def forward(self,
x,
sample=None,
label=None,
device='cpu',
return_rate=False):
rate = []
temp = []
Channel = []
bsz = x.shape[0]
temp.append(sample)
if return_rate:
for layer in self.layer:
x = layer(x)
temp.append(x.detach())
rate.append(RD_fn(T=x, X=sample, Label=label, device=device))
x = self.avgpool(x)
x = self.classifier(x.view(bsz, -1))
for i in range(len(temp) - 1):
Channel.append(mi(temp[i], temp[i + 1], device, 'cov'))
return x, torch.stack(rate), torch.stack(Channel)
else:
for layer in self.layer:
x = layer(x)
x = self.classifier(self.avgpool(x).view(bsz, -1))
return x
def forward(self,
x,
sample=None,
label=None,
device='cpu',
return_rate=False):
# See note [TorchScript super()]
rate = []
temp = []
Channel = []
temp.append(sample)
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
if return_rate:
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.layer1(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.layer2(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.layer3(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.layer4(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.avgpool(x)
x = torch.flatten(x, 1)
x = self.fc(x)
for i in range(len(temp) - 1):
Channel.append(mi(temp[i], temp[i + 1], device, 'cov'))
return x, torch.stack(rate), torch.stack(Channel)
else:
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avgpool(x)
x = torch.flatten(x, 1)
x = self.fc(x)
return x
def forward(self,
x,
sample=None,
label=None,
device='cpu',
return_rate=False):
rate = []
temp = []
Channel = []
temp.append(sample)
if return_rate:
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = F.max_pool2d(F.relu(self.conv2(x)), (2, 2))
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = x.view(-1, self.num_flat_features(x))
x = F.relu(self.fc1(x))
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = F.relu(self.fc2(x))
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.fc3(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
for i in range(len(temp) - 1):
Channel.append(mi(temp[i], temp[i + 1], device, 'cov'))
return x, torch.stack(rate), torch.stack(Channel)
else:
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
x = F.max_pool2d(F.relu(self.conv2(x)), (2, 2))
x = x.view(-1, self.num_flat_features(x))
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def forward(self, x, sample=None, label=None, device='cpu', return_rate=False):
rate = []
temp = []
Channel = []
temp.append(sample)
if return_rate:
x = self.layer1(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.layer2(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.layer3(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.layer4(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.layer5(x)
temp.append(x.detach())
rate.append(RD_fn(X=sample, T=x, Label=label, device=device))
x = self.avgpool(x)
x = torch.flatten(x, 1)
x = self.classifier(x)
for i in range(len(temp)-1):
Channel.append(mi(temp[i], temp[i+1],device,'cov'))
return x, torch.stack(rate), torch.stack(Channel)
else:
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.layer5(x)
x = self.avgpool(x)
x = torch.flatten(x, 1)
x = self.classifier(x)
return x
def forward(self,
x,
sample=None,
label=None,
device='cpu',
return_rate=False):
rate = []
m = x.shape[0]
x = x.view(-1, 784) # Flattern the (n,3,32,32) to (n,3096)
temp = []
Channel = []
temp.append(sample)
if return_rate:
x = self.gate(self.l1(x))
temp.append(x.detach())
rate.append(RD_fn(T=x, X=sample, Label=label, device=device))
x = self.gate(self.l2(x))
temp.append(x.detach())
rate.append(RD_fn(T=x, X=sample, Label=label, device=device))
x = self.gate(self.l3(x))
temp.append(x.detach())
rate.append(RD_fn(T=x, X=sample, Label=label, device=device))
x = self.gate(self.l4(x))
temp.append(x.detach())
rate.append(RD_fn(T=x, X=sample, Label=label, device=device))
x = self.l5(x)
temp.append(x.detach())
rate.append(RD_fn(T=x, X=sample, Label=label, device=device))
for i in range(len(temp) - 1):
Channel.append(mi(temp[i], temp[i + 1], device, 'cov'))
return x, torch.stack(rate), torch.stack(Channel)
else:
x = self.gate(self.l1(x))
x = self.gate(self.l2(x))
x = self.gate(self.l3(x))
x = self.gate(self.l4(x))
return self.l5(x)
def mi_from_experiments(experiments):
cols = transpose([map(lambda x:x[1],concat(row)) for row in experiments])
plt.imshow([[mi(col1,col2,correct=False) for col1 in cols] for col2 in (cols)],
interpolation='none')
plt.colorbar()
plt.show()