def generator(self, z):
with tf.variable_scope("generator"):
s_h, s_w = self.image_dims['height'], self.image_dims['width']
s_h2, s_w2 = conv_out_size_same(s_h, 2), conv_out_size_same(s_w, 2)
s_h4, s_w4 = conv_out_size_same(s_h2, 2), conv_out_size_same(s_w2, 2)
s_h8, s_w8 = conv_out_size_same(s_h4, 2), conv_out_size_same(s_w4, 2)
s_h16, s_w16 = conv_out_size_same(s_h8, 2), conv_out_size_same(s_w8, 2)
# 将随机向量z作为输入，通过全连接生成[s_h16, s_w16, G_F_DIM * 8]大小的输出
g_h0_lin = ops.linear(z, G_F_DIM * 8 * s_h16 * s_w16, 'g_h0_lin')
g_h0_re = tf.reshape(g_h0_lin, [-1, s_h16, s_w16, G_F_DIM * 8])
g_bn0 = tf.layers.batch_normalization(g_h0_re, name='g_bn0', training=self.is_train)
h0 = tf.nn.relu(g_bn0)
g_h1 = ops.deconv2d(h0, [BATCH_SIZE, s_h8, s_w8, G_F_DIM * 4], name='g_h1')
g_bn1 = tf.layers.batch_normalization(g_h1, name='g_bn1', training=self.is_train)
h1 = tf.nn.relu(g_bn1)
g_h2 = ops.deconv2d(h1, [BATCH_SIZE, s_h4, s_w4, G_F_DIM * 2], name='g_h2')
g_bn2 = tf.layers.batch_normalization(g_h2, name='g_bn2', training=self.is_train)
h2 = tf.nn.relu(g_bn2)
g_h3 = ops.deconv2d(h2, [BATCH_SIZE, s_h2, s_w2, G_F_DIM * 1], name='g_h3')
g_bn3 = tf.layers.batch_normalization(g_h3, name='g_bn3', training=self.is_train)
h3 = tf.nn.relu(g_bn3)
h4 = ops.deconv2d(h3, [BATCH_SIZE, s_h, s_w, self.image_dims['channel']], name='g_h4')
# 激活函数使用tanh
return tf.nn.tanh(h4)
def discriminator(self, image, reuse=False):
with tf.variable_scope("discriminator") as scope:
if reuse:
scope.reuse_variables()
# 用5*5，stride为2的filter对输入进行卷积操作
d_h0_conv = ops.conv2d(image, D_F_DIM, name='d_h0_conv')
# 激活函数为leaky relu
h0 = ops.lrelu(d_h0_conv)
d_h1_conv = ops.conv2d(h0, D_F_DIM * 2, name='d_h1_conv')
d_bn1 = tf.layers.batch_normalization(d_h1_conv, name='d_bn1', training=self.is_train)
h1 = ops.lrelu(d_bn1)
d_h2_conv = ops.conv2d(h1, D_F_DIM * 4, name='d_h2_conv')
d_bn2 = tf.layers.batch_normalization(d_h2_conv, name='d_bn2', training=self.is_train)
h2 = ops.lrelu(d_bn2)
d_h3_conv = ops.conv2d(h2, D_F_DIM * 8, name='d_h3_conv')
d_bn3 = tf.layers.batch_normalization(d_h3_conv, name='d_bn3', training=self.is_train)
h3 = ops.lrelu(d_bn3)
# 最后一层使用全连接
h4 = ops.linear(tf.reshape(h3, [BATCH_SIZE, -1]), 1, 'd_h4_lin')
return tf.nn.sigmoid(h4), h4