def build_model(images, num_classes, is_training=True, reuse=None):
model = delf_v1.DelfV1()
net, end_points = model.GetResnet50Subnetwork(images,
global_pool=True,
is_training=is_training,
reuse=reuse)
with slim.arg_scope(
resnet_v1.resnet_arg_scope(weight_decay=0.0001,
batch_norm_scale=True)):
with slim.arg_scope([slim.batch_norm], is_training=True):
feature_map = end_points['resnet_v1_50/block3']
feature_map = slim.conv2d(feature_map,
512,
1,
rate=1,
activation_fn=tf.nn.relu,
scope='conv1')
feature_map = tf.reduce_mean(feature_map, [1, 2])
feature_map = tf.expand_dims(tf.expand_dims(feature_map, 1), 2)
logits = slim.conv2d(feature_map,
num_classes, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='logits')
logits = tf.squeeze(logits, [1, 2], name='spatial_squeeze')
return logits
def _ModelFn(images, normalized_image, reuse):
"""Attention model to get feature map and attention score map.
Args:
images: Image tensor.
normalized_image: Whether or not the images are normalized.
reuse: Whether or not the layer and its variables should be reused.
Returns:
attention: Attention score after the non-linearity.
feature_map: Feature map after ResNet convolution.
"""
if normalized_image:
image_tensor = images
else:
image_tensor = NormalizePixelValues(images)
# Extract features and attention scores.
model = delf_v1.DelfV1(layer_name)
_, attention, _, feature_map, _ = model.GetAttentionPrelogit(
image_tensor,
attention_nonlinear=attention_nonlinear,
attention_type=attention_type,
kernel=[attention_kernel_size, attention_kernel_size],
training_resnet=False,
training_attention=False,
reuse=reuse)
return attention, feature_map
def build_model(images, num_classes, is_training=True):
model = delf_v1.DelfV1()
logits, attention, feature_map = model.AttentionModel(
images,
num_classes=num_classes,
training_resnet=False,
training_attention=is_training)
return logits
def BuildRegModel(images, normalized_image=False, attentive=True):
if normalized_image:
image_tensor = images
else:
image_tensor = NormalizePixelValues(images)
model = delf_v1.DelfV1()
_, _, _, feature_map, end_points = model.GetAttentionPrelogit(image_tensor)
if attentive:
return end_points['attention_feature_map']
else:
return feature_map
def _ModelFn(images, normalized_image, reuse):
image_tensor = images if normalized_image else NormalizePixelValues(
images)
model = delf_v1.DelfV1(layer_name)
_, attention, _, feature_map, _ = model.GetAttentionPrelogit(
image_tensor,
attention_nonlinear=attention_nonlinear,
attention_type=attention_type,
kernel=[attention_kernel_size, attention_kernel_size],
training_resnet=False,
training_attention=False,
reuse=reuse)
return attention, feature_map
def network(self, images, is_training=True, reuse=None):
model = delf_v1.DelfV1()
net, end_points = model.GetResnet50Subnetwork(images, global_pool=True, is_training=is_training, reuse=reuse)
with slim.arg_scope(
resnet_v1.resnet_arg_scope(
weight_decay=0.0001, batch_norm_scale=True)):
with slim.arg_scope([slim.batch_norm], is_training=True):
feature_map = end_points['siamese/resnet_v1_50/block3']
feature_map = slim.conv2d(
feature_map,
512,
1,
rate=1,
activation_fn=tf.nn.relu,
scope='conv1')
feature_map = tf.reduce_mean(feature_map, [1, 2])
feature_map = tf.expand_dims(tf.expand_dims(feature_map, 1), 2)
return feature_map
