def batch_norm(x, name, is_training=True):
decay_rate = 0.99
shape = x.get_shape().as_list()
dim = shape[-1]
if len(shape) == 2:
mean, var = tf.nn.moments(x, [0], name='moments_bn_{}'.format(name))
elif len(shape) == 4:
mean, var = tf.nn.moments(x, [0, 1, 2],
name='moments_bn_{}'.format(name))
avg_mean = get_biases('avg_mean_bn_{}'.format(name), [1, dim], 0.0, False)
avg_var = get_biases('avg_var_bn_{}'.format(name), [1, dim], 1.0, False)
beta = get_biases('beta_bn_{}'.format(name), [1, dim], 0.0)
gamma = get_biases('gamma_bn_{}'.format(name), [1, dim], 1.0)
if is_training:
avg_mean_assign_op = tf.assign(
avg_mean, decay_rate * avg_mean + (1 - decay_rate) * mean)
avg_var_assign_op = tf.assign(
avg_var, decay_rate * avg_var + (1 - decay_rate) * var)
with tf.control_dependencies([avg_mean_assign_op, avg_var_assign_op]):
ret = gamma * (x - mean) / tf.sqrt(1e-6 + var) + beta
else:
ret = gamma * (x - avg_mean) / tf.sqrt(1e-6 + avg_var) + beta
return ret
def set_model(self, figs, is_training, reuse=False):
# return only logits
h = figs
# convolution
with tf.variable_scope(self.name_scope_conv, reuse=reuse):
for i, (in_chan, out_chan) in enumerate(
zip(self.layer_chanels, self.layer_chanels[1:])):
if i == 0:
conved = conv_layer(inputs=h,
out_num=out_chan,
filter_width=5,
filter_hight=5,
stride=1,
l_id=i)
h = tf.nn.relu(conved)
#h = lrelu(conved)
else:
conved = conv_layer(inputs=h,
out_num=out_chan,
filter_width=5,
filter_hight=5,
stride=2,
l_id=i)
bn_conved = batch_norm(conved, i, is_training)
h = tf.nn.relu(bn_conved)
#h = lrelu(bn_conved)
feature_image = h
# full connect
dim = get_dim(h)
h = tf.reshape(h, [-1, dim])
with tf.variable_scope(self.name_scope_fc, reuse=reuse):
weights = get_weights('fc', [dim, self.fc_dim], 0.02)
biases = get_biases('fc', [self.fc_dim], 0.0)
h = tf.matmul(h, weights) + biases
h = batch_norm(h, 'fc', is_training)
h = tf.nn.relu(h)
weights = get_weights('fc2', [self.fc_dim, 1], 0.02)
biases = get_biases('fc2', [1], 0.0)
h = tf.matmul(h, weights) + biases
return h, feature_image
def set_model(self, z, batch_size, is_training, reuse = False):
# reshape z
with tf.variable_scope(self.name_scope_reshape, reuse = reuse):
w_r = get_weights('_r',
[self.z_dim, self.in_dim * self.in_dim * self.layer_chanels[0]],
0.02)
b_r = get_biases('_r',
[self.in_dim * self.in_dim * self.layer_chanels[0]],
0.0)
h = tf.matmul(z, w_r) + b_r
h = batch_norm(h, 'reshape', is_training)
#h = tf.nn.relu(h)
h = lrelu(h)
h = tf.reshape(h, [-1, self.in_dim, self.in_dim, self.layer_chanels[0]])
# deconvolution
layer_num = len(self.layer_chanels) - 1
with tf.variable_scope(self.name_scope_deconv, reuse = reuse):
for i, (in_chan, out_chan) in enumerate(zip(self.layer_chanels, self.layer_chanels[1:])):
deconved = deconv_layer(inputs = h,
out_shape = [batch_size, self.in_dim * 2 ** (i + 1), self.in_dim * 2 **(i + 1), out_chan],
filter_width = 5, filter_hight = 5,
stride = 2, l_id = i)
if i == layer_num -1:
h = tf.nn.tanh(deconved)
else:
bn_deconved = batch_norm(deconved, i, is_training)
#h = tf.nn.relu(bn_deconved)
h = lrelu(bn_deconved)
return h
def set_model(self, figs, is_training, reuse=False):
u'''
return only logits. not sigmoid(logits).
'''
h = figs
# convolution
with tf.variable_scope(self.name_scope_conv, reuse=reuse):
for i, (in_chan, out_chan) in enumerate(
zip(self.layer_chanels, self.layer_chanels[1:])):
conved = conv_layer(inputs=h,
out_num=out_chan,
filter_width=5,
filter_hight=5,
stride=2,
l_id=i)
if i == 0:
h = tf.nn.relu(conved)
#h = lrelu(conved)
else:
bn_conved = batch_norm(conved, i, is_training)
h = tf.nn.relu(bn_conved)
#h = lrelu(bn_conved)
# full connect
dim = get_dim(h)
h = tf.reshape(h, [-1, dim])
with tf.variable_scope(self.name_scope_fc, reuse=reuse):
weights = get_weights('fc', [dim, self.fc_dim], 0.02)
biases = get_biases('fc', [self.fc_dim], 0.0)
h = tf.matmul(h, weights) + biases
h = batch_norm(h, 'en_fc_bn', is_training)
h = tf.nn.relu(h)
weights = get_weights('mu', [self.fc_dim, self.z_dim], 0.02)
biases = get_biases('mu', [self.z_dim], 0.0)
mu = tf.matmul(h, weights) + biases
weights = get_weights('sigma', [self.fc_dim, self.z_dim], 0.02)
biases = get_biases('sigma', [self.z_dim], 0.0)
log_sigma = tf.matmul(h, weights) + biases
return mu, log_sigma