices(error_indices)
# train the instructor model
self.learner.eval()
self.instructor.train()
for images, labels, indices in iter(data_loader_train_and_error):
images = images.to(self.device)
labels = labels.to(self.device)
pred_y, feature = self.learner(images)
input = torch.concat([pred_y, feature], dim=1)
output = self.instructor(input)
id = output.data >= 0.5
target = id != labels.data
loss = cross_entropy(output, target)
self.writer.add_scalar("loss/train_instructor_on_t&e_set", loss.item(), instructor_train_iteration)
self.writer.add_scalar("accuracy/instructor_on_t&e_set", nums_correct / len(images), instructor_train_iteration)
self.optimizer_instructor.zero_grad()
loss.backward()
self.optimizer_instructor.step()
instructor_train_iteration += 1
# target_matrix = torch.zeros([50000, self.class_num]).to(self.device)
# for images, labels, indices in iter(data_loader_train_and_error):
# with torch.no_grad():
# target = self.instructor(self.learner.feature_output(images, labels, indices)).softmax(dim=1)
# target_matrix[indices] = target
# retrain the learner
self.learner.train()
num = 0
if self.method == "baseline":
data_loader = data_loader_train_and_error
elif self.method == "upper_bound":
data_loader = data_loader_val
elif self.method == "our_method":
data_loader = data_loader_train
elif self.method == "our_method_t_e":
data_loader = data_loader_train_and_error
else:
data_loader = data_loader_train
for images, labels, indices in iter(data_loader):
images = images.to(self.device)
labels = labels.to(self.device)
pred_y, feature = self.learner(images, retrain=True)
v_loss = - self.instructor(pred_y, feature).mean()
ce_loss = cross_entropy(pred_y, labels)
if self.method == "baseline":
loss = ce_loss
if self.method == "our_method":
loss = ce_loss + v_loss
self.optimizer_retrain_learner.zero_grad()
loss.backward()
self.optimizer_retrain_learner.step()
retrain_losses.append(loss.item())
_, id = torch.max(pred_y.data, 1)
nums_correct = torch.sum(id == labels.data)
self.writer.add_scalar("loss/retrain_v_loss", v_loss.item(), retrain_iteration)
self.writer.add_scalar("loss/retrain_ce_loss", ce_loss.item(), retrain_iteration)
self.writer.add_scalar("accuracy/retrain_learner", nums_correct / len(images), retrain_iteration)
retrain_iteration += 1
num += len(images)
# test
print(num)
self.learner.eval()
self.instructor.eval()
error_res = self.test(data_loader_error)
test_res = self.test(data_loader_test)
self.writer.add_scalar("accuracy/test_learner", test_res['acc_learner'], epoch)
self.writer.add_scalar("accuracy/error_learner", error_res['acc_learner'], epoch)
self.writer.add_scalar("accuracy/test_instructor", test_res['acc_ins'], epoch)
self.writer.add_scalar("accuracy/error_instructor", error_res['acc_ins'], epoch)
def test(self, data_loader):
result = {}
test_correct_learner = 0
test_correct_ins = 0
num_dataset = 0
for images, labels, indices in iter(data_loader):
images = images.to(self.device)
labels = labels.to(self.device)
with torch.no_grad():
pred_y_learner, _ = self.learner(images)
_, id_learner = torch.max(pred_y_learner.data, 1)
test_correct_learner += torch.sum(id_learner == labels.data)
num_dataset += len(images)
result['acc_learner'] = test_correct_learner / num_dataset
result['acc_ins'] = test_correct_ins / num_dataset
return result
def train_learner(self):
data_loader_train, data_loader_val, data_loader_test = self.data_processor.load_data()
for epoch in range(self.epochs):
# train learner model
for images, labels, indices in iter(data_loader_train):
images = images.to(self.device)
labels = labels.to(self.device)
pred_y, features = self.learner(images)
loss = self.criterion(pred_y, labels)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
print(f"loss:{loss.item()}")
# evaluate the performance on validation set
self.learner.eval()
test_correct = 0
for images, labels, indices in iter(data_loader_test):
images = images.to(self.device)
labels = labels.to(self.device)
outputs = self.learner(images)
_, id = torch.max(outputs.data, 1)
test_correct += torch.sum(id == labels.data)
print("correct:%.3f%%" % (test_correct / len(self.data_processor.data_test)))