def compute_y(self,x,targets,mixup=False, visible_mixup=False):
target_soft = to_one_hot(targets,10)
if visible_mixup:
lam = Variable(torch.from_numpy(np.array([np.random.beta(0.5,0.5)]).astype('float32')).cuda())
x, target_soft = mixup_process(x, target_soft, lam=lam)
h = self.h1(x)
if mixup:
lam = Variable(torch.from_numpy(np.array([np.random.beta(0.5,0.5)]).astype('float32')).cuda())
h, target_soft = mixup_process(h, target_soft, lam=lam)
y = self.sm(self.h2(h))
return y, target_soft
def forward(self, x, lam=None, target=None, target_reweighted=None, layer_mix='rand'):
if layer_mix == 'rand':
if self.mixup_hidden:
layer_mix = random.randint(0,2)
else:
layer_mix = 0
out = x
if lam is not None:
if target_reweighted is None:
target_reweighted = to_one_hot(target,self.num_classes)
else:
assert target is None
if layer_mix == 0:
out, target_reweighted = mixup_process(out, target_reweighted, lam)
out = self.conv1(out)
out = self.layer1(out)
if lam is not None and layer_mix == 1:
out, target_reweighted = mixup_process(out, target_reweighted, lam=lam)
out = self.layer2(out)
if lam is not None and layer_mix == 2:
out, target_reweighted = mixup_process(out, target_reweighted, lam=lam)
out = self.layer3(out)
out = self.layer4(out)
out = F.avg_pool2d(out, 4)
out = out.view(out.size(0), -1)
out = self.linear(out)
if lam is None:
return out
else:
return out, target_reweighted
def forward(self, x, y=None, mixup=False):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
x = F.max_pool2d(x, 2)
x = F.relu(self.conv4(x))
x = F.relu(self.conv5(x))
x = F.relu(self.conv6(x))
x = F.max_pool2d(x, 2)
x = x.view(x.size(0), -1)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
if mixup:
lam = Variable(torch.from_numpy(np.array([np.random.beta(0.5,0.5)]).astype('float32')).cuda())
x, y = utils.mixup_process(x, y, lam=lam)
x = F.relu(self.fc3(x))
x = self.fc4(x)
x = F.log_softmax(x, dim=1)
if mixup:
return x, y
else:
return x
def forward(self,
x,
target=None,
mixup=False,
mixup_hidden=False,
mixup_alpha=None,
layer_num_out=None):
#print x.shape
if self.per_img_std:
x = per_image_standardization(x)
if mixup_hidden:
layer_mix = random.randint(0, 2)
elif mixup:
layer_mix = 0
else:
layer_mix = None
out = x
if mixup_alpha is not None:
lam = get_lambda(mixup_alpha)
lam = torch.from_numpy(np.array([lam]).astype('float32')).cuda()
lam = Variable(lam)
if target is not None:
target_reweighted = to_one_hot(target, self.num_classes)
if layer_mix == 0:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.conv1(out)
out = self.layer1(out)
if layer_mix == 1:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
if layer_num_out == 1:
out_tmp = Variable(out.detach().data, requires_grad=False)
out = self.layer2(out)
if layer_mix == 2:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
if layer_num_out == 2:
out_tmp = Variable(out.detach().data, requires_grad=False)
out = self.layer3(out)
if layer_mix == 3:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
if layer_num_out == 3:
out_tmp = Variable(out.detach().data, requires_grad=False)
out = act(self.bn1(out))
out = F.avg_pool2d(out, 8)
out = out.view(out.size(0), -1)
out = self.linear(out)
if layer_num_out == 4:
out_tmp = Variable(out.detach().data, requires_grad=False)
if layer_num_out is not None:
return out, target_reweighted, out_tmp
if target is not None:
return out, target_reweighted
else:
return out
def forward(self,
x,
target=None,
mixup=False,
mixup_hidden=False,
mixup_alpha=None):
#import pdb; pdb.set_trace()
if self.per_img_std:
x = per_image_standardization(x)
if mixup_hidden:
layer_mix = random.randint(0, 2)
elif mixup:
layer_mix = 0
else:
layer_mix = None
out = x
if mixup_alpha is not None:
lam = get_lambda(mixup_alpha)
lam = torch.from_numpy(np.array([lam]).astype('float32')).cuda()
lam = Variable(lam)
if target is not None:
target_reweighted = to_one_hot(target, self.num_classes)
if layer_mix == 0:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.conv1(out)
out = self.layer1(out)
if layer_mix == 1:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.layer2(out)
if layer_mix == 2:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.layer3(out)
if layer_mix == 3:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.layer4(out)
out = F.avg_pool2d(out, 4)
out = out.view(out.size(0), -1)
out = self.linear(out)
if target is not None:
return out, target_reweighted
else:
return out
def forward(self,
x,
target=None,
mixup=False,
mixup_hidden=False,
mixup_alpha=None):
if self.per_img_std:
x = per_image_standardization(x)
if mixup_hidden:
layer_mix = random.randint(0, 2)
elif mixup:
layer_mix = 0
else:
layer_mix = None
out = x
if mixup_alpha is not None:
lam = get_lambda(mixup_alpha)
lam = torch.from_numpy(np.array([lam]).astype('float32')).cuda()
lam = Variable(lam)
if target is not None:
target_reweighted = to_one_hot(target, self.num_classes)
if layer_mix == 0:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.conv_1_3x3(out)
out = F.relu(self.bn_1(out), inplace=True)
out = self.stage_1(out)
if layer_mix == 1:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.stage_2(out)
if layer_mix == 2:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.stage_3(out)
if layer_mix == 3:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.avgpool(out)
out = out.view(out.size(0), -1)
if self.dropout:
out = F.dropout(out, p=0.5, training=self.training)
out = self.classifier(out)
if target is not None:
return out, target_reweighted
else:
return out
def forward(self, x, target= None, mixup=False, mixup_hidden=False, mixup_alpha=None, MyMixup = False, MyMixup_num_layer = 4):
#import pdb; pdb.set_trace()
if self.per_img_std:
x = per_image_standardization(x)
if mixup_hidden:
layer_mix = random.randint(0,2)
elif mixup:
layer_mix = 0
else:
layer_mix = None
out = x
if mixup_alpha is not None:
lam = get_lambda(mixup_alpha)
lam = torch.from_numpy(np.array([lam]).astype('float32')).cuda()
lam = Variable(lam)
# lam은 tensor형 소수
# tensor([0.7119], device='cuda:0')
###########################################각 layer에 맞게 output과 target값을 얻어냄############################
if target is not None :
target_reweighted = to_one_hot(target,self.num_classes)
if MyMixup:
if MyMixup_num_layer >= 0:
mixup_out, target_reweighted = mixup_process(out, target_reweighted, lam=lam) #shape : torch.Size([100, 3, 32, 32])
if MyMixup_num_layer >= 1:
mixup_out = self.conv1(mixup_out)
mixup_out = self.layer1(mixup_out) # shape : torch.Size([100, 64, 32, 32])
if MyMixup_num_layer >= 2:
mixup_out = self.layer2(mixup_out) # shape : torch.Size([100, 128, 16, 16])
if MyMixup_num_layer >= 3:
mixup_out = self.layer3(mixup_out) #shape : torch.Size([100, 256, 8, 8])
with torch.no_grad():
if MyMixup_num_layer >= 1:
out = self.conv1(out)
out = self.layer1(out) # shape : torch.Size([100, 64, 32, 32])
if MyMixup_num_layer >= 2:
out = self.layer2(out) # shape : torch.Size([100, 128, 16, 16])
if MyMixup_num_layer >= 3:
out = self.layer3(out) #shape : torch.Size([100, 256, 8, 8])
return mixup_out, out
else:
if layer_mix == 0:
out, target_reweighted = mixup_process(out, target_reweighted, lam=lam) #input값을 섞어줌
#out.shape : torch.Size([100, 3, 32, 32])
out = self.conv1(out)
out = self.layer1(out)
# out.shape : torch.Size([100, 64, 32, 32])
if layer_mix == 1:
out, target_reweighted = mixup_process(out, target_reweighted, lam=lam) #1번 layer의 output을 섞어줌
out = self.layer2(out)
# out.shape : torch.Size([100, 128, 16, 16])
if layer_mix == 2:
out, target_reweighted = mixup_process(out, target_reweighted, lam=lam) #2번 layer의 output을 섞어줌
out = self.layer3(out)
# out.shape : torch.Size([100, 256, 8, 8])
if layer_mix == 3:
out, target_reweighted = mixup_process(out, target_reweighted, lam=lam) #3번 layer의 output을 섞어줌
out = self.layer4(out)
out = F.avg_pool2d(out, 4)
out = out.view(out.size(0), -1)
out = self.linear(out)
##########################################################################################################
# out.shape = torch.Size([100, 256, 8, 8])
# target_reweighted.shape = torch.Size([100, 10])
# out.shape(after) = torch.Size([100, 10])
if target is not None:
return out, target_reweighted
else:
return out
def forward(self,
x,
target=None,
mixup=False,
mixup_hidden=False,
mixup_alpha=None,
MyMixup=False,
MyMixup_num_layer=1):
#print x.shape
if self.per_img_std:
x = per_image_standardization(x)
if mixup_hidden:
layer_mix = random.randint(0, 2)
elif mixup:
layer_mix = 0
else:
layer_mix = None
out = x
if mixup_alpha is not None:
lam = get_lambda(mixup_alpha)
lam = torch.from_numpy(np.array([lam]).astype('float32')).cuda()
lam = Variable(lam)
if target is not None:
target_reweighted = to_one_hot(target, self.num_classes)
if MyMixup:
"""
out: image
target_reweigted = label
"""
Aug_out = transform(out)
out = self.conv1(out)
out = self.layer1(out)
if MyMixup_num_layer == 1:
middle_original_output = out
out = self.layer2(out)
if MyMixup_num_layer == 2:
middle_original_output = out
out = self.layer3(out)
if MyMixup_num_layer == 3:
middle_original_output = out
out = act(self.bn1(out))
out = F.avg_pool2d(out, 8)
out = out.view(out.size(0), -1)
out = self.linear(out)
with torch.no_grad():
Aug_out = self.conv1(Aug_out)
Aug_out = self.layer1(Aug_out)
if MyMixup_num_layer == 1:
middle_Aug_output = Aug_out
Aug_out = self.layer2(Aug_out)
if MyMixup_num_layer == 2:
middle_Aug_output = Aug_out
Aug_out = self.layer3(Aug_out)
if MyMixup_num_layer == 3:
middle_Aug_output = Aug_out
Aug_out = act(self.bn1(Aug_out))
Aug_out = F.avg_pool2d(Aug_out, 8)
Aug_out = Aug_out.view(Aug_out.size(0), -1)
Aug_out = self.linear(Aug_out)
return out, middle_original_output, middle_Aug_output, target_reweighted
else:
if layer_mix == 0:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.conv1(out)
out = self.layer1(out)
if layer_mix == 1:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.layer2(out)
if layer_mix == 2:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = self.layer3(out)
if layer_mix == 3:
out, target_reweighted = mixup_process(out,
target_reweighted,
lam=lam)
out = act(self.bn1(out))
out = F.avg_pool2d(out, 8)
out = out.view(out.size(0), -1)
out = self.linear(out)
if target is not None:
return out, target_reweighted
else:
return out