def pretrain():
train = FabricDataset(root='./data/X_fabric_data/tr/')
test = FabricDataset(root='./data/X_fabric_data/te/')
train = DataLoader(train, batch_size=128, shuffle=True, num_workers=4)
test = DataLoader(test, batch_size=128, shuffle=True, num_workers=4)
MAX_EPOCHES = 20
lr = 0.1
model = fabric()
model = model.to(device)
if cuda.is_available():
model = torch.nn.DataParallel(model)
opt1 = torch.optim.Adam(model.parameters())
criterion = nn.CrossEntropyLoss()
for epoch in range(MAX_EPOCHES):
for xx, xt, Y in train:
xx = xx.to(device)
xt = xt.to(device)
c = model(xx, xt)
Y = Y.to(device)
loss = criterion(c, Y)
opt1.zero_grad()
loss.backward()
opt1.step()
print(loss.item())
with torch.no_grad():
right = 0
su = 0
for xx, xt, Y in test:
xx = xx.to(device)
xt = xt.to(device)
c = model(xx, xt)
su = su + Y.shape[0]
right = right + (c.argmax(dim=1)
== Y.to(device)).float().sum().cpu().item()
print('test', epoch, right / su)
if cuda.is_available():
return model.module.feature
else:
return model.feature
def finetune(feature):
xxx = torchvision.models.resnet18(num_classes=4)
classifier = xxx.fc
train = FabricDataset(root='./data/DIY_fabric_data/tr/')
test = FabricDataset(root='./data/DIY_fabric_data/te/')
train = DataLoader(train, batch_size=128, shuffle=True, num_workers=4)
test = DataLoader(test, batch_size=128, shuffle=True, num_workers=4)
MAX_EPOCHES = 30
lr = 0.1
feature = feature.to(device)
classifier = classifier.to(device)
if cuda.is_available():
feature = torch.nn.DataParallel(feature)
classifier = torch.nn.DataParallel(classifier)
opt1 = torch.optim.Adam([{
'params': feature.parameters(),
'lr': 0.0001
}, {
'params': classifier.parameters(),
'lr': 0.01
}])
criterion = nn.CrossEntropyLoss()
avgpool = xxx.avgpool
for epoch in range(MAX_EPOCHES):
for xx, xt, Y in train:
xx = xx.to(device)
xt = xt.to(device)
y = feature(xx)
ytemp = feature(xt)
y = avgpool(y)
ytemp = avgpool(ytemp)
c = classifier(torch.flatten(y - ytemp, 1))
Y[Y == 1] = 0
Y[Y == 2] = 1
Y[Y == 5] = 2
Y[Y == 13] = 3
Y = Y.to(device)
loss = criterion(c, Y)
opt1.zero_grad()
loss.backward()
opt1.step()
print(loss.item())
with torch.no_grad():
right = 0
su = 0
for xx, xt, Y in test:
xx = xx.to(device)
xt = xt.to(device)
y = feature(xx)
ytemp = feature(xt)
y = avgpool(y)
ytemp = avgpool(ytemp)
c = classifier(torch.flatten(y - ytemp, 1))
su = su + Y.shape[0]
Y[Y == 1] = 0
Y[Y == 2] = 1
Y[Y == 5] = 2
Y[Y == 13] = 3
right = right + (c.argmax(dim=1)
== Y.to(device)).float().sum().cpu().item()
print('test', epoch, right / su)
self.fc = nn.Linear(512 * 7 * 7, 4)
def forward(self, xx, xt):
y = self.pretrained(xx)
ytemp = self.pretrained(xt)
c = self.fc(torch.flatten(y-ytemp, 1))
return c
if __name__ == "__main__":
# tainsform未生效
transform = transforms.Compose([transforms.ToTensor(),
transforms.Resize((224, 224)),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
train = FabricDataset(root='./data/DIY_fabric_data/tr/')
test = FabricDataset(root='./data/DIY_fabric_data/te/')
train = DataLoader(train, batch_size=128, shuffle=True, num_workers=4)
test = DataLoader(test, batch_size=128, shuffle=True, num_workers=4)
MAX_EPOCHES = 100
lr = 0.1
model = fabric()
model = model.to(device)
if cuda.is_available():
model = torch.nn.DataParallel(model)
'''
opt1=torch.optim.Adam([
{'params':model.module.pretrained.parameters(),'lr':0.000001},
def finetune(feature, f2):
train = FabricDataset(root='./data/DIY_fabric_data/tr/')
test = FabricDataset(root='./data/DIY_fabric_data/te/')
train = DataLoader(train, batch_size=128, shuffle=True, num_workers=4)
test = DataLoader(test, batch_size=128, shuffle=True, num_workers=4)
MAX_EPOCHES = 30
lr = 0.1
classifier = nn.Sequential(
nn.Dropout(),
nn.Linear(256 * 12 * 6, 4096),
nn.ReLU(inplace=True),
nn.Dropout(),
nn.Linear(4096, 4096),
nn.ReLU(inplace=True),
nn.Linear(4096, 4),
)
feature = feature.to(device)
f2 = f2.to(device)
classifier = classifier.to(device)
if cuda.is_available():
feature = torch.nn.DataParallel(feature)
f2 = torch.nn.DataParallel(f2)
classifier = torch.nn.DataParallel(classifier)
opt1 = torch.optim.Adam([{
'params': feature.parameters(),
'lr': 0.0001
}, {
'params': classifier.parameters(),
'lr': 0.01
}])
criterion = nn.CrossEntropyLoss()
avgpool = torchvision.models.AlexNet(4).avgpool
for epoch in range(MAX_EPOCHES):
for xx, xt, Y in train:
xx = xx.to(device)
xt = xt.to(device)
y1 = feature(xx)
ytemp1 = feature(xt)
y2 = f2(y1)
ytemp2 = f2(ytemp1)
y1 = avgpool(y1)
ytemp1 = avgpool(ytemp1)
y2 = avgpool(y2)
ytemp2 = avgpool(ytemp2)
c = classifier(
torch.cat((torch.flatten(
y1 - ytemp1, 1), torch.flatten(y2 - ytemp2, 1)), 1))
Y[Y == 1] = 0
Y[Y == 2] = 1
Y[Y == 5] = 2
Y[Y == 13] = 3
Y = Y.to(device)
loss = criterion(c, Y)
opt1.zero_grad()
loss.backward()
opt1.step()
print(loss.item())
with torch.no_grad():
right = 0
su = 0
for xx, xt, Y in test:
xx = xx.to(device)
xt = xt.to(device)
y1 = feature(xx)
ytemp1 = feature(xt)
y2 = f2(y1)
ytemp2 = f2(ytemp1)
y1 = avgpool(y1)
ytemp1 = avgpool(ytemp1)
y2 = avgpool(y2)
ytemp2 = avgpool(ytemp2)
c = classifier(
torch.cat((torch.flatten(
y1 - ytemp1, 1), torch.flatten(y2 - ytemp2, 1)), 1))
su = su + Y.shape[0]
Y[Y == 1] = 0
Y[Y == 2] = 1
Y[Y == 5] = 2
Y[Y == 13] = 3
right = right + (c.argmax(dim=1)
== Y.to(device)).float().sum().cpu().item()
print('test', epoch, right / su)