def endpoints(image, is_training):
if image.get_shape().ndims != 4:
raise ValueError('Input must be of size [batch, height, width, 3]')
#image = image - tf.constant(_RGB_MEAN, dtype=tf.float32, shape=(1,1,1,3))
with tf.contrib.slim.arg_scope(
resnet_arg_scope(batch_norm_decay=0.9, weight_decay=0.0)):
_, endpoints = resnet_v1_50(image,
num_classes=None,
is_training=is_training,
global_pool=True)
resnet_out_shape = endpoints['img_var/resnet_v1_50/block4'].get_shape()
num_feature = resnet_out_shape[1] * resnet_out_shape[2]
NetVLAD = lp.NetVLAD(feature_size=2048,
max_samples=num_feature,
cluster_size=16,
output_dim=1000,
gating=True,
add_batch_norm=True,
is_training=is_training)
endpoints['vlad_in'] = tf.reshape(endpoints['img_var/resnet_v1_50/block4'],
[-1, 2048])
endpoints['vlad_out'] = NetVLAD.forward(endpoints['vlad_in'])
return endpoints['vlad_out'], 'resnet_v1_50'
def endpoints(image, pc_trans_feat, is_training):
if image.get_shape().ndims != 4:
raise ValueError('Input must be of size [batch, height, width, 3]')
image = image - tf.constant(
_RGB_MEAN, dtype=tf.float32, shape=(1, 1, 1, 3))
with tf.contrib.slim.arg_scope(
resnet_arg_scope(batch_norm_decay=0.9, weight_decay=0.0)):
_, endpoints = resnet_v1_50(image,
num_classes=None,
is_training=is_training,
global_pool=True)
if not pc_trans_feat is None:
concat_feat = tf.concat(
(endpoints['img_var/resnet_v1_50/block4'], pc_trans_feat), axis=3)
conv_feat = tf.layers.conv2d(concat_feat,
2048,
3,
activation=tf.nn.relu)
img_pc_feat = tf.reduce_mean(conv_feat, [1, 2], name='pool6')
img_pc_feat = tf.layers.dense(img_pc_feat, 1000, activation=tf.nn.relu)
img_pc_feat = tf.nn.l2_normalize(img_pc_feat, 1)
else:
img_pc_feat = None
img_feat = tf.reduce_mean(endpoints['img_var/resnet_v1_50/block4'], [1, 2],
name='pool5')
img_feat = tf.layers.dense(img_feat, 1000, activation=tf.nn.relu)
img_feat = tf.nn.l2_normalize(img_feat, 1)
return img_feat, img_pc_feat
def endpoints(pcai, is_training):
with tf.contrib.slim.arg_scope(
resnet_arg_scope(batch_norm_decay=0.9, weight_decay=0.0)):
_, endpoints = resnet_v1_fusion(pcai,
num_classes=None,
is_training=is_training,
global_pool=False)
#print(endpoints['fusion_var/resnet_v1_fusion/block1'])
endpoints['final_pooling'] = tf.reduce_mean(
endpoints['fusion_var/resnet_v1_fusion/block1'], [1, 2], name='pool5')
endpoints['model_output'] = tf.layers.dense(endpoints['final_pooling'],
1024)
return endpoints, 'resnet_v1_50'
def endpoints(image, is_training):
if image.get_shape().ndims != 4:
raise ValueError('Input must be of size [batch, height, width, 3]')
image = image - tf.constant(
_RGB_MEAN, dtype=tf.float32, shape=(1, 1, 1, 3))
with tf.contrib.slim.arg_scope(
resnet_arg_scope(batch_norm_decay=0.9, weight_decay=0.0)):
_, endpoints = resnet_v1_50(image,
num_classes=None,
is_training=is_training,
global_pool=True)
endpoints['model_output'] = endpoints['global_pool'] = tf.reduce_mean(
endpoints['img_var/resnet_v1_50/block4'], [1, 2], name='pool5')
return endpoints['img_var/resnet_v1_50/block4'], endpoints[
'model_output'], 'resnet_v1_50'