def _shadowdata_discriminator_model(generated_data,
generator_input,
is_training=True):
with slim.arg_scope(
[slim.fully_connected, slim.separable_conv2d, slim.convolution1d],
weights_initializer=initializers.variance_scaling(scale=2.0),
weights_regularizer=slim.l2_regularizer(0.001),
# normalizer_fn=slim.batch_norm,
# normalizer_params={'is_training': is_training, 'decay': 0.999},
# normalizer_fn=slim.instance_norm,
# normalizer_params={'center': True, 'scale': True, 'epsilon': 0.001},
activation_fn=(lambda inp: slim.nn.leaky_relu(inp, alpha=0.1))):
band_size = generated_data.get_shape()[3].value
net = generated_data
net = tf.squeeze(net, axis=[1, 2])
net = tf.expand_dims(net, axis=2)
net1 = slim.convolution1d(net, band_size, band_size, padding='VALID')
net2 = slim.convolution1d(transpose(net1, perm=[0, 2, 1]),
band_size,
band_size,
padding='VALID',
normalizer_fn=None,
normalizer_params=None,
activation_fn=None)
return tf.expand_dims(tf.expand_dims(slim.flatten(net2), axis=1), axis=1)
def _shadowdata_generator_model_simple(netinput, is_training=True):
with slim.arg_scope(
[slim.conv2d, slim.conv2d_transpose, slim.convolution1d],
trainable=is_training,
data_format="NHWC"):
band_size = netinput.get_shape()[3].value
net = tf.expand_dims(tf.squeeze(netinput, axis=[1, 2]), axis=2)
net = slim.convolution1d(net,
1,
band_size,
padding='SAME',
normalizer_fn=None,
normalizer_params=None,
weights_regularizer=None,
activation_fn=None)
return tf.expand_dims(tf.expand_dims(slim.flatten(net), axis=1), axis=1)
def _shadowdata_feature_discriminator_model(generated_data, is_training=True):
with slim.arg_scope(
[slim.fully_connected, slim.separable_conv2d, slim.convolution1d],
weights_initializer=initializers.variance_scaling(scale=2.0),
weights_regularizer=slim.l2_regularizer(0.001),
# normalizer_fn=slim.batch_norm,
# normalizer_params={'is_training': is_training, 'decay': 0.999},
# normalizer_fn=slim.instance_norm,
# normalizer_params={'center': True, 'scale': True, 'epsilon': 0.001},
activation_fn=(lambda inp: slim.nn.leaky_relu(inp, alpha=0.1))):
band_size = generated_data.get_shape()[3].value
net = generated_data
net = slim.flatten(net)
net1 = slim.fully_connected(net, band_size // 2)
net2 = slim.fully_connected(net1, band_size // 4)
net3 = slim.fully_connected(net2, band_size // 8)
return net3
def _build_graph_a(self, state_input, scope_name, train):
# 环境和智能体本地的共同观察
with tf.variable_scope(scope_name, reuse=tf.AUTO_REUSE):
with slim.arg_scope(
[slim.conv2d, slim.fully_connected],
trainable=train,
weights_initializer=tf.truncated_normal_initializer(
stddev=0.1),
weights_regularizer=slim.l2_regularizer(0.05)):
fc1 = slim.fully_connected(state_input,
100,
scope='full_connected1')
relu1 = tf.nn.relu(fc1)
fc2 = slim.fully_connected(relu1, 100, scope='full_connected2')
relu2 = tf.nn.relu(fc2)
fc3 = slim.fully_connected(relu2,
self.action_dim,
scope='full_connected3')
output = tf.nn.softmax(fc3)
return output
def _shadowdata_discriminator_model_simple(generated_data,
generator_input,
is_training=True):
with slim.arg_scope(
[slim.fully_connected, slim.separable_conv2d, slim.convolution1d],
weights_initializer=initializers.variance_scaling(scale=2.0),
activation_fn=(lambda inp: slim.nn.leaky_relu(inp, alpha=0.01))):
band_size = generated_data.get_shape()[3].value
net = tf.concat(axis=3, values=[generated_data, generator_input])
net = tf.squeeze(net, axis=[1, 2])
net = tf.expand_dims(net, axis=2)
size = band_size * 2
net = slim.convolution1d(net,
size,
size,
padding='VALID',
normalizer_fn=None,
normalizer_params=None,
activation_fn=None)
net = tf.expand_dims(tf.expand_dims(slim.flatten(net), axis=1), axis=1)
return net
def _build_graph_c(self, state_input, action, scope_name, train):
# 环境和智能体本地的共同观察
with tf.variable_scope(scope_name, reuse=tf.AUTO_REUSE):
with slim.arg_scope(
[slim.conv2d, slim.fully_connected],
trainable=train,
weights_initializer=tf.truncated_normal_initializer(
stddev=0.1),
weights_regularizer=slim.l2_regularizer(0.05)):
fc1_s = slim.fully_connected(state_input,
100,
scope='full_connected1_s')
fc1_a = slim.fully_connected(action,
100,
scope='full_connected1_a')
concat = tf.concat([fc1_s, fc1_a], axis=1)
fc1 = slim.fully_connected(concat,
100,
scope='full_connected1')
output = slim.fully_connected(fc1, 1, scope='full_connected2')
return output
def _shadowdata_generator_model(netinput, is_training=True):
with slim.arg_scope(
[slim.conv2d, slim.conv2d_transpose, slim.convolution1d],
# weights_initializer=initializers.variance_scaling(scale=2.0),
weights_initializer=initializers.zeros(),
# weights_regularizer=slim.l1_l2_regularizer(),
# normalizer_fn=slim.batch_norm,
# normalizer_params={'is_training': is_training, 'decay': 0.95},
# normalizer_fn=slim.instance_norm,
# normalizer_params={'center': True, 'scale': True, 'epsilon': 0.001},
activation_fn=(lambda inp: slim.nn.leaky_relu(inp, alpha=0.1)),
trainable=is_training,
data_format="NHWC"):
num_filters = 1
band_size = netinput.get_shape()[3].value
kernel_size = band_size
net0 = tf.expand_dims(tf.squeeze(netinput, axis=[1, 2]), axis=2)
net1 = slim.convolution1d(net0,
num_filters,
kernel_size,
padding='SAME')
net1 = net1 + net0
net2 = slim.convolution1d(net1,
num_filters,
kernel_size // 2,
padding='SAME')
net2 = net2 + net1 + net0
net3 = slim.convolution1d(net2,
num_filters,
kernel_size // 4,
padding='SAME')
net3 = net3 + net2 + net1
net4 = slim.convolution1d(net3,
num_filters,
kernel_size // 8,
padding='SAME')
net4 = net4 + net3 + net2
net5 = slim.convolution1d(net4,
num_filters,
kernel_size // 4,
padding='SAME')
net5 = net5 + net4 + net3
net6 = slim.convolution1d(net5,
num_filters,
kernel_size // 2,
padding='SAME')
net6 = net6 + net5 + net4
net7 = slim.convolution1d(net6,
num_filters,
kernel_size,
padding='SAME',
normalizer_fn=None,
normalizer_params=None,
weights_regularizer=None,
activation_fn=None)
flatten = slim.flatten(net7)
# net9 = slim.fully_connected(flatten, band_size, activation_fn=None)
return tf.expand_dims(tf.expand_dims(flatten, axis=1), axis=1)
def create_Q_network(self, name): # 创建Q网络(vgg16结构)
self.state_input = tf.placeholder("float",
shape=self.state_dim,
name='state_input')
with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
with slim.arg_scope(
[slim.conv2d],
activation_fn=tf.nn.relu,
padding='SAME',
weights_initializer=tf.truncated_normal_initializer(
self.mu, self.sigma),
weights_regularizer=slim.l2_regularizer(0.0005),
):
# 112 * 112 * 64
net = slim.conv2d(self.state_input,
64, [7, 7],
stride=2,
scope='conv1')
# 56 * 56 * 64
net = slim.max_pool2d(net, [3, 3], stride=2, scope='pool1')
temp = net
# 第一残差块
net = slim.conv2d(net, 64, [3, 3], scope='conv2_1_1')
net = slim.conv2d(net, 64, [3, 3], scope='conv2_1_2')
# 残差相加
net = tf.nn.relu(tf.add(temp, net))
temp = net
# 残差块
net = slim.conv2d(net, 64, [3, 3], scope='conv2_2_1')
net = slim.conv2d(net, 64, [3, 3], scope='conv2_2_2')
# 残差相加
net = tf.nn.relu(tf.add(temp, net))
temp = net
# 28 * 28 * 128
temp = slim.conv2d(temp, 128, [1, 1], stride=2, scope='r1')
# 第二残差块
net = slim.conv2d(net,
128, [3, 3],
stride=2,
scope='conv3_1_1')
net = slim.conv2d(net, 128, [3, 3], scope='conv3_1_2')
# 残差相加
net = tf.nn.relu(tf.add(temp, net))
temp = net
# 残差块
net = slim.conv2d(net, 128, [3, 3], scope='conv3_2_1')
net = slim.conv2d(net, 128, [3, 3], scope='conv3_2_2')
# 残差相加
net = tf.nn.relu(tf.add(temp, net))
temp = net
# 14 * 14 * 256
temp = slim.conv2d(temp, 256, [1, 1], stride=2, scope='r2')
# 第三残差块
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv4_1_1')
net = slim.conv2d(net, 256, [3, 3], scope='conv4_1_2')
# 残差相加
net = tf.nn.relu(tf.add(temp, net))
temp = net
# 残差块
net = slim.conv2d(net, 256, [3, 3], scope='conv4_2_1')
net = slim.conv2d(net, 256, [3, 3], scope='conv4_2_2')
# 残差相加
net = tf.nn.relu(tf.add(temp, net))
temp = net
# 7 * 7 * 512
temp = slim.conv2d(temp, 512, [1, 1], stride=2, scope='r3')
# 第四残差块
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv5_1_1')
net = slim.conv2d(net, 512, [3, 3], scope='conv5_1_2')
# 残差相加
net = tf.nn.relu(tf.add(temp, net))
temp = net
# 残差块
net = slim.conv2d(net, 512, [3, 3], scope='conv5_2_1')
net = slim.conv2d(net, 512, [3, 3], scope='conv5_2_2')
# 残差相加
net = tf.nn.relu(tf.add(temp, net))
net = slim.avg_pool2d(net, [4, 4], stride=1, scope='pool2')
net = slim.flatten(net, scope='flatten')
fc1 = slim.fully_connected(net, 1000, scope='fc1')
self.logits = slim.fully_connected(fc1,
self.action_dim +
self.parameterdim,
activation_fn=None,
scope='fc2')
self.Q_value = tf.nn.softmax(self.logits)