def build(self, _input, net, store_output_op=False):
assert (isinstance(net, BasicModel))
output = _input
if not self.ready:
return output
with tf.variable_scope(self._id):
self._scope = tf.get_variable_scope().name
param_initializer = self.param_initializer
if self.pre_activation:
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# Pooling
if self._type == 'avg':
output = BasicModel.avg_pool(output,
k=self.kernel_size,
s=self.strides)
elif self._type == 'max':
output = BasicModel.max_pool(output,
k=self.kernel_size,
s=self.strides)
else:
raise ValueError('Do not support the pooling type: %s' %
self._type)
else:
# Pooling
if self._type == 'avg':
output = BasicModel.avg_pool(output,
k=self.kernel_size,
s=self.strides)
elif self._type == 'max':
output = BasicModel.max_pool(output,
k=self.kernel_size,
s=self.strides)
else:
raise ValueError('Do not support the pooling type: %s' %
self._type)
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# dropout
output = BasicModel.dropout(output, self.keep_prob,
net.is_training)
if store_output_op:
self.output_op = output
return output
def build(self, _input, net, store_output_op=False):
assert (isinstance(net, BasicModel))
output = _input
if not self.ready:
return output
with tf.variable_scope(self._id):
self._scope = tf.get_variable_scope().name
param_initializer = self.param_initializer
# flatten if not
output = BasicModel.flatten(output)
if self.pre_activation:
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# FC
output = BasicModel.fc_layer(
output,
self.units,
self.use_bias,
param_initializer=param_initializer)
else:
# FC
output = BasicModel.fc_layer(
output,
self.units,
self.use_bias,
param_initializer=param_initializer)
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# dropout
output = BasicModel.dropout(output, self.keep_prob,
net.is_training)
if store_output_op:
self.output_op = output
return output
def build(self, _input, net, store_output_op=False):
assert (isinstance(net, BasicModel))
output = _input
if not self.ready:
return output
with tf.variable_scope(self._id):
self._scope = tf.get_variable_scope().name
param_initializer = self.param_initializer
if self.pre_activation:
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# convolutional
output = BasicModel.conv2d(output,
self.filter_num,
self.kernel_size,
self.strides,
param_initializer=param_initializer)
else:
# convolutional
output = BasicModel.conv2d(output,
self.filter_num,
self.kernel_size,
self.strides,
param_initializer=param_initializer)
# batch normalization
if self.use_bn:
output = BasicModel.batch_norm(
output,
net.is_training,
net.net_config.bn_epsilon,
net.net_config.bn_decay,
param_initializer=param_initializer)
# activation
output = BasicModel.activation(output, self.activation)
# dropout
output = BasicModel.dropout(output, self.keep_prob,
net.is_training)
if store_output_op:
self.output_op = output
return output
