def build_activator(self, input_tensor, features: int, activator="", leaky_relu_alpha=0.1, base_name=""):
features = int(features)
if activator is None or "":
return
elif activator == "relu":
output = tf.nn.relu(input_tensor, name=base_name + "_relu")
elif activator == "sigmoid":
output = tf.nn.sigmoid(input_tensor, name=base_name + "_sigmoid")
elif activator == "tanh":
output = tf.nn.tanh(input_tensor, name=base_name + "_tanh")
elif activator == "leaky_relu":
output = tf.maximum(input_tensor, leaky_relu_alpha * input_tensor, name=base_name + "_leaky")
elif activator == "prelu":
with tf.variable_scope("prelu"):
alphas = tf.Variable(tf.constant(0.1, shape=[features]), name=base_name + "_prelu")
if self.save_weights:
util.add_summaries("prelu_alpha", self.name, alphas, save_stddev=False, save_mean=False)
output = tf.nn.relu(input_tensor) + tf.multiply(alphas, (input_tensor - tf.abs(input_tensor))) * 0.5
elif activator == "selu":
output = tf.nn.selu(input_tensor, name=base_name + "_selu")
else:
raise NameError('Not implemented activator:%s' % activator)
self.complexity += (self.pix_per_input * features)
return output
def build_conv(self, name, input_tensor, cnn_size, input_feature_num,
output_feature_num):
with tf.variable_scope(name):
w = util.weight(
[cnn_size, cnn_size, input_feature_num, output_feature_num],
stddev=self.weight_dev,
name="conv_W",
initializer=self.initializer)
h = self.conv2d(input_tensor,
w,
self.cnn_stride,
bias=None,
activator=None,
name=name)
if self.save_weights:
util.add_summaries("weight",
self.name,
w,
save_stddev=True,
save_mean=True)
if self.save_images and cnn_size > 1 and input_feature_num == 1:
weight_transposed = tf.transpose(w, [3, 0, 1, 2])
with tf.name_scope("image"):
tf.summary.image(self.name,
weight_transposed,
max_outputs=self.log_weight_image_num)
return w, h
def add_optimizer_op(self, loss, lr_input):
if self.optimizer == "gd":
optimizer = tf.train.GradientDescentOptimizer(lr_input)
elif self.optimizer == "adadelta":
optimizer = tf.train.AdadeltaOptimizer(lr_input)
elif self.optimizer == "adagrad":
optimizer = tf.train.AdagradOptimizer(lr_input)
elif self.optimizer == "adam":
optimizer = tf.train.AdamOptimizer(lr_input, beta1=self.beta1, beta2=self.beta2, epsilon=self.epsilon)
elif self.optimizer == "momentum":
optimizer = tf.train.MomentumOptimizer(lr_input, self.momentum)
elif self.optimizer == "rmsprop":
optimizer = tf.train.RMSPropOptimizer(lr_input, momentum=self.momentum)
else:
print("Optimizer arg should be one of [gd, adadelta, adagrad, adam, momentum, rmsprop].")
return None
if self.clipping_norm > 0 or self.save_weights:
trainables = tf.trainable_variables()
grads = tf.gradients(loss, trainables)
if self.save_weights:
for i in range(len(grads)):
util.add_summaries("", self.name, grads[i], header_name=grads[i].name + "/", save_stddev=True,
save_mean=True)
if self.clipping_norm > 0:
clipped_grads, _ = tf.clip_by_global_norm(grads, clip_norm=self.clipping_norm)
grad_var_pairs = zip(clipped_grads, trainables)
training_optimizer = optimizer.apply_gradients(grad_var_pairs)
else:
training_optimizer = optimizer.minimize(loss)
return training_optimizer
def conv2d(self, x, w, stride, bias=None, activator=None, leaky_relu_alpha=0.1, name=""):
conv = tf.nn.conv2d(x, w, strides=[stride, stride, 1, 1], padding="SAME", name=name + "_conv")
self.complexity += int(w.shape[0] * w.shape[1] * w.shape[2] * w.shape[3])
if bias is not None:
conv = tf.add(conv, bias, name=name + "_add")
self.complexity += int(bias.shape[0])
if activator is not None:
if activator == "relu":
conv = tf.nn.relu(conv, name=name + "_relu")
elif activator == "sigmoid":
conv = tf.nn.sigmoid(conv, name=name + "_sigmoid")
elif activator == "tanh":
conv = tf.nn.tanh(conv, name=name + "_tanh")
elif activator == "leaky_relu":
conv = tf.maximum(conv, leaky_relu_alpha * conv, name=name + "_leaky")
elif activator == "prelu":
with tf.variable_scope("prelu"):
alphas = tf.Variable(tf.constant(0.1, shape=[w.get_shape()[3]]), name=name + "_prelu")
if self.save_weights:
util.add_summaries("prelu_alpha", self.name, alphas, save_stddev=False, save_mean=False)
conv = tf.nn.relu(conv) + tf.multiply(alphas, (conv - tf.abs(conv))) * 0.5
else:
raise NameError('Not implemented activator:%s' % activator)
self.complexity += int(bias.shape[0])
return conv
def build_conv_and_bias(self,
name,
input_tensor,
cnn_size,
input_feature_num,
output_feature_num,
use_activator=True,
use_dropout=True):
with tf.variable_scope(name):
w = util.weight(
[cnn_size, cnn_size, input_feature_num, output_feature_num],
stddev=self.weight_dev,
name="conv_W",
initializer=self.initializer)
b = util.bias([output_feature_num], name="conv_B")
h = self.conv2d(
input_tensor,
w,
self.cnn_stride,
bias=b,
activator=self.activator if use_activator else None,
name=name)
if use_dropout and self.dropout != 1.0:
h = tf.nn.dropout(h, self.dropout_input, name="dropout")
if self.save_weights:
util.add_summaries("weight",
self.name,
w,
save_stddev=True,
save_mean=True)
util.add_summaries("bias",
self.name,
b,
save_stddev=True,
save_mean=True)
if self.save_images and cnn_size > 1 and input_feature_num == 1:
weight_transposed = tf.transpose(w, [3, 0, 1, 2])
with tf.name_scope("image"):
tf.summary.image(self.name,
weight_transposed,
max_outputs=self.log_weight_image_num)
return w, b, h
def build_conv(self, name, input_tensor, cnn_size, input_feature_num, output_feature_num, use_bias=False,
activator=None, use_batch_norm=False, dropout_rate=1.0):
with tf.variable_scope(name):
w = util.weight([cnn_size, cnn_size, input_feature_num, output_feature_num],
stddev=self.weight_dev, name="conv_W", initializer=self.initializer)
b = util.bias([output_feature_num], name="conv_B") if use_bias else None
h = self.conv2d(input_tensor, w, self.cnn_stride, bias=b, use_batch_norm=use_batch_norm, name=name)
if activator is not None:
h = self.build_activator(h, output_feature_num, activator, base_name=name)
if dropout_rate < 1.0:
h = tf.nn.dropout(h, self.dropout, name="dropout")
self.H.append(h)
if self.save_weights:
util.add_summaries("weight", self.name, w, save_stddev=True, save_mean=True)
util.add_summaries("output", self.name, h, save_stddev=True, save_mean=True)
if use_bias:
util.add_summaries("bias", self.name, b, save_stddev=True, save_mean=True)
# todo check
if self.save_images and cnn_size > 1 and input_feature_num == 1:
weight_transposed = tf.transpose(w, [3, 0, 1, 2])
with tf.name_scope("image"):
tf.summary.image(self.name, weight_transposed, max_outputs=self.log_weight_image_num)
if self.receptive_fields == 0:
self.receptive_fields = cnn_size
else:
self.receptive_fields += (cnn_size - 1)
self.features += "%d " % output_feature_num
self.Weights.append(w)
if use_bias:
self.Biases.append(b)
return h
def build_graph(self):
self.x = tf.placeholder(tf.float32,
shape=[None, None, None, self.channels],
name="x")
self.y = tf.placeholder(tf.float32,
shape=[None, None, None, self.output_channels],
name="y")
self.x2 = tf.placeholder(
tf.float32,
shape=[None, None, None, self.output_channels],
name="x2")
self.dropout = tf.placeholder(tf.float32,
shape=[],
name="dropout_keep_rate")
self.is_training = tf.placeholder(tf.bool, name="is_training")
# building feature extraction layers
output_feature_num = self.filters
total_output_feature_num = 0
input_feature_num = self.channels
input_tensor = self.x
if self.save_weights:
with tf.name_scope("X"):
util.add_summaries("output",
self.name,
self.x,
save_stddev=True,
save_mean=True)
for i in range(self.layers):
if self.min_filters != 0 and i > 0:
x1 = i / float(self.layers - 1)
y1 = pow(x1, 1.0 / self.filters_decay_gamma)
output_feature_num = int((self.filters - self.min_filters) *
(1 - y1) + self.min_filters)
self.build_conv("CNN%d" % (i + 1),
input_tensor,
self.cnn_size,
input_feature_num,
output_feature_num,
use_bias=True,
activator=self.activator,
use_batch_norm=self.batch_norm,
dropout_rate=self.dropout_rate)
input_feature_num = output_feature_num
input_tensor = self.H[-1]
total_output_feature_num += output_feature_num
with tf.variable_scope("Concat"):
self.H_concat = tf.concat(self.H, 3, name="H_concat")
self.features += " Total: (%d)" % total_output_feature_num
# building reconstruction layers ---
if self.use_nin:
self.build_conv("A1",
self.H_concat,
1,
total_output_feature_num,
self.nin_filters,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
self.receptive_fields -= (self.cnn_size - 1)
self.build_conv("B1",
self.H_concat,
1,
total_output_feature_num,
self.nin_filters2,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
self.build_conv("B2",
self.H[-1],
3,
self.nin_filters2,
self.nin_filters2,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
self.H.append(
tf.concat([self.H[-1], self.H[-3]], 3, name="Concat2"))
input_channels = self.nin_filters + self.nin_filters2
else:
self.H.append(self.H_concat)
input_channels = total_output_feature_num
# building upsampling layer
if self.pixel_shuffler:
if self.pixel_shuffler_filters != 0:
output_channels = self.pixel_shuffler_filters
else:
output_channels = input_channels
if self.scale == 4:
self.build_pixel_shuffler_layer("Up-PS", self.H[-1], 2,
input_channels, input_channels)
self.build_pixel_shuffler_layer("Up-PS2", self.H[-1], 2,
input_channels,
output_channels)
else:
self.build_pixel_shuffler_layer("Up-PS", self.H[-1],
self.scale, input_channels,
output_channels)
input_channels = output_channels
else:
self.build_transposed_conv("Up-TCNN", self.H[-1], self.scale,
input_channels)
for i in range(self.reconstruct_layers - 1):
self.build_conv("R-CNN%d" % (i + 1),
self.H[-1],
self.cnn_size,
input_channels,
self.reconstruct_filters,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
input_channels = self.reconstruct_filters
self.build_conv("R-CNN%d" % self.reconstruct_layers, self.H[-1],
self.cnn_size, input_channels, self.output_channels)
self.y_ = tf.add(self.H[-1], self.x2, name="output")
if self.save_weights:
with tf.name_scope("Y_"):
util.add_summaries("output",
self.name,
self.y_,
save_stddev=True,
save_mean=True)
logging.info("Feature:%s Complexity:%s Receptive Fields:%d" %
(self.features, "{:,}".format(
self.complexity), self.receptive_fields))
def build_graph(self):
inference = False
# first dimension is always 1 for some reason. Second dimension is image height. Third dimension is image width. Fourth dimension is the number of color channels.
if (self.input_image_height > 0 and self.input_image_width > 0):
if (inference):
self.x = tf.placeholder(tf.float32,
shape=[
1, self.input_image_height,
self.input_image_width,
self.channels
],
name="x")
self.y = tf.placeholder(
tf.float32,
shape=[
1, self.input_image_height * self.scale,
self.input_image_width * self.scale,
self.output_channels
],
name="y")
self.x2 = tf.placeholder(
tf.float32,
shape=[
1, self.input_image_height * self.scale,
self.input_image_width * self.scale,
self.output_channels
],
name="x2")
else:
self.x = tf.placeholder(tf.float32,
shape=[
None, self.input_image_height,
self.input_image_width,
self.channels
],
name="x")
self.y = tf.placeholder(
tf.float32,
shape=[
None, self.input_image_height * self.scale,
self.input_image_width * self.scale,
self.output_channels
],
name="y")
self.x2 = tf.placeholder(
tf.float32,
shape=[
None, self.input_image_height * self.scale,
self.input_image_width * self.scale,
self.output_channels
],
name="x2")
else:
self.x = tf.placeholder(tf.float32,
shape=[None, None, None, self.channels],
name="x")
self.y = tf.placeholder(
tf.float32,
shape=[None, None, None, self.output_channels],
name="y")
self.x2 = tf.placeholder(
tf.float32,
shape=[None, None, None, self.output_channels],
name="x2")
self.dropout = tf.placeholder(tf.float32,
shape=[],
name="dropout_keep_rate")
self.is_training = tf.placeholder(tf.bool, name="is_training")
# building feature extraction layers
output_feature_num = self.filters
total_output_feature_num = 0
input_feature_num = self.channels
input_tensor = self.x
if self.save_weights:
with tf.name_scope("X"):
util.add_summaries("output",
self.name,
self.x,
save_stddev=True,
save_mean=True)
if (self.depthwise_seperable):
for i in range(self.layers):
if (i > 0):
x1 = i / float(self.layers - 1)
y1 = pow(x1, 1.0 / self.filters_decay_gamma)
output_feature_num = int(
(self.filters - self.min_filters) * (1 - y1) +
self.min_filters)
# self.build_conv("Bottleneck%d" % (i + 1), input_tensor, 1, input_feature_num,
# 1, use_bias=True, activator=None,
# use_batch_norm=self.batch_norm, dropout_rate=self.dropout_rate)
# input_tensor = self.H[-1]
# self.build_depthwise_seperable_conv("CNN%d" % (i + 1), input_tensor, self.cnn_size, 1,
# output_feature_num, use_bias=True, activator=self.activator,
# use_batch_norm=self.batch_norm, dropout_rate=self.dropout_rate)
self.build_depthwise_seperable_conv(
"CNN%d" % (i + 1),
input_tensor,
self.cnn_size,
input_feature_num,
output_feature_num,
use_bias=True,
activator=self.activator,
use_batch_norm=self.batch_norm,
dropout_rate=self.dropout_rate)
else:
self.build_conv("CNN%d" % (i + 1),
input_tensor,
self.cnn_size,
input_feature_num,
output_feature_num,
use_bias=True,
activator=self.activator,
use_batch_norm=self.batch_norm,
dropout_rate=self.dropout_rate)
input_feature_num = output_feature_num
input_tensor = self.H[-1]
total_output_feature_num += output_feature_num
else:
# original version
for i in range(self.layers):
if self.min_filters != 0 and i > 0:
x1 = i / float(self.layers - 1)
y1 = pow(x1, 1.0 / self.filters_decay_gamma)
output_feature_num = int(
(self.filters - self.min_filters) * (1 - y1) +
self.min_filters)
self.build_conv("CNN%d" % (i + 1),
input_tensor,
self.cnn_size,
input_feature_num,
output_feature_num,
use_bias=True,
activator=self.activator,
use_batch_norm=self.batch_norm,
dropout_rate=self.dropout_rate)
input_feature_num = output_feature_num
input_tensor = self.H[-1]
total_output_feature_num += output_feature_num
with tf.variable_scope("Concat"):
self.H_concat = tf.concat(self.H, 3, name="H_concat")
self.features += " Total: (%d)" % total_output_feature_num
# building reconstruction layers ---
if self.use_nin:
self.build_conv("A1",
self.H_concat,
1,
total_output_feature_num,
self.nin_filters,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
self.receptive_fields -= (self.cnn_size - 1)
self.build_conv("B1",
self.H_concat,
1,
total_output_feature_num,
self.nin_filters2,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
if (self.depthwise_seperable):
self.build_depthwise_seperable_conv(
"B2",
self.H[-1],
3,
self.nin_filters2,
self.nin_filters2,
use_bias=True,
activator=self.activator,
dropout_rate=self.dropout_rate)
else:
self.build_conv("B2",
self.H[-1],
3,
self.nin_filters2,
self.nin_filters2,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
self.H.append(
tf.concat([self.H[-1], self.H[-3]], 3, name="Concat2"))
input_channels = self.nin_filters + self.nin_filters2
else:
if (self.depthwise_seperable):
self.build_depthwise_seperable_conv(
"C",
self.H_concat,
1,
total_output_feature_num,
self.filters,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
else:
self.build_conv("C",
self.H_concat,
1,
total_output_feature_num,
self.filters,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
input_channels = self.filters
# building upsampling layer
if self.pixel_shuffler:
if self.pixel_shuffler_filters != 0:
output_channels = self.pixel_shuffler_filters
else:
output_channels = input_channels
if self.scale == 4:
if (self.bottleneck):
self.build_conv("Up-Bottleneck",
self.H[-1],
1,
input_channels,
output_channels,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
self.build_pixel_shuffler_layer(
"Up-PS",
self.H[-1],
2,
output_channels,
input_channels,
depthwise_seperable=self.depthwise_seperable)
self.build_conv("Up-Bottleneck2",
self.H[-1],
1,
input_channels,
output_channels,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
self.build_pixel_shuffler_layer(
"Up-PS2",
self.H[-1],
2,
output_channels,
output_channels,
depthwise_seperable=self.depthwise_seperable)
else:
self.build_pixel_shuffler_layer(
"Up-PS",
self.H[-1],
2,
input_channels,
input_channels,
depthwise_seperable=self.depthwise_seperable)
self.build_pixel_shuffler_layer(
"Up-PS2",
self.H[-1],
2,
input_channels,
output_channels,
depthwise_seperable=self.depthwise_seperable)
else:
self.build_pixel_shuffler_layer(
"Up-PS",
self.H[-1],
self.scale,
input_channels,
output_channels,
depthwise_seperable=self.depthwise_seperable)
input_channels = output_channels
else:
self.build_transposed_conv("Up-TCNN", self.H[-1], self.scale,
input_channels)
for i in range(self.reconstruct_layers - 1):
self.build_conv("R-CNN%d" % (i + 1),
self.H[-1],
self.cnn_size,
input_channels,
self.reconstruct_filters,
dropout_rate=self.dropout_rate,
use_bias=True,
activator=self.activator)
input_channels = self.reconstruct_filters
if (self.depthwise_seperable):
self.build_depthwise_seperable_conv(
"R-CNN%d" % self.reconstruct_layers, self.H[-1], self.cnn_size,
input_channels, self.output_channels)
else:
self.build_conv("R-CNN%d" % self.reconstruct_layers, self.H[-1],
self.cnn_size, input_channels,
self.output_channels)
self.y_ = tf.add(self.H[-1], self.x2, name="output")
if self.save_weights:
with tf.name_scope("Y_"):
util.add_summaries("output",
self.name,
self.y_,
save_stddev=True,
save_mean=True)
logging.info("Feature:%s Complexity:%s Receptive Fields:%d" %
(self.features, "{:,}".format(
self.complexity), self.receptive_fields))
