def conv2d_residual_block(network, lastlayer, model_layer):
name = model_layer['name']
filter_size = model_layer['filter_size']
is_training = model_layer['batch_norm']
if 'function' in model_layer:
activation = model_layer['function']
else:
activation = 'relu'
# original residual unit
shape = network[lastlayer].get_output_shape()
num_filters = shape[-1].value
if not isinstance(filter_size, (list, tuple)):
filter_size = (filter_size, 1)
network[name + '_1resid'] = layers.Conv2DLayer(network[lastlayer],
num_filters=num_filters,
filter_size=filter_size,
padding='SAME')
network[name + '_1resid_norm'] = layers.BatchNormLayer(
network[name + '_1resid'], is_training)
network[name + '_1resid_active'] = layers.ActivationLayer(
network[name + '_1resid_norm'], function=activation)
if 'dropout_block' in model_layer:
network[name + '_dropout1'] = layers.DropoutLayer(
network[name + '_1resid_active'],
keep_prob=model_layer['dropout_block'])
lastname = name + '_dropout1'
else:
lastname = name + '_1resid_active'
network[name + '_2resid'] = layers.Conv2DLayer(network[lastname],
num_filters=num_filters,
filter_size=filter_size,
padding='SAME')
network[name + '_2resid_norm'] = layers.BatchNormLayer(
network[name + '_2resid'], is_training)
network[name + '_resid_sum'] = layers.ElementwiseSumLayer(
[network[lastlayer], network[name + '_2resid_norm']])
network[name + '_resid'] = layers.ActivationLayer(network[name +
'_resid_sum'],
function=activation)
return network
def conv1d_residual_block(self, model_layer):
lastlayer = self.lastlayer
name = model_layer['name']
filter_size = model_layer['filter_size']
is_training = model_layer['batch_norm']
if 'function' in model_layer:
activation = model_layer['function']
else:
activation = 'relu'
# original residual unit
shape = self.network[lastlayer].get_output_shape()
num_filters = shape[-1].value
if not isinstance(filter_size, (list, tuple)):
filter_size = (filter_size, 1)
self.network[name+'_1resid'] = layers.Conv2DLayer(self.network[lastlayer], num_filters=num_filters, filter_size=filter_size, padding='SAME')
self.network[name+'_1resid_norm'] = layers.BatchNormLayer(self.network[name+'_1resid'], is_training)
self.network[name+'_1resid_active'] = layers.ActivationLayer(self.network[name+'_1resid_norm'], function=activation)
if 'dropout_block' in model_layer:
dropout = model_layer['dropout_block']
placeholder_name = 'keep_prob'+str(len(self.num_dropout))
exec(placeholder_name+" = tf.placeholder(tf.float32, name='"+placeholder_name+"')")
#exec("self.placeholders["+placeholder_name+"] = " + placeholder_name)
exec("self.network["+name+"+'_dropout1'] = layers.DropoutLayer(self.network["+name+"+'_1resid_active'], keep_prob="+placeholder_name+")")
exec("self.hidden_feed_dict["+placeholder_name+"] ="+str(dropout))
self.num_dropout += 1
lastname = name+'_dropout1'
else:
lastname = name+'_1resid_active'
self.network[name+'_2resid'] = layers.Conv2DLayer(self.network[lastname], num_filters=num_filters, filter_size=filter_size, padding='SAME')
self.network[name+'_2resid_norm'] = layers.BatchNormLayer(self.network[name+'_2resid'], is_training)
self.network[name+'_resid_sum'] = layers.ElementwiseSumLayer([self.network[lastlayer], self.network[name+'_2resid_norm']])
self.network[name+'_resid'] = layers.ActivationLayer(self.network[name+'_resid_sum'], function=activation)
self.lastlayer = name+'_resid'
return network
def dense_residual_block(network, lastlayer, model_layer):
name = model_layer['name']
is_training = model_layer['batch_norm']
if 'function' in model_layer:
activation = model_layer['function']
else:
activation = 'relu'
# original residual unit
shape = network[lastlayer].get_output_shape()
num_units = shape[-1].value
network[name + '_1resid'] = layers.DenseLayer(network[lastlayer],
num_units=num_units,
b=None)
network[name + '_1resid_norm'] = layers.BatchNormLayer(
network[name + '_1resid'], is_training)
network[name + '_1resid_active'] = layers.ActivationLayer(
network[name + '_1resid_norm'], function=activation)
if 'dropout_block' in model_layer:
network[name + '_dropout1'] = layers.DropoutLayer(
network[name + '_1resid_active'],
keep_prob=model_layer['dropout_block'])
else:
lastname = name + '_1resid_active'
network[name + '_2resid'] = layers.DenseLayer(network[lastname],
num_units=num_units,
b=None)
network[name + '_2resid_norm'] = layers.BatchNormLayer(
network[name + '_2resid'], is_training)
network[name + '_resid_sum'] = layers.ElementwiseSumLayer(
[network[lastlayer], network[name + '_2resid_norm']])
network[name + '_resid'] = layers.ActivationLayer(network[name +
'_resid_sum'],
function=activation)
return network
def dense_residual_block(self, model_layer):
lastlayer = self.lastlayer
name = model_layer['name']
is_training = model_layer['batch_norm']
if 'function' in model_layer:
activation = model_layer['function']
else:
activation = 'relu'
# original residual unit
shape = self.network[lastlayer].get_output_shape()
num_units = shape[-1].value
self.network[name+'_1resid'] = layers.DenseLayer(self.network[lastlayer], num_units=num_units, b=None)
self.network[name+'_1resid_norm'] = layers.BatchNormLayer(self.network[name+'_1resid'], is_training)
self.network[name+'_1resid_active'] = layers.ActivationLayer(self.network[name+'_1resid_norm'], function=activation)
if 'dropout_block' in model_layer:
dropout = model_layer['dropout_block']
placeholder_name = 'keep_prob'+str(len(self.num_dropout))
exec(placeholder_name+" = tf.placeholder(tf.float32, name='"+placeholder_name+"')")
#exec("self.placeholders["+placeholder_name+"] = " + placeholder_name)
exec("self.network["+name+"+'_dropout1'] = layers.DropoutLayer(self.network["+name+"+'_1resid_active'], keep_prob="+placeholder_name+")")
exec("self.hidden_feed_dict["+placeholder_name+"] ="+str(dropout))
lastname = name+'_dropout1'
self.num_dropout += 1
else:
lastname = name+'_1resid_active'
self.network[name+'_2resid'] = layers.DenseLayer(self.network[lastname], num_units=num_units, b=None)
self.network[name+'_2resid_norm'] = layers.BatchNormLayer(self.network[name+'_2resid'], is_training)
self.network[name+'_resid_sum'] = layers.ElementwiseSumLayer([self.network[lastlayer], self.network[name+'_2resid_norm']])
self.network[name+'_resid'] = layers.ActivationLayer(self.network[name+'_resid_sum'], function=activation)
return network
def build_layers(model_layers, network=OrderedDict()):
# loop to build each layer of network
lastlayer = ''
for model_layer in model_layers:
layer = model_layer['layer']
name = model_layer['name']
if layer == "input":
# add input layer
network[name] = single_layer(model_layer, network)
lastlayer = name
else:
if layer == 'residual-conv1d':
network = conv1d_residual_block(network, lastlayer,
model_layer)
lastlayer = name + '_resid'
elif layer == 'residual-conv2d':
network = conv2d_residual_block(network, lastlayer,
model_layer)
lastlayer = name + '_resid'
elif layer == 'residual-dense':
network = dense_residual_block(network, lastlayer, model_layer)
lastlayer = name + '_resid'
else:
# add core layer
newlayer = name #'# str(counter) + '_' + name + '_batch'
network[newlayer] = single_layer(model_layer,
network[lastlayer])
lastlayer = newlayer
# add bias layer
if 'b' in model_layer:
newlayer = name + '_bias'
network[newlayer] = layers.BiasLayer(network[lastlayer],
b=model_layer['b'])
lastlayer = newlayer
# add Batch normalization layer
if 'batch_norm' in model_layer:
newlayer = name + '_batch' #str(counter) + '_' + name + '_batch'
network[newlayer] = layers.BatchNormLayer(
network[lastlayer], model_layer['batch_norm'])
lastlayer = newlayer
# add activation layer
if 'activation' in model_layer:
newlayer = name + '_active'
network[newlayer] = layers.ActivationLayer(
network[lastlayer], function=model_layer['activation'])
lastlayer = newlayer
# add max-pooling layer
if 'pool_size' in model_layer:
newlayer = name + '_pool' # str(counter) + '_' + name+'_pool'
if isinstance(model_layer['pool_size'], (tuple, list)):
network[newlayer] = layers.MaxPool2DLayer(
network[lastlayer], pool_size=model_layer['pool_size'])
else:
network[newlayer] = layers.MaxPool2DLayer(
network[lastlayer],
pool_size=(model_layer['pool_size'], 1))
lastlayer = newlayer
# add dropout layer
if 'dropout' in model_layer:
newlayer = name + '_dropout' # str(counter) + '_' + name+'_dropout'
network[newlayer] = layers.DropoutLayer(
network[lastlayer], keep_prob=model_layer['dropout'])
lastlayer = newlayer
return network, lastlayer
def build_layers(self):
# loop to build each layer of network
for model_layer in self.model_layers:
layer = model_layer['layer']
name = model_layer['name']
if layer == "input":
# add input layer
self.single_layer(model_layer)
else:
if layer == 'residual-conv1d':
conv1d_residual_block(model_layer)
elif layer == 'residual-conv2d':
conv2d_residual_block(model_layer)
elif layer == 'residual-dense':
dense_residual_block(model_layer)
else:
# add core layer
self.single_layer(model_layer)
# add bias layer
if 'b' in model_layer:
newlayer = name+'_bias'
self.network[newlayer] = layers.BiasLayer(self.network[self.lastlayer], b=model_layer['b'])
self.lastlayer = newlayer
# add Batch normalization layer
if 'norm' in model_layer:
if 'batch' in model_layer['norm']:
newlayer = name + '_batch' #str(counter) + '_' + name + '_batch'
self.network[newlayer] = layers.BatchNormLayer(self.network[self.lastlayer], self.is_training)
self.lastlayer = newlayer
# add activation layer
if 'activation' in model_layer:
newlayer = name+'_active'
self.network[newlayer] = layers.ActivationLayer(self.network[self.lastlayer], function=model_layer['activation'])
self.lastlayer = newlayer
# add max-pooling layer
if 'pool_size' in model_layer:
newlayer = name+'_pool' # str(counter) + '_' + name+'_pool'
if isinstance(model_layer['pool_size'], (tuple, list)):
self.network[newlayer] = layers.MaxPool2DLayer(self.network[self.lastlayer], pool_size=model_layer['pool_size'])
else:
self.network[newlayer] = layers.MaxPool2DLayer(self.network[self.lastlayer], pool_size=(model_layer['pool_size'], 1))
self.lastlayer = newlayer
# add dropout layer
if 'dropout' in model_layer:
newlayer = name+'_dropout' # str(counter) + '_' + name+'_dropout'
if model_layer['dropout']:
dropout = model_layer['dropout']
placeholder_name = 'keep_prob'+str(self.num_dropout)
exec(placeholder_name+" = tf.placeholder(tf.float32, name='"+placeholder_name+"')")
#exec("self.placeholders["+placeholder_name+"] = " + placeholder_name)
exec("self.hidden_feed_dict[" + placeholder_name+"] = "+str(dropout))
self.num_dropout += 1
self.network[newlayer] = layers.DropoutLayer(self.network[self.lastlayer], keep_prob=model_layer['dropout'])
self.lastlayer = newlayer