class CryptoNet_fivelayer(Module):
def __init__(self, in_dim, n_class):
super(CryptoNet_fivelayer, self).__init__()
self.conv = Conv_sec(in_dim,
5,
5,
5,
zero_padding=1,
stride=2,
method='SAME',
bit_length=bit_length)
self.sq1 = Activators_sec.Square(bit_length=bit_length)
self.fc1 = FullyConnect_sec(845, 100, bit_length=bit_length)
self.sq2 = Activators_sec.Square(bit_length=bit_length)
self.fc2 = FullyConnect_sec(100, n_class, bit_length=bit_length)
# self.logsoftmax = Logsoftmax()
self.logsoftmax = Logsoftmax_sec(bit_length=bit_length)
def forward(self, x1, x2):
in_size = x1.shape[0]
# start_time = time.time()
out_c1, out_c2 = self.conv.forward(x1, x2)
out_11, out_12 = self.sq1.forward(out_c1, out_c2)
self.conv_out_shape = out_11.shape
# print('out1shape: ',self.conv_out_shape)
out_11 = out_11.reshape(in_size, -1) # 将输出拉成一行
out_12 = out_12.reshape(in_size, -1) # 将输出拉成一行
out_fc11, out_fc12 = self.fc1.forward(out_11, out_12)
out_21, out_22 = self.sq2.forward(out_fc11, out_fc12)
out_31, out_32 = self.fc2.forward(out_21, out_22)
# end_time = time.time()
# print('time consume: ', (end_time-start_time)*1000)
out_logsoftmax_1, out_logsoftmax_2 = self.logsoftmax.forward(
out_31, out_32)
return out_logsoftmax_1, out_logsoftmax_2
# out_logsoftmax = self.logsoftmax.forward(out_31+out_32)
# return out_logsoftmax
def backward(self, dy1, dy2):
dy_logsoftmax_1, dy_logsoftmax_2 = self.logsoftmax.gradient(dy1, dy2)
dy_f31, dy_f32 = self.fc2.gradient(dy_logsoftmax_1, dy_logsoftmax_2)
dy_sq21, dy_sq22 = self.sq2.gradient(dy_f31, dy_f32)
dy_f21, dy_f22 = self.fc1.gradient(dy_sq21, dy_sq22)
dy_f21 = dy_f21.reshape(self.conv_out_shape)
dy_f22 = dy_f22.reshape(self.conv_out_shape)
dy_sq11, dy_sq12 = self.sq1.gradient(dy_f21, dy_f22)
self.conv.gradient(dy_sq11, dy_sq12)
def __init__(self, in_dim, n_class):
super(Minionn_fivelayer, self).__init__()
self.conv = Conv_sec(in_dim,
5,
5,
5,
zero_padding=2,
stride=2,
method='SAME',
bit_length=bit_length)
self.relu1 = Activators_sec.ReLU(bit_length=bit_length)
self.fc1 = FullyConnect_sec(980, 100, bit_length=bit_length)
self.relu2 = Activators_sec.ReLU(bit_length=bit_length)
self.fc2 = FullyConnect_sec(100, n_class, bit_length=bit_length)
self.logsoftmax = Logsoftmax_sec(bit_length=bit_length)
def __init__(self, in_dim, n_class):
super(CryptoNet_fivelayer, self).__init__()
self.conv = Conv_sec(in_dim,
5,
5,
5,
zero_padding=1,
stride=2,
method='SAME',
bit_length=bit_length)
self.sq1 = Activators_sec.Square(bit_length=bit_length)
self.fc1 = FullyConnect_sec(845, 100, bit_length=bit_length)
self.sq2 = Activators_sec.Square(bit_length=bit_length)
self.fc2 = FullyConnect_sec(100, n_class, bit_length=bit_length)
# self.logsoftmax = Logsoftmax()
self.logsoftmax = Logsoftmax_sec(bit_length=bit_length)
def __init__(self):
super(Discriminator_sec, self).__init__()
# 输入1*28*28 MNIST
# 1*28*28 -> 64*16*16
self.conv1 = Conv_sec(nc,
ndf,
4,
4,
zero_padding=1,
stride=2,
method='SAME',
bias_required=False,
bit_length=bit_length)
self.lrelu1 = Activators_sec.LeakyReLU(0.2, bit_length=bit_length)
# 64*16*16 -> 128*8*8
self.conv2 = Conv_sec(ndf,
ndf * 2,
4,
4,
zero_padding=1,
stride=2,
method='SAME',
bias_required=False,
bit_length=bit_length)
self.bn1 = BatchNorm_sec(ndf * 2, bit_length=bit_length)
self.lrelu2 = Activators_sec.LeakyReLU(0.2, bit_length=bit_length)
# 128*8*8 -> 256*4*4
self.conv3 = Conv_sec(ndf * 2,
ndf * 4,
4,
4,
zero_padding=1,
stride=2,
method='SAME',
bias_required=False,
bit_length=bit_length)
self.bn2 = BatchNorm_sec(ndf * 4, bit_length=bit_length)
self.lrelu3 = Activators_sec.LeakyReLU(0.2, bit_length=bit_length)
# 256*4*4 -> 1*1
self.conv4 = Conv_sec(ndf * 4,
1,
4,
4,
zero_padding=0,
stride=1,
method='VALID',
bias_required=False,
bit_length=bit_length)
self.sigmoid = Activators_sec.Sigmoid_CE_sec(bit_length=bit_length)
class Minionn_fivelayer(Module):
def __init__(self, in_dim, n_class):
super(Minionn_fivelayer, self).__init__()
self.conv = Conv_sec(in_dim,
5,
5,
5,
zero_padding=2,
stride=2,
method='SAME',
bit_length=bit_length)
self.relu1 = Activators_sec.ReLU(bit_length=bit_length)
self.fc1 = FullyConnect_sec(980, 100, bit_length=bit_length)
self.relu2 = Activators_sec.ReLU(bit_length=bit_length)
self.fc2 = FullyConnect_sec(100, n_class, bit_length=bit_length)
self.logsoftmax = Logsoftmax_sec(bit_length=bit_length)
def forward(self, x1, x2):
in_size = x1.shape[0]
# start_time_sum = time.time()
# start_conv = time.time()
out_c1, out_c2 = self.conv.forward(x1, x2)
# end_conv = time.time()
out_11, out_12, dfc_time1 = self.relu1.forward(out_c1, out_c2)
# start_fc1 = time.time()
self.conv_out_shape = out_11.shape
# print('out1shape: ',self.conv_out_shape)
out_11 = out_11.reshape(in_size, -1) # 将输出拉成一行
out_12 = out_12.reshape(in_size, -1) # 将输出拉成一行
out_fc11, out_fc12 = self.fc1.forward(out_11, out_12)
# end_fc1 = time.time()
out_21, out_22, dfc_time2 = self.relu2.forward(out_fc11, out_fc12)
# start_fc2 = time.time()
out_31, out_32 = self.fc2.forward(out_21, out_22)
# end_fc2 = time.time()
# end_time_sum = time.time()
# print('time consume: ', (end_conv-start_conv)*1000+(end_fc1-start_fc1)*1000+(end_fc1-start_fc1)*1000+self.relu1.offline_time+self.relu1.online_time+self.relu2.offline_time+self.relu2.online_time)
# print('time consume sum: ', (end_time_sum-start_time_sum)*1000)
out_logsoftmax_1, out_logsoftmax_2 = self.logsoftmax.forward(
out_31, out_32)
return out_logsoftmax_1, out_logsoftmax_2, dfc_time1 + dfc_time2
# out_logsoftmax = self.logsoftmax.forward(out_31+out_32)
# return out_logsoftmax
def backward(self, dy1, dy2):
dy_logsoftmax_1, dy_logsoftmax_2 = self.logsoftmax.gradient(dy1, dy2)
dy_f31, dy_f32 = self.fc2.gradient(dy_logsoftmax_1, dy_logsoftmax_2)
dy_relu21, dy_relu22 = self.relu2.gradient(dy_f31, dy_f32)
dy_f21, dy_f22 = self.fc1.gradient(dy_relu21, dy_relu22)
dy_f21 = dy_f21.reshape(self.conv_out_shape)
dy_f22 = dy_f22.reshape(self.conv_out_shape)
dy_relu11, dy_relu12 = self.relu1.gradient(dy_f21, dy_f22)
self.conv.gradient(dy_relu11, dy_relu12)
def conv_test():
bit_length = 32
# (1,28,28)*(5,5,5)
# x_numpy = np.random.randn(1,1,28,28).astype(np.float32)
# w_numpy = np.random.randn(5,1,5,5).astype(np.float32)
# b_numpy = np.random.randn(5).astype(np.float32)
# # (1,28,28)*(5,5,5)
# x_numpy_1 = np.random.randn(1,1,28,28).astype(np.float32)
# x_numpy_2 = x_numpy-x_numpy_1
# w_numpy_1 = np.random.randn(5,1,5,5).astype(np.float32)
# w_numpy_2 = w_numpy-w_numpy_1
# b_numpy_1 = np.random.randn(5).astype(np.float32)
# b_numpy_2 = b_numpy-b_numpy_1
## (3,32,32)*(64,2,2)
# x_numpy = np.random.randn(1,3,32,32).astype(np.float32)
# w_numpy = np.random.randn(64,3,2,2).astype(np.float32)
# b_numpy = np.random.randn(64).astype(np.float32)
# x = torch.tensor(x_numpy, requires_grad=True)
# x_numpy_1 = np.random.randn(1,3,32,32).astype(np.float32)
# x_numpy_2 = x_numpy-x_numpy_1
# w_numpy_1 = np.random.randn(64,3,2,2).astype(np.float32)
# w_numpy_2 = w_numpy-w_numpy_1
# b_numpy_1 = np.random.randn(64).astype(np.float32)
# b_numpy_2 = b_numpy-b_numpy_1
x_numpy = np.random.randn(1,32,32,32).astype(np.float32)
w_numpy = np.random.randn(128,32,3,3).astype(np.float32)
b_numpy = np.random.randn(128).astype(np.float32)
x = torch.tensor(x_numpy, requires_grad=True)
x_numpy_1 = np.random.randn(1,32,32,32).astype(np.float32)
x_numpy_2 = x_numpy-x_numpy_1
w_numpy_1 = np.random.randn(128,32,3,3).astype(np.float32)
w_numpy_2 = w_numpy-w_numpy_1
b_numpy_1 = np.random.randn(128).astype(np.float32)
b_numpy_2 = b_numpy-b_numpy_1
print('input_shape: ', x_numpy.shape)
print('w_shape: ', w_numpy.shape)
# padding=0, stride=2
# cl1 = Conv_sec(1, 5, 5, 5, zero_padding=0, stride=2, method='SAME')
# cl1 = Conv_sec(3, 64, 2, 2, zero_padding=0, stride=2, method='SAME')
cl1 = Conv_sec(32, 128, 3, 3, zero_padding=0, stride=2, method='SAME')
cl_ori = ConvLayer(1, 5, 5, 5, zero_padding=1, stride=2, method='SAME')
cl_tensor = torch.nn.Conv2d(1, 5, kernel_size=5, stride=2, padding=1)
## 设置参数
cl_ori.set_weight(Parameter(w_numpy, requires_grad=True))
cl_ori.set_bias(Parameter(b_numpy, requires_grad=True))
cl1.set_weight_1(Parameter(w_numpy_1, requires_grad=True))
cl1.set_bias_1(Parameter(b_numpy_1, requires_grad=True))
cl1.set_weight_2(Parameter(w_numpy_2, requires_grad=True))
cl1.set_bias_2(Parameter(b_numpy_2, requires_grad=True))
# print('param_error: \n', w_numpy-(w_numpy_1+w_numpy_2))
# print('param_error: \n', cl_ori.weights.data-(cl1.weights_1.data+cl1.weights_2.data))
'''前向传播'''
# start_time_tensor = time.time()
# conv_out = cl_tensor(x)
# end_time_tensor = time.time()
# start_time = time.time()
# conv_out = cl_ori.forward(x_numpy)
# end_time = time.time()
test_num = 10
time_avg = 0
for i in range(test_num):
start_time_sec = time.time()
conv_out_1, conv_out_2 = cl1.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)
class Discriminator_sec(Module):
def __init__(self):
super(Discriminator_sec, self).__init__()
# 输入1*28*28 MNIST
# 1*28*28 -> 64*16*16
self.conv1 = Conv_sec(nc,
ndf,
4,
4,
zero_padding=1,
stride=2,
method='SAME',
bias_required=False,
bit_length=bit_length)
self.lrelu1 = Activators_sec.LeakyReLU(0.2, bit_length=bit_length)
# 64*16*16 -> 128*8*8
self.conv2 = Conv_sec(ndf,
ndf * 2,
4,
4,
zero_padding=1,
stride=2,
method='SAME',
bias_required=False,
bit_length=bit_length)
self.bn1 = BatchNorm_sec(ndf * 2, bit_length=bit_length)
self.lrelu2 = Activators_sec.LeakyReLU(0.2, bit_length=bit_length)
# 128*8*8 -> 256*4*4
self.conv3 = Conv_sec(ndf * 2,
ndf * 4,
4,
4,
zero_padding=1,
stride=2,
method='SAME',
bias_required=False,
bit_length=bit_length)
self.bn2 = BatchNorm_sec(ndf * 4, bit_length=bit_length)
self.lrelu3 = Activators_sec.LeakyReLU(0.2, bit_length=bit_length)
# 256*4*4 -> 1*1
self.conv4 = Conv_sec(ndf * 4,
1,
4,
4,
zero_padding=0,
stride=1,
method='VALID',
bias_required=False,
bit_length=bit_length)
self.sigmoid = Activators_sec.Sigmoid_CE_sec(bit_length=bit_length)
def forward(self, x_input_1, x_input_2):
# 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)
# output_sigmoid = self.sigmoid.forward(l4)
# return output_sigmoid
start_l1 = time.time()
conv11, conv12 = self.conv1.forward(x_input_1, x_input_2)
# end_conv1 = time.time()
lrelu11, lrelu12, dfc_time_1 = self.lrelu1.forward(
conv11, conv12) # 还是relu最消耗时间= =
# end_l1 = time.time()
# print('conv1 (ms): ', (end_conv1-start_l1)*1000)
# print('L_1 (ms): ', (end_l1-start_l1)*1000)
conv21, conv22 = self.conv2.forward(lrelu11, lrelu12)
bn11, bn12 = self.bn1.forward(conv21, conv22)
lrelu21, lrelu22, dfc_time_2 = self.lrelu2.forward(bn11, bn12)
conv31, conv32 = self.conv3.forward(lrelu21, lrelu22)
bn21, bn22 = self.bn2.forward(conv31, conv32)
lrelu31, lrelu32, dfc_time_3 = self.lrelu3.forward(bn21, bn22)
conv41, conv42 = self.conv4.forward(lrelu31, lrelu32)
# print('conv4: ', (conv41+conv42)[0][0])
output_sig_1, output_sig_2 = self.sigmoid.forward(conv41, conv42)
end_sig = time.time()
print('sigmoid input shape: ', conv41.shape)
print('dfc_time (s): ', (dfc_time_1 + dfc_time_2 + dfc_time_3) * 60)
print('total SD time (s): ', (end_sig - start_l1) -
(dfc_time_1 + dfc_time_2 + dfc_time_3) * 60)
print('total SD time real (s): ', (end_sig - start_l1))
return output_sig_1, output_sig_2
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