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
| 定义基础类 BaseModel
class BaseModel:
def __init__(self, args, net, weights):
self.args = args
self.weights = weights
self.net = net.to(args.device)
def load(self):
try:
if os.path.exists(self.weights):
print('Loading classifier...')
self.net.load_state_dict(torch.load(self.weights))
except:
print('Failed to load pre-trained network')
def save(self):
print('Saving model to "{}"'.format(self.weights))
torch.save(self.net.state_dict(), self.weights)
继承该class:
class Model(BaseModel):
def __init__(self, args, net):
super(Model, self).__init__(args, net, args.net_name)
def train(self):
train_loss和validation_loss根据epochs数量初始化
train_loss = np.zeros(self.args.epochs)
声明Dataset和Dataloader
定义criterion、optimizer以及学习率调整策略scheduler
for epoch in range(self.args.epochs):
self.net.train()
学习率迭代 scheduler.step()
for index, (images, labels) in enumerate(train_loader):
images, labels = images.to(self.args.device), labels.to(self.args.device)
optimizer.zero_grad()
output = self.net(images)
loss = criterion(output, labels)
loss.backward()
optimizer.step()
_, predicted = torch.max(output.data, 1)
correct += torch.sum(predicted == labels)
train_loss[epoch] += loss.item() * self.args.batch_size
total += len(images)
if index > 0 and index % self.args.log_interval == 0:
print('Epoch: {}/{} [{}/{} ({:.0f}%)]\tLoss: {:.6f}, Accuracy: {:.2f}%'.format(
epoch + 1,
self.args.epochs,
index * len(images),
len(train_loader.dataset),
100. * index / len(train_loader),
loss.item(),
100 * correct / total
))
train_loss[epoch] = train_loss[epoch] / total
validate_loss[epoch], validate_acc = self.validate()
if validate_acc > best_acc:
best_acc = validate_acc
self.save()
print('\nEnd of training, best class accuracy: {}'.format(best_acc))
plt.plot(train_loss,'b-',label='train_loss')
plt.plot(validate_loss,'r--',label='validate_loss')
plt.xlabel('epoches')
plt.title("Loss curve")
plt.legend()
plt.savefig(self.args.net_name+'.jpg')
def validate(self, verbose=False):
self.net.eval()
loss = 0
correct = 0
dataset=TrainDataset(path=self.args.dataset_path + VALIDATION_PATH,
stride=self.args.validate_stride)
validate_loader = DataLoader(
dataset=dataset,
batch_size=self.args.batch_size,
shuffle=False,
num_workers=0
)
criterion = nn.CrossEntropyLoss(size_average=False)
print('\nEvaluating classifier')
with torch.no_grad():
for images, labels in validate_loader:
images = images.to(self.args.device)
output = self.net(images)
loss += criterion(output.cpu(), labels).item()
_, predicted = torch.max(output.cpu(), 1)
correct += (predicted == labels).sum().item()
loss /= len(dataset)
acc = correct / len(dataset)
print('Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)'.format(
loss,
correct,
len(dataset),
100. * correct / len(dataset)
))
return loss, acc
def ctest(self, path, augment=False):
self.load()
self.net.eval()
dataset = cTestDataset(path=path, stride=self.args.test_stride, augment=augment)
data_loader = DataLoader(dataset=dataset, batch_size=1, shuffle=False)
stime = datetime.datetime.now()
print('\t sum\t\t max\t\t maj')
labels = []
maj_pred = []
prop_pred = np.zeros((len(dataset), len(LABELS)), dtype=float)
maj_correct = 0
with torch.no_grad():
for index, (image, label) in enumerate(data_loader):
image = image.squeeze()
image = image.to(self.args.device)
output = self.net(image)
_, predicted = torch.max(output.data, 1)
sum_prob = 3 - np.argmax(np.sum(np.exp(output.data.cpu().numpy()), axis=0)[::-1])
max_prob = 3 - np.argmax(np.max(np.exp(output.data.cpu().numpy()), axis=0)[::-1])
maj_prob = 3 - np.argmax(np.sum(np.eye(4)[np.array(predicted).reshape(-1)], axis=0)[::-1])
labels.append(label.item())
maj_pred.append(maj_prob)
prop_pred[index, :] = np.sum(np.eye(4)[np.array(predicted).reshape(-1)], axis=0) / len(image)
print('{}) \t {} \t {} \t {} \t {}'.format(
str(index + 1).rjust(2, '0'),
LABELS[sum_prob].ljust(8),
LABELS[max_prob].ljust(8),
LABELS[maj_prob].ljust(8),
str(label.data)))
maj_correct += torch.sum(maj_prob == label).item()
maj_acc = maj_correct / len(dataset)
print('Maj_image accuracy: {}/{} ({:.2f}%)'.format(
maj_correct,
len(data_loader.dataset),
100. * maj_acc
))
print('\nInference time: {}\n'.format(datetime.datetime.now() - stime))
return prop_pred, labels
|
