def LapSRNSingleLevel(net_image, net_feature, reuse=False):
with tf.variable_scope("Model_level", reuse=reuse):
tl.layers.set_name_reuse(reuse)
net_tmp = net_feature
# recursive block
for d in range(config.model.resblock_depth):
net_tmp = PReluLayer(net_tmp, name='prelu_D%s' % (d))
net_tmp = Conv2dLayer(
net_tmp,
shape=[3, 3, 64, 64],
strides=[1, 1, 1, 1],
name='conv_D%s' % (d),
W_init=tf.contrib.layers.xavier_initializer())
# for r in range(1,config.model.recursive_depth):
# for d in range(config.model.resblock_depth):
# net_tmp = PReluLayer(net_tmp, name='prelu_R%s_D%s'%(r,d))
# net_tmp = Conv2dLayer(net_tmp,shape=[3,3,64,64],strides=[1,1,1,1],
# name='conv_R%s_D%s'%(r,d), W_init=tf.contrib.layers.xavier_initializer())
net_feature = ElementwiseLayer(layer=[net_feature, net_tmp],
combine_fn=tf.add,
name='add_feature')
net_feature = PReluLayer(net_feature, name='prelu_feature')
net_feature = Conv2dLayer(
net_feature,
shape=[3, 3, 64, 256],
strides=[1, 1, 1, 1],
name='upconv_feature',
W_init=tf.contrib.layers.xavier_initializer())
net_feature = SubpixelConv2d(net_feature,
scale=2,
n_out_channel=64,
name='subpixel_feature')
# add image back
gradient_level = Conv2dLayer(
net_feature,
shape=[3, 3, 64, 3],
strides=[1, 1, 1, 1],
act=lrelu,
name='grad',
W_init=tf.contrib.layers.xavier_initializer())
net_image = Conv2dLayer(net_image,
shape=[3, 3, 3, 12],
strides=[1, 1, 1, 1],
name='upconv_image',
W_init=tf.contrib.layers.xavier_initializer())
net_image = SubpixelConv2d(net_image,
scale=2,
n_out_channel=3,
name='subpixel_image')
net_image = ElementwiseLayer(layer=[gradient_level, net_image],
combine_fn=tf.add,
name='add_image')
return net_image, net_feature, gradient_level
def build_classifier(self, im, inf_norm, reuse=False):
with tf.variable_scope('C', reuse=reuse) as vs:
tensorlayer.layers.set_name_reuse(reuse)
x = tf.reshape(im, [-1, 3, 32, 32])
x = tf.transpose(x, [0, 2, 3, 1])
xmin = tf.clip_by_value(x - inf_norm, 0., 1.)
xmax = tf.clip_by_value(x + inf_norm, 0., 1.)
x = tf.random_uniform(tf.shape(x), xmin, xmax, dtype=tf.float32)
# Crop the central [height, width] of the image.
# x = tf.image.resize_image_with_crop_or_pad(x, 24, 24)
x = tf.map_fn(lambda frame: tf.image.per_image_standardization(frame), x)
net = InputLayer(x)
net = Conv2dLayer(net, \
act=tf.nn.relu, \
shape=[5,5,3,64], \
name="conv1")
net = MaxPool2d(net, \
filter_size=(3,3), \
strides=(2,2), \
name="pool1")
net = LocalResponseNormLayer(net, \
depth_radius=4, \
bias=1.0, \
alpha = 0.001/9.0, \
beta = 0.75, \
name="norm1")
net = Conv2dLayer(net, \
act=tf.nn.relu, \
shape=[5,5,64,64], \
name="conv2")
net = LocalResponseNormLayer(net, \
depth_radius=4, \
bias=1.0, \
alpha=0.001/9.0, \
beta = 0.75, \
name="norm2")
net = MaxPool2d(net, \
filter_size=(3,3), \
strides=(2,2), \
name="pool2")
net = FlattenLayer(net, name="flatten_1")
net = DenseLayer(net, n_units=384, name="local3", act=tf.nn.relu)
net = DenseLayer(net, n_units=192, name="local4", act=tf.nn.relu)
net = DenseLayer(net, n_units=10, name="softmax_linear", act=tf.identity)
cla_vars = tf.contrib.framework.get_variables(vs)
def name_fixer(var):
return var.op.name.replace("W", "weights") \
.replace("b", "biases") \
.replace("weights_conv2d", "weights") \
.replace("biases_conv2d", "biases")
cla_vars = {name_fixer(var): var for var in cla_vars}
return net.outputs, cla_vars
def cifar10_classifier(im, reuse):
with tf.variable_scope('C', reuse=reuse) as vs:
net = InputLayer(im)
net = Conv2dLayer(net, \
act=tf.nn.relu, \
shape=[5,5,3,64], \
name="conv1")
net = MaxPool2d(net, \
filter_size=(3,3), \
strides=(2,2), \
name="pool1")
net = LocalResponseNormLayer(net, \
depth_radius=4, \
bias=1.0, \
alpha = 0.001/9.0, \
beta = 0.75, \
name="norm1")
net = Conv2dLayer(net, \
act=tf.nn.relu, \
shape=[5,5,64,64], \
name="conv2")
net = LocalResponseNormLayer(net, \
depth_radius=4, \
bias=1.0, \
alpha=0.001/9.0, \
beta = 0.75, \
name="norm2")
net = MaxPool2d(net, \
filter_size=(3,3), \
strides=(2,2), \
name="pool2")
net = FlattenLayer(net, name="flatten_1")
net = DenseLayer(net, n_units=384, name="local3", act=tf.nn.relu)
net = DenseLayer(net, n_units=192, name="local4", act=tf.nn.relu)
net = DenseLayer(net,
n_units=10,
name="softmax_linear",
act=tf.identity)
cla_vars = tf.contrib.framework.get_variables(vs)
def name_fixer(var):
return var.op.name.replace("W", "weights") \
.replace("b", "biases") \
.replace("weights_conv2d", "weights") \
.replace("biases_conv2d", "biases")
cla_vars = {name_fixer(var): var for var in cla_vars}
if not reuse:
return net.outputs, tf.argmax(net.outputs, axis=1), cla_vars
return net.outputs, tf.argmax(net.outputs, axis=1)
def build_classifier(self, im, inf_norm, reuse=False):
with tf.variable_scope("C", reuse=reuse) as vs:
x = tf.reshape(im, [-1, 64, 64, 3])
xmin = tf.clip_by_value(x - inf_norm, 0., 1.)
xmax = tf.clip_by_value(x + inf_norm, 0., 1.)
x = tf.random_uniform(tf.shape(x), xmin, xmax, dtype=tf.float32)
#x = tf.map_fn(lambda frame: tf.image.per_image_standardization(frame), x)
net = InputLayer(x)
n_filters = 3
for i in range(2):
net = Conv2dLayer(net, \
act=tf.nn.relu, \
shape=[5,5,n_filters,64], \
name="conv_" + str(i))
net = MaxPool2d(net, \
filter_size=(3,3), \
strides=(2,2), \
name="mpool_" + str(i))
net = LocalResponseNormLayer(net, \
depth_radius=4, \
bias=1.0, \
alpha=0.001 / 9.0, \
beta=0.75, \
name="lrn_" + str(i))
n_filters = 64
net = FlattenLayer(net)
net = DenseLayer(net, n_units=384, act=tf.nn.relu, name="d1")
net = DenseLayer(net, n_units=192, act=tf.nn.relu, name="d2")
net = DenseLayer(net, n_units=2, act=tf.identity, name="final")
cla_vars = tf.contrib.framework.get_variables(vs)
return net.outputs, cla_vars
def celebA_classifier(ims, reuse):
with tf.variable_scope("C", reuse=reuse) as vs:
net = InputLayer(ims)
n_filters = 3
for i in range(2):
net = Conv2dLayer(net, \
act=tf.nn.relu, \
shape=[5,5,n_filters,64], \
name="conv_" + str(i))
net = MaxPool2d(net, \
filter_size=(3,3), \
strides=(2,2), \
name="mpool_" + str(i))
net = LocalResponseNormLayer(net, \
depth_radius=4, \
bias=1.0, \
alpha=0.001 / 9.0, \
beta=0.75, \
name="lrn_" + str(i))
n_filters = 64
net = FlattenLayer(net)
net = DenseLayer(net, n_units=384, act=tf.nn.relu, name="d1")
net = DenseLayer(net, n_units=192, act=tf.nn.relu, name="d2")
net = DenseLayer(net, n_units=2, act=tf.identity, name="final")
cla_vars = tf.contrib.framework.get_variables(vs)
if not reuse:
return net.outputs, tf.argmax(net.outputs, axis=1), cla_vars
return net.outputs, tf.argmax(net.outputs, axis=1)
def conv2d(layer, out_channels, filter_size=3, stride=1, act=tf.identity, padding='SAME', W_init=w_init, b_init=b_init, name='conv2d'):
in_channels = layer.outputs.get_shape().as_list()[-1]
return Conv2dLayer(layer = layer,
act = act,
shape = [filter_size, filter_size, in_channels, out_channels],
strides=[1, stride, stride, 1],
padding=padding,
W_init = w_init,
b_init = b_init,
name =name)
def LapSRN(inputs, is_train=False, reuse=False):
n_level = int(np.log2(config.model.scale))
if not n_level >= 1:
raise ValueError('n_level >=1 expected but not satisfied')
with tf.variable_scope("LapSRN", reuse=reuse) as vs:
tl.layers.set_name_reuse(reuse)
shapes = tf.shape(inputs)
inputs_level = InputLayer(inputs, name='input_level')
net_feature = Conv2dLayer(
inputs_level,
shape=[3, 3, 3, 64],
strides=[1, 1, 1, 1],
W_init=tf.contrib.layers.xavier_initializer(),
name='init_conv')
net_image = inputs_level
net_image1, net_feature1, net_gradient1 = LapSRNSingleLevel(
net_image, net_feature, reuse=reuse)
net_image2, net_feature2, net_gradient2 = LapSRNSingleLevel(
net_image1, net_feature1, reuse=True)
return net_image2, net_gradient2, net_image1, net_gradient1
def conv_layers(net_in):
with tf.name_scope('preprocess'):
# Notice that we include a preprocessing layer that takes the RGB image
# with pixels values in the range of 0-255 and subtracts the mean image
# values (calculated over the entire ImageNet training set).
mean = tf.constant([123.68, 116.779, 103.939],
dtype=tf.float32,
shape=[1, 1, 1, 3],
name='img_mean')
net_in.outputs = net_in.outputs - mean
# conv1
net = Conv2dLayer(
net_in,
act=tf.nn.relu,
shape=[3, 3, 3, 64], # 64 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv1_1')
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 64, 64], # 64 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv1_2')
net = PoolLayer(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name='pool1')
# conv2
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 64, 128], # 128 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv2_1')
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 128, 128], # 128 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv2_2')
net = PoolLayer(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name='pool2')
# conv3
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 128, 256], # 256 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv3_1')
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 256, 256], # 256 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv3_2')
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 256, 256], # 256 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv3_3')
net = PoolLayer(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name='pool3')
# conv4
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 256, 512], # 512 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv4_1')
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 512, 512], # 512 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv4_2')
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 512, 512], # 512 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv4_3')
net = PoolLayer(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name='pool4')
# conv5
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 512, 512], # 512 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv5_1')
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 512, 512], # 512 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv5_2')
net = Conv2dLayer(
net,
act=tf.nn.relu,
shape=[3, 3, 512, 512], # 512 features for each 3x3 patch
strides=[1, 1, 1, 1],
padding='SAME',
name='conv5_3')
net = PoolLayer(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name='pool5')
return net
def _build_arch(self, net_in):
network = Conv2dLayer(net_in,
act=tf.nn.relu,
shape=[3, 3, 3, 64],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv1_1')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 64, 64],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv1_2')
network = PoolLayer(network,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name=self.name + '_pool1')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 64, 128],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv2_1')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 128, 128],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv2_2')
network = PoolLayer(network,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name=self.name + '_pool2')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 128, 256],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv3_1')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 256, 256],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv3_2')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 256, 256],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv3_3')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 256, 256],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv3_4')
network = PoolLayer(network,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name=self.name + '_pool3')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 256, 512],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv4_1')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 512, 512],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv4_2')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 512, 512],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv4_3')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 512, 512],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv4_4')
network = PoolLayer(network,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name=self.name + '_pool4')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 512, 512],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv5_1')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 512, 512],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv5_2')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 512, 512],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv5_3')
network = Conv2dLayer(network,
act=tf.nn.relu,
shape=[3, 3, 512, 512],
strides=[1, 1, 1, 1],
padding='SAME',
name=self.name + '_conv5_4')
network = PoolLayer(network,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
pool=tf.nn.max_pool,
name=self.name + '_pool5')
network = FlattenLayer(network, name=self.name + '_flatten')
network = DenseLayer(network,
n_units=4096,
act=tf.nn.relu,
name=self.name + '_fc6')
network = DenseLayer(network,
n_units=4096,
act=tf.nn.relu,
name=self.name + '_fc7')
network = DenseLayer(network,
n_units=1000,
act=tf.identity,
name=self.name + '_fc8')
return network
