def __init__(self):
super(Discriminator, self).__init__()
# 输入1*28*28 MNIST
# 1*28*28 -> 64*16*16
self.conv1 = ConvLayer(nc,
ndf,
4,
4,
zero_padding=1,
stride=2,
method='SAME',
bias_required=False)
self.lrelu1 = Activators.LeakyReLU(0.2)
# 64*16*16 -> 128*8*8
self.conv2 = ConvLayer(ndf,
ndf * 2,
4,
4,
zero_padding=1,
stride=2,
method='SAME',
bias_required=False)
self.bn1 = BatchNorm(ndf * 2)
self.lrelu2 = Activators.LeakyReLU(0.2)
# 128*8*8 -> 256*4*4
self.conv3 = ConvLayer(ndf * 2,
ndf * 4,
4,
4,
zero_padding=1,
stride=2,
method='SAME',
bias_required=False)
self.bn2 = BatchNorm(ndf * 4)
self.lrelu3 = Activators.LeakyReLU(0.2)
# 256*4*4 -> 1*1
self.conv4 = ConvLayer(ndf * 4,
1,
4,
4,
zero_padding=0,
stride=1,
method='VALID',
bias_required=False)
self.sigmoid = Activators.Sigmoid_CE()
class Discriminator(Module):
def __init__(self):
super(Discriminator, self).__init__()
# 输入1*28*28 MNIST
# 1*28*28 -> 64*16*16
self.conv1 = ConvLayer(nc, ndf, 4,4, zero_padding=1, stride=2,method='SAME', bias_required=False)
self.lrelu1 = Activators.LeakyReLU(0.2)
# 64*16*16 -> 128*8*8
self.conv2 = ConvLayer(ndf, ndf*2, 4,4, zero_padding=1, stride=2, method='SAME', bias_required=False)
self.bn1 = BatchNorm(ndf*2)
self.lrelu2 = Activators.LeakyReLU(0.2)
# 128*8*8 -> 256*4*4
self.conv3 = ConvLayer(ndf*2, ndf*4, 4,4, zero_padding=1, stride=2, method='SAME', bias_required=False)
self.bn2 = BatchNorm(ndf*4)
self.lrelu3 = Activators.LeakyReLU(0.2)
# 256*4*4 -> 1*1
self.conv4 = ConvLayer(ndf*4, 1, 4,4, zero_padding=0, stride=1, method='VALID', bias_required=False)
self.sigmoid = Activators.Sigmoid_CE()
def forward(self, x_input):
l1 = self.lrelu1.forward(self.conv1.forward(x_input))
l2 = self.lrelu2.forward(self.bn1.forward(self.conv2.forward(l1)))
l3 = self.lrelu3.forward(self.bn2.forward(self.conv3.forward(l2)))
l4 = self.conv4.forward(l3)
# print('D l1 shape: ',l1.shape)
# print('D l2 shape: ',l2.shape)
# print('D l3 shape: ',l3.shape)
# print('D l4 shape: ',l4.shape)
output_sigmoid = self.sigmoid.forward(l4)
return output_sigmoid
def backward(self, dy):
# print('dy.shape: ', dy.shape)
dy_sigmoid = self.sigmoid.gradient(dy)
# print('dy_sigmoid.shape: ', dy_sigmoid.shape)
dy_l4 = self.conv4.gradient(dy_sigmoid)
dy_l3 = self.conv3.gradient(self.bn2.gradient(self.lrelu3.gradient(dy_l4)))
dy_l2 = self.conv2.gradient(self.bn1.gradient(self.lrelu2.gradient(dy_l3)))
dy_l1 = self.conv1.gradient(self.lrelu1.gradient(dy_l2))
# print('D_backward output shape: ',dy_l1.shape)
return dy_l1
def __init__(self):
super(Generator, self).__init__()
# 构建反向传播网络组建
# 输入Z=[100,]
# 100*1 -> 256*4*4
self.deconv1 = Deconv(nz, ngf*4, 4, zero_padding=0, stride=1, method='VALID', bias_required=False)
self.bn1 = BatchNorm(ngf*4)
self.relu1 = Activators.ReLU()
# 256*4*4 -> 128*8*8
self.deconv2 = Deconv(ngf*4, ngf*2, 4, zero_padding=1, stride=2, method='SAME', bias_required=False)
self.bn2 = BatchNorm(ngf*2)
self.relu2 = Activators.ReLU()
# 128*8*8 -> 64*16*16
self.deconv3 = Deconv(ngf*2, ngf, 4, zero_padding=1, stride=2, method='SAME', bias_required=False)
self.bn3 = BatchNorm(ngf)
self.relu3 = Activators.ReLU()
# 64*16*16 -> 1*32*32
self.deconv4 = Deconv(ngf, nc, 4, zero_padding=1, stride=2, method='SAME', bias_required=False)
self.tanh = Activators.Tanh()
class Generator(Module):
def __init__(self):
super(Generator, self).__init__()
# 构建反向传播网络组建
# 输入Z=[100,]
# 100*1 -> 256*4*4
self.deconv1 = Deconv(nz, ngf*4, 4, zero_padding=0, stride=1, method='VALID', bias_required=False)
self.bn1 = BatchNorm(ngf*4)
self.relu1 = Activators.ReLU()
# 256*4*4 -> 128*8*8
self.deconv2 = Deconv(ngf*4, ngf*2, 4, zero_padding=1, stride=2, method='SAME', bias_required=False)
self.bn2 = BatchNorm(ngf*2)
self.relu2 = Activators.ReLU()
# 128*8*8 -> 64*16*16
self.deconv3 = Deconv(ngf*2, ngf, 4, zero_padding=1, stride=2, method='SAME', bias_required=False)
self.bn3 = BatchNorm(ngf)
self.relu3 = Activators.ReLU()
# 64*16*16 -> 1*32*32
self.deconv4 = Deconv(ngf, nc, 4, zero_padding=1, stride=2, method='SAME', bias_required=False)
self.tanh = Activators.Tanh()
def forward(self, x_input):
# print('G input shape: ',x_input.shape)
l1 = self.relu1.forward(self.bn1.forward(self.deconv1.forward(x_input)))
l2 = self.relu2.forward(self.bn2.forward(self.deconv2.forward(l1)))
l3 = self.relu3.forward(self.bn3.forward(self.deconv3.forward(l2)))
l4 = self.deconv4.forward(l3)
# print('G l1 shape: ',l1.shape)
# print('G l2 shape: ',l2.shape)
# print('G l3 shape: ',l3.shape)
# print('G l4 shape: ',l4.shape)
output_tanh = self.tanh.forward(l4)
return output_tanh
def backward(self, dy):
dy_tanh = self.tanh.gradient(dy)
dy_l4 = self.deconv4.gradient(dy_tanh)
dy_l3 = self.deconv3.gradient(self.bn3.gradient(self.relu3.gradient(dy_l4)))
dy_l2 = self.deconv2.gradient(self.bn2.gradient(self.relu2.gradient(dy_l3)))
self.deconv1.gradient(self.bn1.gradient(self.relu1.gradient(dy_l2)))
def bn_test():
bit_length = 64
width = 8
height = 8
channel = 256
x_numpy = np.random.randn(1,channel,height,width).astype(np.float64)
x_numpy_1 = np.random.randn(1,channel,height,width).astype(np.float64)
x_numpy_2 = x_numpy - x_numpy_1
# print('input: ',x_numpy)
w_numpy = np.random.normal(1.0, 0.02, size=(channel)).astype(np.float64)
b_numpy = np.zeros(channel).astype(np.float64)
bn = BatchNorm(channel)
bn_sec = BatchNorm_sec(channel)
# 设置参数
bn.set_gamma(Parameter(w_numpy, requires_grad=True))
bn.set_beta(Parameter(b_numpy, requires_grad=True))
bn_sec.set_gamma(Parameter(w_numpy, requires_grad=True))
bn_sec.set_beta(Parameter(b_numpy, requires_grad=True))
bn_out = bn.forward(x_numpy)
bn_out_sec_1, bn_out_sec_2 = bn_sec.forward(x_numpy_1, x_numpy_2)
# print('error sum: ',bn_out-(bn_out_sec_1+bn_out_sec_2))
test_num = 10
time_avg = 0
for i in range(test_num):
start_time_sec = time.time()
bn_out_sec_1, bn_out_sec_2 = bn_sec.forward(x_numpy_1, x_numpy_2)
end_time_sec = time.time()
time_avg+=(end_time_sec-start_time_sec)*1000
print('time avg sec: \n', time_avg/test_num)