def _pre_conv(x, is_training,
num_outputs=None,
kernel_size=1,
stride=1,
padding='VALID',
activation='linear',
trainable=True,
scope='preconv',
reuse=None):
'''
Args:
num_outputs: If num_outputs is None, the input dimension is used.
'''
# TODO: Support multi-scale.
x = cnn.as_tensor(x)
if not num_outputs:
num_outputs = x.value.shape[-1].value
return cnn.slim_conv2d(x, num_outputs, kernel_size,
stride=stride,
padding=padding,
activation_fn=helpers.get_act(activation),
normalizer_fn=None, # No batch-norm.
trainable=trainable,
scope=scope,
reuse=reuse)
def slim_alexnet_v2(x,
is_training,
trainable=True,
variables_collections=None,
weight_decay=0.0005,
conv_padding='VALID',
pool_padding='VALID',
conv1_stride=4,
output_layer='conv5',
output_act='linear',
freeze_until_layer=None):
if not trainable:
raise NotImplementedError('trainable not supported')
# TODO: Support variables_collections.
with slim.arg_scope(
alexnet_pkg.alexnet_v2_arg_scope(weight_decay=weight_decay,
conv_padding=conv_padding,
pool_padding=pool_padding)):
return alexnet_pkg.alexnet_v2(
x,
is_training=is_training,
conv1_stride=conv1_stride,
output_layer=output_layer,
output_activation_fn=helpers.get_act(output_act),
freeze_until_layer=freeze_until_layer)
def alexnet(x,
is_training,
trainable=True,
variables_collections=None,
weight_decay=0,
output_layer='conv5',
output_act='linear',
freeze_until_layer=None,
padding='VALID',
enable_bnorm=True):
with slim.arg_scope(
feature_arg_scope(weight_decay=weight_decay,
enable_bnorm=enable_bnorm,
padding=padding)):
return _alexnet_layers(
x,
is_training,
trainable,
variables_collections,
output_layer=output_layer,
output_activation_fn=helpers.get_act(output_act),
freeze_until_layer=freeze_until_layer)
def _extra_conv(x,
is_training,
trainable,
variables_collections,
num_outputs=None,
kernel_size=1,
stride=1,
padding='VALID',
activation='linear'):
if not trainable:
raise NotImplementedError('trainable not supported')
x = cnn.as_tensor(x)
if num_outputs is None:
num_outputs = x.value.shape[-1].value
with slim.arg_scope([slim.batch_norm], is_training=is_training):
with slim.arg_scope([cnn.slim_conv2d],
variables_collections=variables_collections):
return cnn.slim_conv2d(x,
num_outputs,
kernel_size,
stride=stride,
padding=padding,
activation_fn=helpers.get_act(activation))