def init_model(self):
with tf.variable_scope ('RGB', reuse=tf.AUTO_REUSE):
if model_type == 'resnet50':
self.rgb_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'resnet101':
self.rgb_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_101',
unit_num=[3, 4, 23, 3])
elif model_type == 'resnet152':
self.rgb_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_152',
unit_num=[3, 8, 36, 3])
elif model_type == 'inception_v1':
self.rgb_model = inception_v1.InceptionV1 (num_classes=self._NUM_CLASSES)
with tf.variable_scope ('Flow', reuse=tf.AUTO_REUSE):
if model_type == 'resnet50':
self.flow_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'resnet101':
self.flow_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_101',
unit_num=[3, 4, 23, 3])
elif model_type == 'resnet152':
self.flow_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_152',
unit_num=[3, 8, 36, 3])
elif model_type == 'inception_v1':
self.flow_model = inception_v1.InceptionV1 (num_classes=self._NUM_CLASSES)
with tf.variable_scope('Fusion',reuse=tf.AUTO_REUSE):
if fusion_type == 'fhn':
self.fusion_model = fusion_net.FeatureHierachyNetwork (num_classes=self._NUM_CLASSES,fusion_mode = fusion_mode)
elif fusion_type == 'non_local':
self.fusion_model = non_local_fusion_net.space_cross_correlation_Network(num_classes=self._NUM_CLASSES,fusion_mode = fusion_mode)
elif fusion_type == 'channel':
self.fusion_model = non_local_fusion_net.channel_cross_correlation_Network(num_classes=self._NUM_CLASSES,fusion_mode = fusion_mode)
elif fusion_type == 'space_channel':
self.fusion_model = non_local_fusion_net.cross_correlation_Network(num_classes=self._NUM_CLASSES,fusion_mode = fusion_mode)
def init_model(self):
if eval_type in ['rgb', 'joint']:
with tf.variable_scope ('RGB', reuse=tf.AUTO_REUSE):
if model_type == 'resnet50':
self.rgb_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
print ('resnet50 process successfully')
elif model_type == 'resnet50_attention':
self.rgb_model = resnet_attention.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'inception_v1':
self.rgb_model = inception_v1.InceptionV1 (num_classes=self._NUM_CLASSES)
elif model_type == 'vgg_16':
self.rgb_model = vgg.vgg (num_classes=self._NUM_CLASSES)
elif model_type == 'resnet101':
self.rgb_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_101',
unit_num=[3, 4, 23, 3])
elif model_type == 'resnet152':
self.rgb_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_152',
unit_num=[3, 8, 36, 3])
elif model_type == 'inception_v2':
self.rgb_model = inception_v2.InceptionV2 (num_classes=self._NUM_CLASSES)
elif model_type == 'se_resnet50':
self.rgb_model = SE.SE_Resnet(num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'inception_v3':
self.rgb_model = inception_v3.InceptionV3 (num_classes=self._NUM_CLASSES)
elif model_type == 'inception_resnet_v2':
self.rgb_model = inception_resnet_v2.InceptionResnetV2(num_classes=self._NUM_CLASSES)
if eval_type in ['flow', 'joint']:
with tf.variable_scope ('Flow', reuse=tf.AUTO_REUSE):
if model_type == 'resnet50':
self.flow_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
print ('resnet50 process successfully')
elif model_type == 'resnet50_attention':
self.flow_model = resnet_attention.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'inception_v1':
self.flow_model = inception_v1.InceptionV1 (num_classes=self._NUM_CLASSES)
elif model_type == 'vgg_16':
self.flow_model = vgg.vgg (num_classes=self._NUM_CLASSES)
elif model_type == 'resnet101':
self.flow_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_101',
unit_num=[3, 4, 23, 3])
elif model_type == 'resnet152':
self.flow_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_152',
unit_num=[3, 8, 36, 3])
elif model_type == 'inception_v2':
self.flow_model = inception_v2.InceptionV2 (num_classes=self._NUM_CLASSES)
elif model_type == 'se_resnet50':
self.flow_model = SE.SE_Resnet(num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'inception_v3':
self.flow_model = inception_v3.InceptionV3 (num_classes=self._NUM_CLASSES)
elif model_type == 'inception_resnet_v2':
self.flow_model = inception_resnet_v2.InceptionResnetV2(num_classes=self._NUM_CLASSES)
def init_model(self):
if eval_type in ['rgb', 'joint']:
with tf.variable_scope('RGB', reuse=tf.AUTO_REUSE):
if model_type == 'resnet50':
self.rgb_model = resnet.Resnet(
num_classes=self._NUM_CLASSES,
name='resnet_v1_50',
eval_type=eval_type)
print('resnet50 process successfully')
elif model_type == 'inception_v1':
self.rgb_model = inception_v1.InceptionV1(
num_classes=self._NUM_CLASSES)
elif model_type == 'vgg_16':
self.rgb_model = vgg.vgg(num_classes=self._NUM_CLASSES)
elif model_type == 'resnet152':
self.rgb_model = resnet.Resnet(
num_classes=self._NUM_CLASSES,
name='resnet_v1_152',
eval_type=eval_type,
unit_num=[3, 8, 36, 3])
elif model_type == 'inception_v2':
self.rgb_model = inception_v2.InceptionV2(
num_classes=self._NUM_CLASSES)
elif model_type == 'resnet18':
self.rgb_model = resnet.Vanilla_Resnet(
num_classes=self._NUM_CLASSES, name='resnet_v1_18')
if eval_type in ['flow', 'joint']:
with tf.variable_scope('Flow', reuse=tf.AUTO_REUSE):
if model_type == 'resnet50':
self.flow_model = resnet.Resnet(
num_classes=self._NUM_CLASSES,
name='resnet_v1_50',
eval_type=eval_type)
print('resnet50 process successfully')
elif model_type == 'inception_v1':
self.flow_model = inception_v1.InceptionV1(
num_classes=self._NUM_CLASSES)
elif model_type == 'vgg_16':
self.flow_model = vgg.vgg(num_classes=self._NUM_CLASSES)
elif model_type == 'resnet152':
self.flow_model = resnet.Resnet(
num_classes=self._NUM_CLASSES,
name='resnet_v1_152',
eval_type=eval_type,
unit_num=[3, 8, 36, 3])
elif model_type == 'inception_v2':
self.flow_model = inception_v2.InceptionV2(
num_classes=self._NUM_CLASSES)
elif model_type == 'resnet18':
self.rgb_model = resnet.Vanilla_Resnet(
num_classes=self._NUM_CLASSES, name='resnet_v1_18')
endpoints['logits'] = logits
return logits, endpoints
if __name__ == "__main__":
try:
from model import resnet, inception_v1, inception_v2, vgg, SE, densenet
except:
import resnet, inception_v1, inception_v2, vgg, SE, densenet
# tf.enable_eager_execution()
rgb_input = tf.placeholder(tf.float32, [None, 224, 224, 3])
flow_input = tf.placeholder(tf.float32, [None, 224, 224, 20])
rgb_logits, rgb_endpoints = inception_v1.InceptionV1()(
rgb_input, is_training=False, dropout_keep_prob=1.0)
flow_logits, flow_endpoints = inception_v1.InceptionV1()(
flow_input, is_training=False, dropout_keep_prob=1.0)
feature_list = [rgb_endpoints, flow_endpoints]
with tf.variable_scope('fusion'):
fusion_logits, endpoints = FeatureHierachyNetwork(
fusion_mode='concat')(feature_list,
is_training=True,
dropout_keep_prob=1.0)
model_logits = fusion_logits
print(model_logits)
# sess = tf.Session()