def discriminator_resnet_stl10(x, x_shape, dim=64, ss_task = 0,\
name = 'discriminator', reuse=False):
relu = tf.nn.relu
with tf.variable_scope(name, reuse=reuse):
image = tf.reshape(
x, [-1, x_shape[0], x_shape[1], x_shape[2]]) #48 * 48 * 3
h0 = optimized_block(image, dim, 'd_optimized_block1', \
act=relu) # 24 * 24 * dim
h1 = d_block(h0, dim * 2, 'd_block2', act=relu) # 12 * 12 * dim * 2
h2 = d_block(h1, dim * 4, 'd_block3', act=relu) # 6 * 6 * dim * 4
h3 = d_block(h2, dim * 8, 'd_block4', act=relu) # 3 * 3 * dim * 8
h4 = d_block(h3, dim * 16, 'd_block5', None,
act=relu) # 3 * 3 * dim * 16
h4_act = relu(h4)
feat = tf.reshape(h4_act, [-1, dim * 16 * 3 * 3])
h5 = tf.reduce_sum(h4_act, [1, 2])
out = ops.linear(h5, 1, scope='linear_out')
if ss_task == 1:
k = 4
elif ss_task == 2:
k = 5
else:
k = -1
if ss_task > 0:
print(
'[net_resnet.py -- discriminator_resnet_stl10] SS task = %d with k = %d classes'
% (ss_task, k))
cls = ops.linear(feat, k)
return tf.sigmoid(out), out, feat, cls
else:
return tf.sigmoid(out), out, feat
def discriminator_resnet_cifar(x, x_shape, dim=128, ss_task = 0,\
name = 'discriminator', reuse=False):
relu = tf.nn.relu
with tf.variable_scope(name, reuse=reuse):
image = tf.reshape(x, [-1, x_shape[0], x_shape[1], x_shape[2]])
h0 = optimized_block(image, dim, 'd_optimized_block1', \
act=relu) # 16 x 16 x dim
h1 = d_block(h0, dim, 'd_block2', act=relu) # 8 x 8 x dim
h2 = d_block(h1, dim, 'd_block3', None, act=relu) # 8 x 8 x dim
h3 = d_block(h2, dim, 'd_block4', None, act=relu) # 8 x 8 x dim
h3_act = relu(h3)
feat = tf.reshape(h3_act, [-1, 8 * 8 * dim])
h4 = tf.reduce_sum(h3_act, [1, 2])
out = ops.linear(h4, 1, scope='linear_out')
if ss_task == 1:
k = 4
elif ss_task == 2:
k = 5
else:
k = -1
if ss_task > 0:
print(
'[net_resnet.py -- discriminator_dcgan_cifar] SS task = %d with k = %d classes'
% (ss_task, k))
cls = ops.linear(feat, k)
return tf.sigmoid(out), out, feat, cls
else:
return tf.sigmoid(out), out, feat
def generator_resnet_cifar(z, x_shape, dim=128, name = 'generator', \
reuse=False, training=True):
dim = dim * 2 # 256 like sn-gan paper
x_dim = x_shape[0] * x_shape[1] * x_shape[2]
with tf.variable_scope(name, reuse=reuse):
act0 = ops.linear(z, dim * 4 * 4, scope='g_linear0')
act0 = tf.reshape(act0, [-1, 4, 4, dim])
act1 = g_block(act0, dim, training, 'g_block1') # 8 * 8
act2 = g_block(act1, dim, training, 'g_block2') # 16 * 16
act3 = g_block(act2, dim, training, 'g_block3') # 32 * 32
bn = ops.batch_norm(name='g_bn')
act3 = tf.nn.relu(bn(act3, training))
act4 = ops.conv2d(act3, 3, 3, 3, 1, 1, name='g_conv_last')
out = tf.nn.sigmoid(act4)
return tf.reshape(out, [-1, x_dim])
def generator_resnet_stl10(z, x_shape, dim=64, name = 'generator', \
reuse=False, training=True):
x_dim = x_shape[0] * x_shape[1] * x_shape[2]
with tf.variable_scope(name, reuse=reuse):
act0 = ops.linear(z, dim * 8 * 6 * 6, scope='g_linear0')
act0 = tf.reshape(act0, [-1, 6, 6, dim * 8]) # 6 * 6 * dim * 8
act1 = g_block(act0, dim * 4, training,
'g_block1') # 12 * 12 * dim * 4
act2 = g_block(act1, dim * 2, training,
'g_block2') # 24 * 24 * dim * 2
act3 = g_block(act2, dim * 1, training,
'g_block3') # 48 * 48 * dim * 1
bn = ops.batch_norm(name='g_bn')
act3 = tf.nn.relu(bn(act3, training))
act4 = ops.conv2d(act3, 3, 3, 3, 1, 1, name='g_conv_last')
out = tf.nn.sigmoid(act4)
return tf.reshape(out, [-1, x_dim])
def encoder_resnet_stl10(x, x_shape, z_dim=128, dim=64, \
name = 'encoder', reuse=False, training=True):
act = lrelu
with tf.variable_scope(name, reuse=reuse):
image = tf.reshape(x, [-1, x_shape[0], x_shape[1], x_shape[2]])
image = ops.conv2d(image, dim, 3, 3, 1, 1, \
name='e_conv0') # 48 * 48 * dim
act0 = e_block(image, dim * 2, training = training,\
name = 'e_block1', act=act) # 24 * 24 * dim * 2
act1 = e_block(act0, dim * 4, training, \
name = 'e_block2', act=act) # 12 * 12 * dim * 4
act2 = e_block(act1, dim * 8, training, \
name = 'e_block3', act=act)# 6 * 6 * dim * 8
bn = ops.batch_norm(name='e_bn')
act2 = act(bn(act2, training))
act2 = tf.reshape(act2, [-1, 6 * 6 * dim * 8])
out = ops.linear(act2, z_dim)
return out
def encoder_resnet_cifar(x, x_shape, z_dim=128, dim=128, \
name = 'encoder', reuse=False, training=True):
dim = dim * 2 # 256 like sn-gan paper
act = lrelu
with tf.variable_scope(name, reuse=reuse):
image = tf.reshape(x, [-1, x_shape[0], x_shape[1], x_shape[2]])
image = ops.conv2d(image, dim, 3, 3, 1, 1, \
name='e_conv0') # 32 * 32
act0 = e_block(image, dim, training = training,\
name = 'e_block1', act=act) # 16 * 16
act1 = e_block(act0, dim, training, \
name = 'e_block2', act=act) # 8 * 8
act2 = e_block(act1, dim, training, \
name = 'e_block3', act=act)# 4 * 4
bn = ops.batch_norm(name='e_bn')
act2 = act(bn(act2, training))
act2 = tf.reshape(act2, [-1, 4 * 4 * dim])
out = ops.linear(act2, z_dim)
return out