def dynamic_L(x):
'''Attention machanism to combine the infomation,
from https://arxiv.org/pdf/1612.01627.pdf.
Args:
x: a tensor with shape [batch, time, dimension]
Returns:
a tensor with shape [batch, dimension]
Raises:
'''
key_0 = tf.get_variable(
name='key',
shape=[x.shape[-1]],
dtype=tf.float32,
initializer=tf.random_uniform_initializer(
-tf.sqrt(6./tf.cast(x.shape[-1], tf.float32)),
tf.sqrt(6./tf.cast(x.shape[-1], tf.float32))))
key = op.dense(x, add_bias=False) #[batch, time, dimension]
weight = tf.reduce_sum(tf.multiply(key, key_0), axis=-1) #[batch, time]
weight = tf.expand_dims(tf.nn.softmax(weight), -1) #[batch, time, 1]
L = tf.reduce_sum(tf.multiply(x, weight), axis=1) #[batch, dimension]
return L
def FFN(x, out_dimension_0=None, out_dimension_1=None):
'''Add two dense connected layer, max(0, x*W0+b0)*W1+b1.
Args:
x: a tensor with shape [batch, time, dimension]
out_dimension: a number which is the output dimension
Returns:
a tensor with shape [batch, time, out_dimension]
Raises:
'''
with tf.variable_scope('FFN_1'):
y = op.dense(x, out_dimension_0) #x*W0+b0
y = tf.nn.relu(y)
with tf.variable_scope('FFN_2'): #max(0, x*W0+b0)*W1+b1
z = op.dense(y, out_dimension_1) #, add_bias=False) #!!!!
return z
def __call__(self, z, reuse=False):
with tf.variable_scope(self.name) as vs:
if reuse:
vs.reuse_variables()
ch=self.fmd[0]*self.fmd[0]*self.channela[0]
hf = dense(z, ch,use_wscale=False)
hc=tf.nn.relu(tf.reshape(hf,[-1, self.fmd[0], self.fmd[0], self.channela[0]])) # 4x4
fmd1=tf.nn.relu(deconv2d(hc,self.channela[1],k=self.kersize[0],use_wscale=False,name='deconv1')) #8x8
fmd2=tf.nn.relu(deconv2d(fmd1,self.channela[2],k=self.kersize[1],use_wscale=False,name='deconv2')) #16x16
fmd3=tf.nn.relu(deconv2d(fmd2,self.channela[3],k=self.kersize[2],use_wscale=False,name='deconv3')) #32x32
gx = tf.nn.sigmoid(deconv2d(fmd3,3,k=self.kersize[3],use_wscale=False,name='deconv4')) #64x64
return gx
def __call__(self, z, reuse=False):
with tf.variable_scope(self.name) as vs:
if reuse:
vs.reuse_variables()
ch=self.fmd[0]*self.fmd[0]*self.channelg[0]
hf = dense(z, ch,use_wscale=False)
hc=tf.nn.relu(tf.reshape(hf,[-1, self.fmd[0], self.fmd[0], self.channelg[0]])) # 4x4
fmd1=tf.nn.relu(deconv2d(hc,self.channelg[1],k=self.kersize[0],use_wscale=False,name='deconv1')) #8x8
fmd2=tf.nn.relu(deconv2d(fmd1,self.channelg[2],k=self.kersize[1],use_wscale=False,name='deconv2')) #16x16
fmd3=tf.nn.relu(deconv2d(fmd2,self.channelg[3],k=self.kersize[2],use_wscale=False,name='deconv3')) #32x32
gdfq = tf.nn.tanh(deconv2d(fmd3,2,k=self.kersize[3],use_wscale=False,name='deconv4')) #64x64
gdf = tf.reshape(gdfq,(-1,self.imgsize*self.imgsize,2))
gdf = tf.transpose(gdf,[0,2,1])
return gdf