def forward(self, input, label):
network = self.conv1(input)
network = LeakyReLU(negative_slope=0.2)(network)
network = self.bn1(network)
network = self.conv2(network)
network = LeakyReLU(negative_slope=0.2)(network)
network = self.bn2(network)
network = self.conv3(network)
network = LeakyReLU(negative_slope=0.2)(network)
network = self.bn3(network)
network = self.conv4(network)
network = LeakyReLU(negative_slope=0.2)(network)
network = self.bn4(network)
network = network.view(-1, 512)
network = torch.cat([network, label], 1)
network = self.d_fc(network)
network = LeakyReLU(negative_slope=0.2)(network)
network = self.merge_layer(network)
network = Sigmoid()(network)
return network
def decode(self, z):
#implement
x = LeakyReLU(0.1)(self.fc1(z))
x = x.view(-1,1,2,2)
x = LeakyReLU(0.1)(self.convt1(x))
x = LeakyReLU(0.1)(self.convt2(x))
x = LeakyReLU(0.1)(self.convt3(x))
x = LeakyReLU(0.1)(self.convt4(x))
return x
def forward(self, x):
ALPHA = self.CFG["ALPHA"]
output = self.layer1(x)
output = LeakyReLU(negative_slope=ALPHA)(self.layer2(output))
output = self.layer3(output)
output = LeakyReLU(negative_slope=ALPHA)(self.layer4(output))
output = self.layer5(output)
output = LeakyReLU(negative_slope=ALPHA)(self.layer6(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer7(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer8(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer9(output))
output = self.layer10(output)
output = LeakyReLU(negative_slope=ALPHA)(self.layer11(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer12(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer13(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer14(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer15(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer16(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer17(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer18(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer19(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer20(output))
output = self.layer21(output)
output = LeakyReLU(negative_slope=ALPHA)(self.layer22(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer23(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer24(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer25(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer26(output))
output = self.layer27(output)
output = LeakyReLU(negative_slope=ALPHA)(self.layer28(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer29(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer30(output))
output = output.view(output.size(0), -1)
output = LeakyReLU(negative_slope=ALPHA)(self.layer33(output))
output = LeakyReLU(negative_slope=ALPHA)(self.layer34(output))
output = self.layer35(output)
output = LeakyReLU(negative_slope=ALPHA)(self.layer36(output))
output = Sigmoid()(output)
output = output.view(
-1, self.CFG["CELL_NUMS"], self.CFG["CELL_NUMS"],
5 * self.CFG["BOXES_EACH_CELL"] + self.CFG["CLASS_NUMS"])
return output
def q_z(self, x):
#(1,28,28) -> (8,26,26) -> (8,13,13)
x = F.max_pool2d(LeakyReLU(0.1)(self.conv1(x)), (2, 2))
#(8,13,13) -> (16,12,12) -> (16,6,6)
x = F.max_pool2d(LeakyReLU(0.1)(self.conv2(x)), (2, 2))
#(16,6,6) -> (32,4,4)
x = LeakyReLU(0.1)(self.conv3(x))
#(32,4,4) -> (64,2,2)
x = LeakyReLU(0.1)(self.conv4(x))
x = x.view(-1, 64 * self.finalConvLength * self.finalConvLength)
z_q_mean = self.mean(x)
z_q_logvar = self.logvar(x)
return z_q_mean, z_q_logvar
def forward(self, input, label):
network_1 = self.input_conv1(input)
network_1 = LeakyReLU(negative_slope=0.2, inplace=True)(network_1)
network_1 = self.input_conv2(network_1)
network_1 = LeakyReLU(negative_slope=0.2, inplace=True)(network_1)
network_1 = self.input_conv3(network_1)
network_1 = self.input_bn3(network_1)
network_1 = LeakyReLU(negative_slope=0.2, inplace=True)(network_1)
network_1 = self.input_conv4(network_1)
network_1 = self.input_bn4(network_1)
network_1 = LeakyReLU(negative_slope=0.2, inplace=True)(network_1)
network_1 = self.input_conv5(network_1)
network_1 = network_1.view(-1, self.features_num * 8)
network = torch.cat([network_1, label], 1)
network = self.cat_conv1(network)
network = LeakyReLU(negative_slope=0.2, inplace=True)(network)
network = self.cat_merge(network)
network = Sigmoid()(network)
return network
def forward(self, x):
#(1,28,28) -> (8,26,26) -> (8,13,13)
x = F.max_pool2d(LeakyReLU(0.1)(self.conv1(x)), (2,2))
#(8,13,13) -> (16,12,12) -> (16,6,6)
x = F.max_pool2d(LeakyReLU(0.1)(self.conv2(x)), (2,2))
#(16,6,6) -> (32,4,4)
x = LeakyReLU(0.1)(self.conv3(x))
#(32,4,4) -> (64,2,2)
x = LeakyReLU(0.1)(self.conv4(x))
#(64,2,2) -> lsdim mean and logvar
mu, logvar = self.encode(x.view(-1, 64*2*2))
#get code
z = self.reparameterize(mu, logvar)
#decode code
return self.decode(z), mu, logvar, z
