def generator(noise, reuse=False, alpha=0.2, training=True):
"""
Generator model that takes `noise` as input. `alpha` is the
Leaky-ReLU paramter for slope.
"""
with tf.variable_scope('generator', reuse=reuse):
x = dense(noise, 4 * 4 * 512)
x = tf.reshape(x, (-1, 4, 4, 512))
x = batch_normalization(x, training=training)
x = tf.maximum(alpha * x, x)
x = conv2d_transpose(x, 256, 5, 2, padding='same')
x = batch_normalization(x, training=training)
x = tf.maximum(alpha * x, x)
x = conv2d_transpose(x, 128, 5, 2, padding='same')
x = batch_normalization(x, training=training)
x = tf.maximum(alpha * x, x)
x = conv2d_transpose(x, 64, 5, 2, padding='same')
x = batch_normalization(x, training=training)
x = tf.maximum(alpha * x, x)
logits = conv2d_transpose(x, 3, 5, 2, padding='same')
out = tf.tanh(logits)
return out
def generator(noise, reuse=False, alpha=0.2, training=True):
with tf.variable_scope('generator', reuse=reuse):
x = dense(noise, 4 * 4 * 512)
x = tf.reshape(x, (-1, 4, 4, 512))
x = batch_normalization(x, training=training)
x = tf.maximum(alpha * x, x)
x = conv2d_transpose(x, 256, 5, 2, padding='same')
x = batch_normalization(x, training=training)
x = tf.maximum(alpha * x, x)
x = conv2d_transpose(x, 128, 5, 2, padding='same')
x = batch_normalization(x, training=training)
x = tf.maximum(alpha * x, x)
x = conv2d_transpose(x, 64, 5, 2, padding='same')
x = batch_normalization(x, training=training)
x = tf.maximum(alpha * x, x)
logits = conv2d_transpose(x, 3, 5, 2, padding='same')
out = tf.tanh(logits)
return out, logits
def decode(self, z, x_dim, training=False):
im_h, im_w, im_c = self.image_shape
with tf.variable_scope("decoder", reuse=tf.AUTO_REUSE):
h = z
h = tfl.dense(h, units=self.h_dim, activation=tf.nn.relu)
h = tfl.dense(h, units=self.h_dim, activation=tf.nn.relu)
stride = 16
h = tfl.dense(
h, units=im_h // stride * im_w // stride * self.kernel_num * 2,
activation=tf.nn.relu)
new_shape = (-1, im_h // stride, im_w // stride, self.kernel_num * 2)
h = tf.reshape(h, new_shape)
h = tfl.conv2d_transpose(
h, self.kernel_num * 2, self.kernel_size,
strides=2, padding="same", activation=tf.nn.relu)
h = tfl.conv2d_transpose(
h, self.kernel_num, self.kernel_size,
strides=2, padding="same", activation=tf.nn.relu)
h = tfl.conv2d_transpose(
h, self.kernel_num, self.kernel_size,
strides=2, padding="same", activation=tf.nn.relu)
h = tfl.conv2d_transpose(
h, im_c, self.kernel_size,
strides=2, padding="same", activation=tf.nn.sigmoid)
h = tfl.flatten(h)
y_mean = h
return y_mean
def generatorNet(name, inputs, is_training, use_batchNorm, reuse=None):
idx = 0
f = inputs
with tf.variable_scope(name, reuse=reuse):
f = layers.dense(f, 1024, None, name="dense_%d" % idx)
if use_batchNorm:
f = layers.batch_normalization(f, training=is_training, name="bn_%d" % idx)
f = tf.nn.relu(f, "relu_%d" % idx)
idx += 1
f = layers.dense(f, 7 * 7 * 128, None, name="dense_%d" % idx) # 6272
if use_batchNorm:
f = layers.batch_normalization(f, training=is_training, name="bn_%d" % idx)
f = tf.nn.relu(f, "relu_%d" % idx)
f = tf.reshape(f, [-1, 7, 7, 128], name="reshape_%d" % idx)
idx += 1
f = layers.conv2d_transpose(f, 64, kernel_size=4, strides=2, padding="SAME", name="deconv_%d" % idx)
if use_batchNorm:
f = layers.batch_normalization(f, training=is_training, name="bn_%d" % idx)
f = tf.nn.relu(f, "relu_%d" % idx)
idx += 1
f = layers.conv2d_transpose(f, 1, kernel_size=4, strides=2, padding="SAME", name="deconv_%d" % idx)
f = tf.nn.sigmoid(f, "sigmoid_%d" % idx)
return f
def deconv_net(inputs):
# *** WARNING: HARDCODED SHAPES, CHANGE ACCORDINGLY ***
dense = layers.dense(inputs, units=7 * 7 * 16, activation=tf.nn.relu)
# input(dense)
dense = tf.reshape(dense, [16, 7, 7, 16])
# 1st deconv layer
deconv1 = layers.conv2d_transpose(dense,
filters=32,
kernel_size=[5, 5],
padding='same',
activation=tf.nn.relu)
# input(deconv1)
# 2nd deconv layer
deconv2 = layers.conv2d_transpose(deconv1,
filters=16,
kernel_size=[5, 5],
padding='same',
activation=tf.nn.relu)
# input(deconv2)
out = layers.conv2d_transpose(deconv2,
filters=16,
kernel_size=[3, 3],
padding='same')
return out
def __init__(self, sparse, guidance_map, params, reuse=False):
with tf.variable_scope("Local", reuse=reuse) as scope:
sparse = tf.reshape(sparse, [-1, 200, 200, 1])
guidance_map = tf.reshape(guidance_map, [-1, 200, 200, 1])
x = sparse + guidance_map
x = conv2d(x, 32, (3, 3), strides=(2, 2), padding='same')
x = relu(x)
x = conv2d(x, 64, (3, 3), strides=(2, 2), padding='same')
x = relu(x)
x = conv2d_transpose(x,
64, (5, 5),
strides=(2, 2),
padding='same',
use_bias=False)
x = batch_normalization(x)
x = relu(x)
x = conv2d_transpose(x,
2, (5, 5),
strides=(2, 2),
padding='same',
use_bias=False)
self.output = x
self.parameters = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope="Local")
def __init__(self, images, sparse, params, reuse=False):
with tf.variable_scope("Global", reuse=reuse) as scope:
sparse = tf.reshape(sparse, [-1, 200, 200, 1])
images = tf.reshape(images, [-1, 200, 200, 3])
x = tf.concat([images, sparse], axis=3)
x = tf.cast(x, dtype=tf.float32)
x = conv2d(x, 32, (3, 3), strides=(2, 2), padding='same')
x = relu(x)
x = conv2d(x, 64, (3, 3), strides=(2, 2), padding='same')
x = relu(x)
x = conv2d_transpose(x,
64, (5, 5),
strides=(2, 2),
padding='same',
use_bias=False)
x = batch_normalization(x)
x = relu(x)
x = conv2d_transpose(x,
3, (5, 5),
strides=(2, 2),
padding='same',
use_bias=False)
self.output = x
self.parameters = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope="Global")
def generator_fn(input_image, mode):
with tf.name_scope('generator'):
#4*4
dense_1 = layers.dense(inputs=input_image, units=batch_size*16)
batch_norm_1 = layers.batch_normalization(inputs=dense_1)
reshape_1 = tf.reshape(batch_norm_1, shape=(batch_size, 4, 4, batch_size))
relu_1 = tf.nn.relu(reshape_1)
# 8*8
conv_T_1 = layers.conv2d_transpose(inputs=relu_1, filters=64, kernel_size=(2, 2), strides=(2, 2), padding='same')
batch_norm_2 = layers.batch_normalization(inputs=conv_T_1)
relu_2 = tf.nn.relu(batch_norm_2)
# 16*16
conv_T_2 = layers.conv2d_transpose(inputs=relu_2, filters=32, kernel_size=(2, 2), strides=(2, 2), padding='same')
batch_norm_3 = layers.batch_normalization(inputs=conv_T_2)
relu_3 = tf.nn.relu(batch_norm_3)
# 32*32
conv_T_3 = layers.conv2d_transpose(inputs=relu_3, filters=16, kernel_size=(2, 2), strides=(2, 2), padding='same')
batch_norm_4 = layers.batch_normalization(inputs=conv_T_3)
relu_4 = tf.nn.relu(batch_norm_4)
# 64*64
conv_T_4 = layers.conv2d_transpose(
inputs=relu_4, filters=3, kernel_size=(2, 2), strides=(2, 2), padding='same')
tanh_1 = tf.nn.tanh(conv_T_4)
print(tanh_1)
return tanh_1
def generator(inp0, dim, name, reuse):
# inp0 = tf.placeholder(tf.float32, [None, G.size, G.size, 1]) # G.size == 2 ** 8
with tf.variable_scope(name, reuse=reuse):
ten1 = tl.conv2d(inp0, 1 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten2 = tl.conv2d(ten1, 2 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten3 = tl.conv2d(ten2, 4 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten4 = tl.conv2d(ten3, 8 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten5 = tl.conv2d(ten4, 16 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten6 = tl.conv2d(ten5, 16 * dim, 3, 1, 'same', activation=tf.nn.leaky_relu)
ten6 = tl.conv2d(ten6, 16 * dim, 3, 1, 'same', activation=tf.nn.leaky_relu)
ten6 += ten5
ten7 = tl.conv2d(ten6, 16 * dim, 3, 1, 'same', activation=tf.nn.leaky_relu)
ten7 = tl.conv2d(ten7, 16 * dim, 3, 1, 'same', activation=tf.nn.leaky_relu)
ten7 += ten6
ten8 = tl.conv2d_transpose(ten7, 16 * dim, 3, 1, 'same', activation=leRU_batch_norm)
ten8 = tl.conv2d_transpose(ten8, 16 * dim, 3, 1, 'same', activation=leRU_batch_norm)
ten8 += ten7
ten9 = tl.conv2d_transpose(ten8, 16 * dim, 3, 1, 'same', activation=leRU_batch_norm)
ten9 = tl.conv2d_transpose(ten9, 16 * dim, 3, 1, 'same', activation=leRU_batch_norm)
ten9 += ten8
ten5 = tl.conv2d_transpose(ten9, 16 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten4 = tl.conv2d_transpose(ten5, 8 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten3 = tl.conv2d_transpose(ten4, 4 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten2 = tl.conv2d_transpose(ten3, 2 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten1 = tl.conv2d_transpose(ten2, 1 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten1 = tf.concat((ten1, inp0), axis=3)
ten1 = tl.conv2d(ten1, 1 * dim, 3, 1, 'same', activation=leRU_batch_norm)
ten1 = tl.conv2d(ten1, 3, 3, 1, 'same', activation=tf.nn.tanh)
ten1 = ten1 * 0.505 + 0.5
return ten1
def generator_net(inputs, scope, reuse=None, rgb=False): output_channels = 3 if rgb else 1 with tf.variable_scope(scope, reuse=reuse): # branch 1 ( color reconstruction) cv1 = conv2d(inputs, filters=16, kernel_size=3, strides=1, padding='same', activation=None, name='conv9_i') cv1_r = leaky_relu(cv1) res1_c = conv2d(cv1_r, filters=16, kernel_size=5, strides=1, padding='same', activation=None, name='conv3a_1') res1_b = batch_normalization(res1_c) res1_r = leaky_relu(res1_b) res1_d = conv2d(res1_r, filters=16, kernel_size=3, strides=1, padding='same', activation=None, name='conv3b_1') res1 = batch_normalization(res1_d) sum1 = cv1 + res1 res2_c = conv2d(sum1, filters=16, kernel_size=3, strides=1, padding='same', activation=None, name='conv3a_2') res2_b = batch_normalization(res2_c) res2_r = leaky_relu(res2_b) res2_d = conv2d(res2_r, filters=16, kernel_size=3, strides=1, padding='same', activation=None, name='conv3b_2') res2 = batch_normalization(res2_d) br1 = sum1 + res2 # branch 2 (features extraction) br2 = conv2d(inputs, filters=16, kernel_size=5, strides=1, padding='same', activation=tf.nn.leaky_relu, name='conv_bf1') br2 = max_pooling2d(br2, pool_size=2, strides=2, name='maxpool1') br2 = conv2d(br2, filters=16, kernel_size=3, strides=1, padding='same', activation=tf.nn.leaky_relu, name='conv_bf2') br2 = max_pooling2d(br2, pool_size=2, strides=2, name='maxpool2a') br2 = conv2d(br2, filters=16, kernel_size=3, strides=1, padding='same', activation=tf.nn.leaky_relu, name='conv_bf3') br2 = max_pooling2d(br2, pool_size=2, strides=2, name='maxpool2') print(br2.shape) br2 = conv2d_transpose(br2, filters=16, kernel_size=3, padding='same', strides=2, activation=tf.nn.leaky_relu, name="deconv_1") print(br2.shape) br2 = conv2d_transpose(br2, filters=16, kernel_size=3, padding='same', strides=2, activation=tf.nn.leaky_relu, name="deconv_2") print(br2.shape) br2 = conv2d_transpose(br2, filters=16, kernel_size=3, padding='same', strides=2, activation=tf.nn.leaky_relu, name="deconv_3") print(br2.shape) # concatenate branches and reconstruct image sum3 = tf.concat((br1, br2), axis=3); model = conv2d(sum3, filters=output_channels, kernel_size=3, strides=1, padding='same', activation=None, name='conv9_f') return model
def _build_generator(input_data, name='generator'):
with tf.variable_scope(name):
net = layers.dense(input_data, 128)
net = tf.nn.relu(net)
net = tf.reshape(net, [-1, 4, 4, 8])
net = layers.conv2d_transpose(net, 128, [5, 5], activation=tf.nn.relu, strides=[2, 2], padding='same') # 8x8
net = layers.batch_normalization(net, momentum=0.9, training=True)
net = layers.conv2d_transpose(net, 64, [5, 5], activation=tf.nn.relu, strides=[2, 2]) # 19x19
net = layers.batch_normalization(net, momentum=0.9, training=True)
net = layers.conv2d_transpose(net, 32, [5, 5], activation=tf.nn.relu) # 23x23
net = layers.batch_normalization(net, momentum=0.9, training=True)
net = layers.conv2d_transpose(net, 16, [5, 5], activation=tf.nn.relu) # 27x27
net = layers.batch_normalization(net, momentum=0.9, training=True)
net = layers.conv2d_transpose(net, 1, [2, 2], activation=tf.nn.relu) # 28x28
return net
def generator(
self,
noise,
label,
training,
reuse=False,
):
with tf.variable_scope('generator', reuse=reuse):
net = tf.concat([label, noise], axis=-1) # (batch_size, 445)
net = layers.dense(
net,
self.g_dim * 3 * 3 * 4, # (batch_size, 32768)
kernel_initializer=tf.truncated_normal_initializer(
mean=0, stddev=0.02))
net = tf.nn.leaky_relu(net)
net = tf.reshape(
net, [-1, 3, 3, self.g_dim * 4]) # Reshape (3 x 3 x 512)
net = deconv2d(net,
self.g_dim * 2,
strides=[1, 1],
padding='valid',
training=training) # (7 x 7 x 256)
net = deconv2d(net, self.g_dim,
training=training) # (14 x 14 x 128)
net = layers.conv2d_transpose(
net,
1,
kernel_size=(5, 5),
strides=(2, 2),
activation=tf.nn.tanh,
padding='same',
kernel_initializer=tf.truncated_normal_initializer(
mean=0, stddev=0.02)) # (28 x 28 x 1)
return net # (28, 28, 1)
def deconv_block(self, input, filters, conv, padding, scope):
with tf.variable_scope(scope):
deconv1 = conv2d_transpose(input,
filters,
kernel_size=(3, 3),
strides=[2, 2],
padding=padding)
deconv_shape = tf.shape(deconv1)
conv_shape = tf.shape(conv)
offsets = [
0, (conv_shape[1] - deconv_shape[1]) // 2,
(conv_shape[2] - deconv_shape[2]) // 2, 0
]
size = [-1, deconv_shape[1], deconv_shape[2], filters]
conv_crop = tf.slice(conv, offsets, size)
conv1 = tf.concat([deconv1, conv_crop], 3)
bn = batch_normalization(conv1)
drop = dropout(bn, .25)
conv2 = conv2d(drop,
filters,
kernel_size=(3, 3),
activation=tf.nn.relu,
name='middle1',
padding="SAME")
conv3 = conv2d(conv2,
filters,
kernel_size=(3, 3),
activation=tf.nn.relu,
name='middle2',
padding="SAME")
return conv3
def restoring_branch(net,
depth,
activation,
is_training,
init=he_init,
data_format='channels_first',
name=''):
net_ = L.conv2d_transpose(net,
depth, [2, 2],
strides=2,
padding='SAME',
data_format=data_format,
kernel_initializer=init,
name='{}_W1'.format(name))
net_ = activation(net_, name='{}_A1'.format(name))
net_ = batch_norm(net_, is_training)
net_ = L.conv2d(net_,
depth, [3, 3],
strides=1,
padding='SAME',
data_format=data_format,
kernel_initializer=init,
name='{}_W2'.format(name))
net_ = activation(net_, name='{}_A2'.format(name))
net_ = batch_norm(net_, is_training)
return net_
def generator(inp0, mask, dim, name, reuse):
# inp0 = tf.placeholder(tf.float32, [None, G.size, G.size, 3]) # G.size > 2 ** 5 + 1
with tf.variable_scope(name, reuse=reuse):
inp0 = tf.concat((inp0, mask), axis=3)
ten1 = conv_mask(inp0, mask, dim, 0)
ten2 = conv_mask(ten1, mask, dim, 1)
ten3 = conv_mask(ten2, mask, dim, 2)
ten4 = conv_mask(ten3, mask, dim, 3)
ten5 = conv_mask(ten4, mask, dim, 4)
ten6 = conv_tp_res(ten5, dim, 4)
ten6 = conv_tp_res(ten6, dim, 4)
ten6 = conv_tp_res(ten6, dim, 4)
ten6 = conv_tp_res(ten6, dim, 4)
ten5 = conv_tp_concat(ten6, ten5, dim, 5)
ten4 = conv_tp_concat(ten5, ten4, dim, 4)
ten3 = conv_tp_concat(ten4, ten3, dim, 3)
ten2 = conv_tp_concat(ten3, ten2, dim, 2)
ten1 = conv_tp_concat(ten2, ten1, dim, 1)
ten1 = tf.concat((ten1, inp0), axis=3)
ten1 = tl.conv2d_transpose(ten1, dim, 3, 1, 'valid', activation=leRU_batch_norm)
ten1 = tl.conv2d(ten1, 3, 3, 1, 'valid', activation=tf.nn.tanh)
ten1 = ten1 * 0.505 + 0.5
return ten1
def conv_tp(ten, dim, idx):
filters = (2**idx) * dim
return tl.conv2d_transpose(ten,
filters,
3,
2,
'same',
activation=leru_batch_norm)
def conv_tp_conv(ten, filters):
ten = tl.conv2d_transpose(ten,
filters,
3,
1,
'valid',
activation=leru_batch_norm)
ten = tl.conv2d(ten, filters, 3, 1, 'valid', activation=tf.nn.leaky_relu)
return ten
def inference(self, input):
conv = conv2d_transpose(input,
self.noutput,
kernel_size=(3, 3),
strides=(2, 2),
padding="SAME")
bn = batch_normalization(conv, self.training)
return tf.nn.relu(bn)
def UpSampler(input, num_outputs = None): ''' 上采样单元，使用转置卷积来使尺寸扩大两倍 ''' if not num_outputs: num_outputs = input.shape.as_list()[-1] net = conv2d_transpose(input, num_outputs, kernel_size = (3, 3), strides = (2, 2), padding = "SAME") return net
def deconv_net(self,input1,input2,filters,
kernel_size=[4,4],strides=[2,2],
activation=tf.nn.relu,name=None):
concat = self.crop_and_concat(input1,input2)
deconv2 = conv2d_transpose(inputs=concat, filters=filters,kernel_size=kernel_size, strides=strides,activation=activation ,data_format="channels_last", name=name)
# tensor = batch_normalization(tensor)
if name !=None : print("{}:{}".format(name,deconv2.shape))
return deconv2
def build_generator(noise, reuse=False):
with tf.variable_scope("generator") as scope:
if reuse:
tf.get_variable_scope().reuse_variables()
#
t = noise
t = dense(inputs=t, units=7 * 7 * 256, activation=my_leaky_relu)
t = batch_normalization(inputs=t)
t = tf.reshape(t, shape=[-1, 7, 7, 256])
#
t = conv2d_transpose(inputs=t,
filters=128,
kernel_size=[5, 5],
strides=1,
padding="same",
activation=my_leaky_relu)
t = batch_normalization(inputs=t)
t = conv2d_transpose(inputs=t,
filters=64,
kernel_size=[5, 5],
strides=2,
padding="same",
activation=my_leaky_relu)
t = batch_normalization(inputs=t)
t = conv2d_transpose(inputs=t,
filters=1,
kernel_size=[5, 5],
strides=2,
padding="same",
activation=tf.tanh)
image = t
#print("\nGen output image shape: {}".format(image.shape))
return image
def default_conv2d_transpose(inputs, filters):
return layers.conv2d_transpose(
inputs,
filters=filters,
kernel_size=4,
strides=(2, 2),
padding='same',
data_format='channels_last',
use_bias=False,
)
def generator(self, embed_input, noise_input, training=True, reuse=False):
def deconv2d(x,
filters,
kernels=(5, 5),
strides=(2, 2),
output_shape=None):
x = layers.conv2d_transpose(
x,
filters,
kernels,
strides,
padding='same',
kernel_initializer=tf.random_normal_initializer(mean=0,
stddev=0.02))
# x = layers.batch_normalization(x, training=training)
x = leaky_relu(x)
return x
s = self.image_size
s2, s4, s8, s16 = int(s / 2), int(s / 4), int(s / 8), int(s / 16)
with tf.variable_scope('Generator', reuse=reuse):
encoding_vec = layers.dense(
embed_input,
self.encode_dim,
name='embedding/dense',
kernel_initializer=tf.truncated_normal_initializer(
mean=0, stddev=0.02))
encoding_vec = leaky_relu(encoding_vec)
net = tf.concat([encoding_vec, noise_input], axis=-1)
net = layers.dense(
net,
self.gen_dim * 8 * s16 * s16,
name='concat/dense',
kernel_initializer=tf.truncated_normal_initializer(
mean=0, stddev=0.02))
# net = layers.batch_normalization(net)
net = leaky_relu(net)
net = tf.reshape(
net, [-1, s16, s16, self.gen_dim * 8]) # Conv1 (4 x 4 x 1024)
net = deconv2d(net, self.gen_dim * 4,
output_shape=[s8, s8]) # Conv1 (8 x 8 x 512)
net = deconv2d(net, self.gen_dim * 2,
output_shape=[s4, s4]) # Conv2 (16 x 16 x 256)
net = deconv2d(net, self.gen_dim,
output_shape=[s2, s2]) # Conv3 (32 x 32 x 128)
net = layers.conv2d_transpose(
net,
3,
kernel_size=(5, 5),
strides=(2, 2),
activation=tf.nn.tanh,
padding='same',
kernel_initializer=tf.truncated_normal_initializer(
mean=0, stddev=0.02))
return net
def create_generator(self, z, rate, is_training, reuse=None):
# hard code all layer params
upsample_power = 5 # 2**5=32
momentum = 0.99
with tf.variable_scope("generator", reuse=reuse):
# Since "mode collapse" issue, feature of FC needs to be small
# each layer come with dropout and batch_norm to improve the performance
x = fully_connected(z, 4 * 4 * 1, activation_fn=tf.nn.leaky_relu)
# x = dropout(x, rate = rate,training=is_training)
x = dropout(x, rate=rate)
x = tf.contrib.layers.batch_norm(x,
is_training=is_training,
decay=momentum)
x = tf.reshape(x, shape=(-1, 4, 4, 1))
for _ in range(upsample_power):
x = conv2d_transpose(x,
128, (5, 5),
strides=(2, 2),
padding='same',
data_format='channels_last',
activation=tf.nn.leaky_relu,
use_bias=True)
# x = dropout(x, rate = rate,training=is_training)
x = dropout(x, rate=rate)
x = tf.contrib.layers.batch_norm(x,
is_training=is_training,
decay=momentum)
x = conv2d_transpose(x,
1, (5, 5),
strides=(1, 1),
padding='same',
data_format='channels_last',
activation=tf.sigmoid,
use_bias=True)
# assert tf.shape(x) == (128,128,1)
# output is a matrix with -1 to 1
return x
def transpose_conv1d_layer(x,
n_out,
kernel_size,
stride=1,
activation=ACTIVATION,
regularize=True,
use_bias=True,
drop_rate=0.0,
batch_norm=BATCH_NORM,
training=True,
name=None,
reuse=None):
if batch_norm:
if name:
x = batch_norm_layer(x, training, name=name + '_bn', reuse=reuse)
else:
x = batch_norm_layer(x, training, name=name, reuse=reuse)
#wt_init = tf.truncated_normal_initializer(stddev=0.2)
#bi_init = tf.truncated_normal_initializer(mean=BIAS_SHIFT,stddev=0.01)
wt_init = None
bi_init = None
if regularize:
wt_reg = WT_REG
bi_reg = BI_REG
else:
wt_reg = None
bi_reg = None
x_tmp = tf.expand_dims(x, 3)
x_tmp = tf.transpose(x_tmp, (0, 1, 3, 2))
y_tmp = layers.conv2d_transpose(x_tmp,
n_out,
kernel_size=[kernel_size, 1],
strides=(stride, 1),
padding='same',
data_format='channels_last',
activation=activation,
use_bias=add_bias,
kernel_initializer=wt_init,
bias_initializer=bi_init,
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
trainable=True,
name=None,
reuse=None)
y = tf.reduce_sum(y_tmp, axis=2)
y = layers.dropout(y, rate=drop_rate, training=training)
return y
def __generator__(self, noise_input):
with tf.variable_scope("generator"):
noise_feature = layers.dense(noise_input, 4 * 4 * 256)
noise_feature = tf.nn.relu(noise_feature)
noise_feature = tf.reshape(noise_feature, shape=(-1, 4, 4, 256))
decnn1 = layers.conv2d_transpose(noise_feature,
filters=128,
kernel_size=5,
strides=(2, 2),
padding="SAME")
decnn1 = tf.nn.relu(decnn1)
decnn2 = layers.conv2d_transpose(decnn1,
filters=64,
kernel_size=5,
strides=(2, 2),
padding="SAME")
decnn2 = tf.nn.relu(decnn2)
if cfg.params["dataset"] == "mnist":
output = layers.conv2d_transpose(decnn2,
filters=1,
kernel_size=5,
strides=(2, 2),
padding="SAME")
elif cfg.params["dataset"] == "cifar":
output = layers.conv2d_transpose(decnn2,
filters=3,
kernel_size=5,
strides=(2, 2),
padding="SAME")
output = tf.nn.tanh(output)
#print_Arch:
print("Generator-Architecture")
print("input:{}".format(noise_input.shape))
print("project:{}".format(noise_feature.shape))
print("decnn1:{}".format(decnn1.shape))
print("decnn2:{}".format(decnn2.shape))
print("output:{}".format(output.shape))
return output
def generator(inp0, dim, name, reuse):
# inp0 = tf.placeholder(tf.float32, [None, G.size, G.size, 4]) # G.size == 2 ** 8
with tf.variable_scope(name, reuse=reuse):
ten1 = tl.conv2d(inp0, 1 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten2 = tl.conv2d(ten1, 2 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten3 = tl.conv2d(ten2, 4 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten4 = tl.conv2d(ten3, 8 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten5 = tl.conv2d(ten4, 16 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten6 = tl.conv2d(ten5, 32 * dim, 4, 2, 'same', activation=tf.nn.leaky_relu)
ten7 = tf.pad(ten6, paddings=tf.constant([(0, 0), (2, 2), (2, 2), (0, 0)]), mode='REFLECT')
ten7 = tl.conv2d(ten7, 32 * dim, 3, 1, 'valid', activation=leRU_batch_norm)
ten7 = tl.conv2d(ten7, 32 * dim, 3, 1, 'valid', activation=leRU_batch_norm)
ten7 = ten6 + ten7
ten6 = tl.conv2d_transpose(ten7, 64 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten5 = tl.conv2d_transpose(ten6, 32 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten4 = tl.conv2d_transpose(ten5, 16 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten3 = tl.conv2d_transpose(ten4, 8 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten2 = tl.conv2d_transpose(ten3, 4 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten1 = tl.conv2d_transpose(ten2, 2 * dim, 4, 2, 'same', activation=leRU_batch_norm)
ten1 = tl.conv2d(tf.concat((ten1, inp0), axis=3), 1 * dim, 3, 1, 'same', activation=leRU_batch_norm)
ten1 = tl.conv2d(ten1, 3, 3, 1, 'same', activation=tf.nn.tanh)
ten1 = ten1 * 0.505 + 0.5
return ten1
def _conv2d_transpose(input,
name,
filters=64,
kernel_size=4,
strides=2,
padding="same"):
with tf.variable_scope(name):
return conv2d_transpose(inputs=input,
name=name,
filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding)
def inference(self, input):
output = input
for layer in self.layers:
output = layer.inference(output)
output = conv2d_transpose(output,
self.num_classes,
kernel_size=(3, 3),
strides=2,
padding="SAME",
activation=None)
return output
def CONV(inputs, filters, kernel_size, strides, padding, is_transpose):
if is_transpose:
conv = conv2d_transpose(inputs=inputs,
filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding)
else:
conv = conv2d(inputs=inputs,
filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding)
return conv
