def discriminator(x, z_size, is_training, reuse, n_hid=500, isize=28 * 28):
fc = dense(x, n_hid, is_training, reuse, w_normalized=True, name='fc1')
fc = tf.nn.relu(fc, name='fc1/relu')
fc = dense(fc, n_hid, is_training, reuse, w_normalized=True, name='fc2')
fc = tf.nn.relu(fc, name='fc2/relu')
out = dense(fc, 1, is_training, reuse, name='d_out')
return out
def generator(z, is_training, reuse, n_hid=500, isize=28 * 28, use_bn=False):
fc = dense(z, n_hid, is_training, reuse, w_regularizer=None, name='fc1')
if use_bn:
fc = tf.nn.relu(
batch_norm(fc, is_training=is_training, reuse=reuse,
name='fc1/bn'))
else:
fc = tf.nn.relu(hid, name='fc1/prelu')
fc = dense(fc, n_hid, is_training, reuse, w_regularizer=None, name='fc2')
if use_bn:
fc = tf.nn.relu(
batch_norm(fc, is_training=is_training, reuse=reuse,
name='fc2/bn'))
else:
fc = rtf.nn.elu(fc, name='fc2/prelu')
out = tf.nn.sigmoid(
dense(fc, isize, is_training, reuse, w_regularizer=None, name='g_out'))
return out
def discriminator_from_params(x,
params,
is_training,
reuse,
isize=28 * 28,
n_hid=100):
fc = dense(x,
n_hid,
is_training,
reuse,
w_normalized=True,
name='fc1',
params=params[:2])
fc = tf.nn.relu(fc, name='fc1/relu')
fc = dense(fc,
n_hid,
is_training,
reuse,
w_normalized=True,
name='fc2',
params=params[2:4])
fc = tf.nn.relu(fc, name='fc2/relu')
out = dense(fc, 1, is_training, reuse, name='d_out', params=params[4:])
return out