def forward(self, x, task_idx: int = -1):
x = F.pad(x, pad=(3, 3, 3, 3), mode='constant',
value=0) # other modes are reflect, replicate, constant
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
# x = F.pad(x, (0,1,0,1), 'constant', 0)
self.x_prelayer1 = self.maxpool(x) # (64,64,64)
if self.bsp:
self.x_layer1 = forward_sequential(self.x_prelayer1, self.layer1,
task_idx)
self.x_layer2 = forward_sequential(self.x_layer1, self.layer2,
task_idx)
self.x_layer3 = forward_sequential(self.x_layer2, self.layer3,
task_idx)
self.x_layer4 = forward_sequential(self.x_layer3, self.layer4,
task_idx)
x = forward_sequential(self.x_layer4, self.compress1, task_idx)
else:
self.x_layer1 = self.layer1(self.x_prelayer1) # (256,32,32)
self.x_layer2 = self.layer2(self.x_layer1) # (512,16,16)
self.x_layer3 = self.layer3(self.x_layer2) # (1024,16,16)
self.x_layer4 = self.layer4(self.x_layer3) # (2048,16,16)
x = self.compress1(self.x_layer4) # (8,16,16)
if self.final_act:
x = self.compress_bn(x)
x = self.relu1(x)
if self.normalize_outputs:
x = self.groupnorm(x)
return x
def forward(self, x, time: int = -1):
out = F.relu(self.bn1(self.conv1(x)))
if self.bsp: # bsp mode
out = forward_sequential(out, self.layer1, time)
out = forward_sequential(out, self.layer2, time)
out = forward_sequential(out, self.layer3, time)
else:
out = self.layer1(out)
out = self.layer2(out)
if self.final_act:
out = self.layer3(out)
else:
for i in range(6):
out = self.layer3[i](out)
basic_block = self.layer3[6]
block_input = out
out = F.relu(basic_block.bn1(basic_block.conv1(out)))
# out = basic_block.bn2(basic_block.conv2(out))
out = basic_block.conv2(out)
out += basic_block.shortcut(block_input)
# out = F.relu(out)
out = F.avg_pool2d(out, out.size()[3])
out = out.view(out.size(0), -1)
return out
def forward(self, x, task_idx: int = -1):
out = F.relu(self.bn1(self.conv1(x)))
if self.bsp: # bsp mode
out = forward_sequential(out, self.layer1, task_idx)
out = forward_sequential(out, self.layer2, task_idx)
out = forward_sequential(out, self.layer3, task_idx)
else:
out = self.layer1(out)
out = self.layer2(out)
out = self.layer3(out)
out = F.avg_pool2d(out, out.size()[3])
out = out.view(out.size(0), -1)
out = self.linear(out)
return out
def forward(self, x, task_idx:int=-1): if self.bsp: x = forward_sequential(x, self.conv1, task_idx) x = forward_sequential(x, self.conv2, task_idx) x = forward_sequential(x, self.conv3, task_idx) x = forward_sequential(x, self.conv5, task_idx) else: x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) if self.final_act: x = self.conv5(x) else: x = self.conv5[0](x) if self.use_residual: res = self.residual(x) x = x + res return x
def forward(self, x, task_idx:int=-1, cache={}):
x0 = x
if self.bsp:
x = forward_sequential(x, self.conv1, task_idx)
x = forward_sequential(x, self.conv2, task_idx)
else:
x = self.conv1(x)
x = self.conv2(x)
if self.use_residual:
x = x + self.residual(x0)
x2 = x
if self.bsp:
x = forward_sequential(x, self.conv3, task_idx)
x = forward_sequential(x, self.conv4, task_idx)
x = forward_sequential(x, self.conv5, task_idx)
x = x + forward_sequential(x2, self.skip, task_idx)
else:
x = self.conv3(x)
x = self.conv4(x)
if self.final_act:
x = self.conv5(x)
else:
x = self.conv5[0](x)
x = x + self.skip(x2)
# x = self.conv5(x)
# if self.final_act:
# x = x + self.skip(x2)
# else:
# x = x + self.skip[0](x2)
if self.normalize_outputs:
x = self.groupnorm(x)
return x
def forward(self, x, time):
residual = x
out = self.conv1(x, time)
out = self.bn1(out)
out = self.relu(out)
# out = F.pad(out, pad=(1,1,1,1), mode='constant', value=0) # other modes are reflect, replicate
out = self.conv2(out, time)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out, time)
out = self.bn3(out)
if self.downsample is not None:
residual = forward_sequential(x, self.downsample, time)
out += residual
out = self.relu(out)
return out
def forward(self, x, time):
out = F.relu(self.bn1(self.conv1(x, time)))
out = self.bn2(self.conv2(out, time))
out += forward_sequential(x, self.shortcut, time)
out = F.relu(out)
return out