def testFPNOutputShape(self, min_level, max_level):
backbone_min_level = 2
backbone_max_level = 5
fpn_feat_dims = 256
image_size = 256
inputs = {}
for level in range(backbone_min_level, backbone_max_level + 1):
inputs[level] = tf.zeros([
1, image_size // 2**level, image_size // 2**level,
fpn_feat_dims
])
with tf.name_scope('min_level_%d_max_level_%d' %
(min_level, max_level)):
fpn_fn = fpn.Fpn(min_level=min_level,
max_level=max_level,
fpn_feat_dims=fpn_feat_dims)
features = fpn_fn(inputs)
for level in range(min_level, max_level):
self.assertEqual(features[level].get_shape().as_list(), [
1, image_size // 2**level, image_size // 2**level,
fpn_feat_dims
])
self.assertEqual(sorted(features.keys()),
range(min_level, max_level + 1))
def multilevel_features_generator(params):
"""Generator function for various FPN models."""
if params.architecture.multilevel_features == 'fpn':
fpn_params = params.fpn
fpn_fn = fpn.Fpn(min_level=fpn_params.min_level,
max_level=fpn_params.max_level,
fpn_feat_dims=fpn_params.fpn_feat_dims,
batch_norm_relu=batch_norm_relu_generator(
fpn_params.batch_norm))
else:
raise ValueError('The multi-level feature model %s is not supported.' %
params.architecture.multilevel_features)
return fpn_fn
def multilevel_features_generator(params):
"""Generator function for various FPN models."""
if params.architecture.multilevel_features == 'fpn':
fpn_params = params.fpn
fpn_fn = fpn.Fpn(min_level=params.architecture.min_level,
max_level=params.architecture.max_level,
fpn_feat_dims=fpn_params.fpn_feat_dims,
use_separable_conv=fpn_params.use_separable_conv,
activation=params.norm_activation.activation,
use_batch_norm=fpn_params.use_batch_norm,
norm_activation=norm_activation_generator(
params.norm_activation))
elif params.architecture.multilevel_features == 'identity':
fpn_fn = identity.Identity()
else:
raise ValueError(
'The multi-level feature model `{}` is not supported.'.format(
params.architecture.multilevel_features))
return fpn_fn