def discriminate(input_data, scope="discriminate", reuse=False, is_training=True):
batch_norm_params = {
"is_training": is_training,
"decay": 0.9997,
"epsilon": 0.001,
"updates_collections": tf.GraphKeys.UPDATE_OPS,
"variables_collections": {
"beta": None,
"gamma": None,
"moving_mean": ['moving_vars'],
'moving_variance': ['moving_vars'],
}
}
with tf.variable_scope(scope, 'discriminate', [input_data]):
with slim.arg_scope([slim.conv2d],
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
weights_initializer=tf.truncated_normal_initializer(0.0, 0.1),
weights_regularizer=slim.l1_l2_regularizer()):
conv1 = slim.conv2d(input_data, 32, [3, 3], padding="SAME", scope="d_conv1", reuse=reuse)
conv2 = slim.conv2d(conv1, 64, [5, 5], padding="SAME", scope="d_conv2", reuse=reuse)
conv3 = slim.conv2d(conv2, 32, [3, 3], padding="SAME", scope="d_conv3", reuse=reuse)
out = slim.conv2d(conv3, 1, [28, 28], padding="VALID", scope="d_conv4", reuse=reuse)
return out
def discriminate(input_data, reuse=False):
with tf.variable_scope("discriminate"):
with slim.arg_scope(
[slim.fully_connected],
weights_initializer=tf.truncated_normal_initializer(
0.0, 0.1),
weights_regularizer=slim.l1_l2_regularizer(),
activation_fn=None):
fc1 = slim.fully_connected(inputs=input_data,
num_outputs=length,
scope="d_fc1",
reuse=reuse)
fc1 = tf.tanh(fc1)
fc2 = slim.fully_connected(inputs=fc1,
num_outputs=length,
scope="d_fc2",
reuse=reuse)
fc2 = tf.tanh(fc2)
fc3 = slim.fully_connected(inputs=fc2,
num_outputs=1,
scope="d_fc3",
reuse=reuse)
fc3 = tf.tanh(fc3)
fc3 = tf.sigmoid(fc3)
return fc3
def Encode(self, input_img):
# I will use conv layer to imporove performance
with tf.variable_scope('encoder', reuse=tf.AUTO_REUSE) as scope:
conv_l = conv2d(input_img,
32, [3, 3],
2,
weights_regularizer=slim.l1_l2_regularizer(.001),
scope='conv_1')
conv_l = conv2d(conv_l,
64, [3, 3],
2,
weights_regularizer=slim.l1_l2_regularizer(.005),
scope='conv_2')
conv_l = conv2d(conv_l,
128, [3, 3],
4,
weights_regularizer=slim.l1_regularizer(.005),
scope=scope)
return conv_l
def generate(input_data, reuse=False):
with tf.variable_scope("generate"):
with slim.arg_scope([slim.fully_connected],
weights_initializer=tf.truncated_normal_initializer(0.0, 0.1),
weights_regularizer=slim.l1_l2_regularizer(),activation_fn=None
):
fc1 = slim.fully_connected(inputs=input_data, num_outputs=length, scope="g_fc1", reuse=reuse)
fc1 = tf.nn.softplus(fc1,name="g_softplus")
fc2 = slim.fully_connected(inputs=fc1, num_outputs=length, scope="g_fc2", reuse=reuse)
return fc2
def Decode(self, Decoded_img):
with tf.variable_scope('decoder', reuse=tf.AUTO_REUSE) as scope:
conv_d = slim.conv2d_transpose(
Decoded_img,
64, [3, 3],
4,
weights_regularizer=slim.l1_l2_regularizer(.001),
scope='dconv_1')
conv_d = slim.conv2d_transpose(
conv_d,
32, [3, 3],
2,
weights_regularizer=slim.l1_l2_regularizer(.005),
scope='dconv_2')
conv_d = slim.conv2d_transpose(
conv_d,
3, [3, 3],
2,
weights_regularizer=slim.l1_regularizer(.005),
scope=scope)
return conv_d
def ds_cnn_arg_scope(scale_l1, scale_l2): """Defines the default ds_cnn argument scope. Args: weight_decay: The weight decay to use for regularizing the model. Returns: An `arg_scope` to use for the DS-CNN model. """ with slim.arg_scope( [slim.convolution2d, slim.separable_convolution2d], weights_initializer=slim.initializers.xavier_initializer(), biases_initializer=slim.init_ops.zeros_initializer(), weights_regularizer=slim.l1_l2_regularizer(scale_l1=scale_l1, scale_l2=scale_l2)) as sc: return sc
def create_model(self,
model_input,
vocab_size,
num_mixtures=None,
l2_penalty=1e-6,
**unused_params):
num_mixtures = num_mixtures or FLAGS.MoNN_num_experts
gate_activations = slim.fully_connected(
model_input,
vocab_size * (num_mixtures + 1),
activation_fn=None,
biases_initializer=None,
weights_regularizer=slim.l2_regularizer(l2_penalty),
scope="gates")
h1Units = 4096
A1 = slim.fully_connected(
model_input,
h1Units,
activation_fn=tf.nn.relu,
weights_regularizer=slim.l2_regularizer(l2_penalty),
scope='FC_H1')
h2Units = 4096
A2 = slim.fully_connected(
A1,
h2Units,
activation_fn=tf.nn.relu,
weights_regularizer=slim.l1_l2_regularizer(l2_penalty),
scope='FC_H2')
#
expert_activations = slim.fully_connected(
A2,
vocab_size * num_mixtures,
activation_fn=None,
weights_regularizer=slim.l2_regularizer(l2_penalty),
scope="experts")
gating_distribution = tf.nn.softmax(
tf.reshape(gate_activations,
[-1, num_mixtures + 1
])) # (Batch * #Labels) x (num_mixtures + 1)
expert_distribution = tf.nn.sigmoid(
tf.reshape(expert_activations,
[-1, num_mixtures])) # (Batch * #Labels) x num_mixtures
final_probabilities_by_class_and_batch = tf.reduce_sum(
gating_distribution[:, :num_mixtures] * expert_distribution, 1)
final_probabilities = tf.reshape(
final_probabilities_by_class_and_batch, [-1, vocab_size])
return {"predictions": final_probabilities}
def generate(inputData,reuse=False):
"""
生成器
:param inputData:
:param reuse:
:return:
"""
with tf.variable_scope("generate"):
"""
arg_scope操作，这一操作符可以让定义在这一scope中的操作共享参数，即如不制定参数的话，则使用默认参数。且参数可以被局部覆盖。使得代码更加简洁
"""
with slim.arg_scope([slim.fully_connected],
weights_initializer=tf.truncated_normal_initializer(0.0,0.1),
weights_regularizer=slim.l1_l2_regularizer(),
activation_fn=None):
fc1=slim.fully_connected(inputs=inputData,num_outputs=length,scope="g_fc1",reuse=reuse)
fc1=tf.nn.softplus(features=fc1,name="g_softplus")
fc2=slim.fully_connected(inputs=fc1,num_outputs=length,scope="g_fc2",reuse=reuse)
return fc2