def fusion_net(feature_list, num_class, out_shape=(8,3)):
num=len(feature_list)
input = None
with tf.variable_scope('fuse-input') as scope:
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height==0 or pool_width==0): continue
roi_features, roi_idxs = tf_roipooling(feature,roi, pool_height, pool_width, pool_scale, name='%d/pool'%n)
roi_features = flatten(roi_features)
if input is None:
input = roi_features
else:
input = concat([input,roi_features], axis=1, name='%d/cat'%n)
with tf.variable_scope('fuse-block-1') as scope:
block = linear_bn_relu(input, num_hiddens=256, name='1')
block = linear_bn_relu(block, num_hiddens=256, name='2')
#include background class
with tf.variable_scope('fuse') as scope:
dim = np.product([*out_shape])
scores = linear(block, num_hiddens=num_class, name='score')
probs = tf.nn.softmax (scores, name='prob')
deltas = linear(block, num_hiddens=dim*num_class, name='box')
deltas = tf.reshape(deltas,(-1,num_class,*out_shape))
return scores, probs, deltas
def fusion_net(feature_list, num_class, out_shape=(8,3)):
test_num = 1
print('\n\n !!!! test_num ={}\n\n'.format(test_num))
if test_num==0:
num=len(feature_list)
input = None
with tf.variable_scope('fuse-input') as scope:
feature_names=['top_feature','front_feature','rgb_feature']
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height==0 or pool_width==0): continue
roi_features, roi_idxs = tf_roipooling(feature,roi, pool_height, pool_width,
pool_scale, name='%d/pool'%n)
roi_features = flatten(roi_features)
tf.summary.histogram(feature_names[n],roi_features)
if input is None:
input = roi_features
else:
input = concat([input,roi_features], axis=1, name='%d/cat'%n)
with tf.variable_scope('fuse-block-1') as scope:
block = linear_bn_relu(input, num_hiddens=512, name='1')
block = linear_bn_relu(block, num_hiddens=512, name='2')
block = linear_bn_relu(block, num_hiddens=512, name='3')
block = linear_bn_relu(block, num_hiddens=512, name='4')
elif test_num==1:
with tf.variable_scope('fuse-net') as scope:
num = len(feature_list)
feature_names = ['top', 'front', 'rgb']
roi_features_list = []
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height == 0 or pool_width == 0): continue
with tf.variable_scope(feature_names[n] + '-roi-pooling'):
roi_features, roi_idxs = tf_roipooling(feature, roi, pool_height, pool_width,
pool_scale, name='%s-roi_pooling' % feature_names[n])
with tf.variable_scope(feature_names[n]+ '-feature-conv'):
with tf.variable_scope('block1') as scope:
block = conv2d_bn_relu(roi_features, num_kernels=128, kernel_size=(3, 3),
stride=[1, 1, 1, 1], padding='SAME',name=feature_names[n]+'_conv_1')
residual=block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1],
padding='SAME',name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('block2') as scope:
block = conv2d_bn_relu(block, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('block3') as scope:
block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
roi_features = flatten(block)
tf.summary.histogram(feature_names[n], roi_features)
roi_features_list.append(roi_features)
with tf.variable_scope('rois-feature-concat'):
block = concat(roi_features_list, axis=1, name='concat')
elif test_num==2:
num = len(feature_list)
input = None
with tf.variable_scope('fuse-input') as scope:
feature_names = ['top_feature', 'front_feature', 'rgb_feature']
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height == 0 or pool_width == 0): continue
roi_features, roi_idxs = tf_roipooling(feature, roi, pool_height, pool_width,
pool_scale, name='%d/pool' % n)
with tf.variable_scope('roipooling-conv1') as scope:
block = conv2d_bn_relu(roi_features, num_kernels=128, kernel_size=(3, 3),
stride=[1, 1, 1, 1], padding='SAME',name=feature_names[n]+'_conv_1')
residual=block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1],
padding='SAME',name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('roipooling-conv2') as scope:
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('roipooling-conv3') as scope:
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
roi_features = flatten(block)
tf.summary.histogram(feature_names[n], roi_features)
if input is None:
input = roi_features
else:
input = concat([input, roi_features], axis=1, name='%d/cat' % n)
block = linear_bn_relu(input,128)
# include background class
with tf.variable_scope('fuse') as scope:
dim = np.product([*out_shape])
scores = linear(block, num_hiddens=num_class, name='score')
probs = tf.nn.softmax(scores, name='prob')
deltas = linear(block, num_hiddens=dim * num_class, name='box')
deltas = tf.reshape(deltas, (-1, num_class, *out_shape))
return block
def fusion_net(feature_list, num_class, out_shape=(8,3)):
stride=8
num=len(feature_list)
input = None
with tf.variable_scope('fuse-input') as scope:
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height==0 or pool_width==0): continue
# feature = conv2d_bn_relu(feature, num_kernels=512, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='%d'%n)
roi_features, roi_idxs = tf_roipooling(feature,roi, pool_height, pool_width, pool_scale, name='%d/pool'%n)
# pdb.set_trace()
roi_features=flatten(roi_features)
with tf.variable_scope('fuse-block-1-%d'%n):
tf.summary.histogram('fuse-block_input_%d'%n, roi_features)
block = linear_bn_relu(roi_features, num_hiddens=512, name='1')#512, so small?
tf.summary.histogram('fuse-block1_%d'%n, block)
block = tf.nn.dropout(block, keep_prob, name='drop1')
block = linear_bn_relu(block, num_hiddens=512, name='2')
tf.summary.histogram('fuse-block2_%d'%n, block)
block = tf.nn.dropout(block, keep_prob, name='drop2')
block = linear_bn_relu(roi_features, num_hiddens=512, name='3')#512, so small?
block = tf.nn.dropout(block, keep_prob, name='drop3')
if input is None:
input = block
else:
input = concat([input,block], axis=1, name='%d/cat'%n)
# if input is None:
# input = roi_features
# else:
# input = concat([input,roi_features], axis=1, name='%d/cat'%n)
# with tf.variable_scope('fuse-block-1') as scope:
# input_=None
# for i in len(feat):
# with tf.variable_scope('fc%d'%i):
# tf.summary.histogram('fuse-block_input_%d'%i, feat[i])
# block = linear_bn_relu(feat[i], num_hiddens=4096, name='1')#512, so small?
# tf.summary.histogram('fuse-block1_%d'%i, block)
# block = tf.nn.dropout(block, keep_prob, name='drop1')
# block = linear_bn_relu(block, num_hiddens=4096, name='2')
# tf.summary.histogram('fuse-block2_%d'%i, block)
# block = tf.nn.dropout(block, keep_prob, name='drop2')
# if input_ is None:
# input_ = block
# else:
# input_ = concat([input_,block], axis=1, name='%d/cat'%n)
# block = linear_bn_relu(block, num_hiddens=512, name='3')
# block = linear_bn_relu(block, num_hiddens=512, name='4')
#include background class
with tf.variable_scope('fuse') as scope:
block = linear_bn_relu(input, num_hiddens=512, name='4')#512, so small?
block = tf.nn.dropout(block, keep_prob, name='drop4')
dim = np.product([*out_shape])
scores = linear(block, num_hiddens=num_class, name='score')
probs = tf.nn.softmax (scores, name='prob')
deltas = linear(block, num_hiddens=dim*num_class, name='box')
deltas = tf.reshape(deltas,(-1,num_class,*out_shape))
return scores, probs, deltas
def fusion_net(feature_list, num_class, out_shape=(8,3)):
with tf.variable_scope('fuse-net') as scope:
num = len(feature_list)
feature_names = ['top', 'front', 'rgb']
roi_features_list = []
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height == 0 or pool_width == 0): continue
with tf.variable_scope(feature_names[n] + '-roi-pooling'):
roi_features, roi_idxs = tf_roipooling(feature, roi, pool_height, pool_width,
pool_scale, name='%s-roi_pooling' % feature_names[n])
with tf.variable_scope(feature_names[n]+ '-feature-conv'):
with tf.variable_scope('block1') as scope:
block = conv2d_bn_relu(roi_features, num_kernels=128, kernel_size=(3, 3),
stride=[1, 1, 1, 1], padding='SAME',name=feature_names[n]+'_conv_1')
residual=block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1],
padding='SAME',name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('block2') as scope:
block = conv2d_bn_relu(block, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('block3') as scope:
block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
roi_features = flatten(block)
tf.summary.histogram(feature_names[n], roi_features)
roi_features_list.append(roi_features)
with tf.variable_scope('rois-feature-concat'):
block = concat(roi_features_list, axis=1, name='concat')
with tf.variable_scope('fusion-feature-fc'):
print('\nUse fusion-feature-2fc')
block = linear_bn_relu(block, num_hiddens=512, name='1')
block = linear_bn_relu(block, num_hiddens=512, name='2')
return block
def fusion_net(feature_list, num_class, out_shape=(8,3)):
test_num = 1
print('\n\n !!!! test_num ={}\n\n'.format(test_num))
if test_num==0:
num=len(feature_list)
input = None
with tf.variable_scope('fuse-input') as scope:
feature_names=['top_feature','front_feature','rgb_feature']
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height==0 or pool_width==0): continue
roi_features, roi_idxs = tf_roipooling(feature,roi, pool_height, pool_width,
pool_scale, name='%d/pool'%n)
roi_features = flatten(roi_features)
tf.summary.histogram(feature_names[n],roi_features)
if input is None:
input = roi_features
else:
input = concat([input,roi_features], axis=1, name='%d/cat'%n)
with tf.variable_scope('fuse-block-1') as scope:
block = linear_bn_relu(input, num_hiddens=512, name='1')
block = linear_bn_relu(block, num_hiddens=512, name='2')
block = linear_bn_relu(block, num_hiddens=512, name='3')
block = linear_bn_relu(block, num_hiddens=512, name='4')
elif test_num==1:
with tf.variable_scope('fuse-net') as scope:
num = len(feature_list)
feature_names = ['top', 'front', 'rgb']
roi_features_list = []
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height == 0 or pool_width == 0): continue
with tf.variable_scope(feature_names[n] + '-roi-pooling'):
roi_features, roi_idxs = tf_roipooling(feature, roi, pool_height, pool_width,
pool_scale, name='%s-roi_pooling' % feature_names[n])
with tf.variable_scope(feature_names[n]+ '-feature-conv'):
with tf.variable_scope('block1') as scope:
block = conv2d_bn_relu(roi_features, num_kernels=128, kernel_size=(3, 3),
stride=[1, 1, 1, 1], padding='SAME',name=feature_names[n]+'_conv_1')
residual=block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1],
padding='SAME',name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('block2') as scope:
block = conv2d_bn_relu(block, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('block3') as scope:
block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
roi_features = flatten(block)
tf.summary.histogram(feature_names[n], roi_features)
roi_features_list.append(roi_features)
with tf.variable_scope('rois-feature-concat'):
block = concat(roi_features_list, axis=1, name='concat')
with tf.variable_scope('fusion-feature-fc'):
block = linear_bn_relu(block, num_hiddens=512, name='1')
block = linear_bn_relu(block, num_hiddens=512, name='2')
block = linear_bn_relu(block, num_hiddens=512, name='3')
block = linear_bn_relu(block, num_hiddens=512, name='4')
elif test_num==2:
num = len(feature_list)
input = None
with tf.variable_scope('fuse-input') as scope:
feature_names = ['top_feature', 'front_feature', 'rgb_feature']
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height == 0 or pool_width == 0): continue
roi_features, roi_idxs = tf_roipooling(feature, roi, pool_height, pool_width,
pool_scale, name='%d/pool' % n)
with tf.variable_scope('roipooling-conv1') as scope:
block = conv2d_bn_relu(roi_features, num_kernels=128, kernel_size=(3, 3),
stride=[1, 1, 1, 1], padding='SAME',name=feature_names[n]+'_conv_1')
residual=block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1],
padding='SAME',name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('roipooling-conv2') as scope:
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('roipooling-conv3') as scope:
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
roi_features = flatten(block)
tf.summary.histogram(feature_names[n], roi_features)
if input is None:
input = roi_features
else:
input = concat([input, roi_features], axis=1, name='%d/cat' % n)
block = linear_bn_relu(input,128)
# include background class
with tf.variable_scope('fuse') as scope:
dim = np.product([*out_shape])
scores = linear(block, num_hiddens=num_class, name='score')
probs = tf.nn.softmax(scores, name='prob')
deltas = linear(block, num_hiddens=dim * num_class, name='box')
deltas = tf.reshape(deltas, (-1, num_class, *out_shape))
return block
def fusion_net(feature_list, num_class, out_shape=(2, 2)):
num = len(feature_list)
input = None
with tf.variable_scope('fuse-input') as scope:
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height == 0 or pool_width == 0): continue
# feature = conv2d_bn_relu(feature, num_kernels=512, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='%d'%n)
roi_features, roi_idxs = tf_roipooling(feature,
roi,
pool_height,
pool_width,
pool_scale,
name='%d/pool' % n)
# pdb.set_trace()
roi_features = flatten(roi_features)
# roi_features_ = tf.stop_gradient(roi_features)
with tf.variable_scope('fuse-block-1-%d' % n):
tf.summary.histogram('fuse-block_input_%d' % n, roi_features)
block = linear_bn_relu(roi_features,
num_hiddens=2048,
name='1') #512, so small?
tf.summary.histogram('fuse-block1_%d' % n, block)
block = tf.nn.dropout(block, CFG.KEEPPROBS, name='drop1')
if input is None:
input = block
else:
# input = concat([input,block], axis=1, name='%d/cat'%n)
input = tf.add(input, block, name='fuse_feature')
# input_ = tf.stop_gradient(input)
input_ = input
#include background class
with tf.variable_scope('fuse') as scope:
block = linear_bn_relu(input_, num_hiddens=512,
name='4') #512, so small?
# block = tf.stop_gradient(block)
block = tf.nn.dropout(block, CFG.KEEPPROBS, name='drop4')
with tf.variable_scope('2D') as sc:
dim = np.product([*out_shape])
scores_3d = linear(block, num_hiddens=num_class, name='score')
probs_3d = tf.nn.softmax(scores_3d, name='prob')
deltas_3d = linear(block, num_hiddens=dim * num_class, name='box')
deltas_3d = tf.reshape(deltas_3d, (-1, num_class, *out_shape))
with tf.variable_scope('3D') as sc_:
block3D = linear_bn_relu(roi_features, num_hiddens=2048,
name='1') #512, so small?
block3D_1 = tf.nn.dropout(block3D, CFG.KEEPPROBS, name='drop1')
block = linear_bn_relu(block3D_1, num_hiddens=512, name='3D')
# block = tf.nn.dropout(block, CFG.KEEPPROBS , name='drop4')
dim = np.product(16)
deltas_2d = linear(block, num_hiddens=dim * num_class, name='box')
deltas_2d = tf.reshape(deltas_2d, (-1, num_class, 16))
# scores_3d = tf.stop_gradient(scores_3d)
# probs_3d = tf.stop_gradient(probs_3d)
# deltas_3d = tf.stop_gradient(deltas_3d)
return scores_3d, probs_3d, deltas_3d, deltas_2d
def fusion_net(feature_list, num_class, out_shape=(8,3)):
with tf.variable_scope('fuse-net') as scope:
num = len(feature_list)
feature_names = ['top', 'front', 'rgb']
roi_features_list = []
for n in range(num):
feature = feature_list[n][0]
roi = feature_list[n][1]
pool_height = feature_list[n][2]
pool_width = feature_list[n][3]
pool_scale = feature_list[n][4]
if (pool_height == 0 or pool_width == 0): continue
with tf.variable_scope(feature_names[n] + '-roi-pooling'):
roi_features, roi_idxs = tf_roipooling(feature, roi, pool_height, pool_width,
pool_scale, name='%s-roi_pooling' % feature_names[n])
with tf.variable_scope(feature_names[n]+ '-feature-conv'):
with tf.variable_scope('block1') as scope:
block = conv2d_bn_relu(roi_features, num_kernels=128, kernel_size=(3, 3),
stride=[1, 1, 1, 1], padding='SAME',name=feature_names[n]+'_conv_1')
residual=block
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1],
padding='SAME',name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('block2') as scope:
block = conv2d_bn_relu(block, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
with tf.variable_scope('block3') as scope:
block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_1')
residual = block
block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME',
name=feature_names[n]+'_conv_2')+residual
block = avgpool(block, kernel_size=(2, 2), stride=[1, 2, 2, 1],
padding='SAME', name=feature_names[n]+'_max_pool')
roi_features = flatten(block)
tf.summary.histogram(feature_names[n], roi_features)
roi_features_list.append(roi_features)
with tf.variable_scope('rois-feature-concat'):
block = concat(roi_features_list, axis=1, name='concat')
with tf.variable_scope('fusion-feature-fc'):
print('\nUse fusion-feature-2fc')
block = linear_bn_relu(block, num_hiddens=512, name='1')
block = linear_bn_relu(block, num_hiddens=512, name='2')
return block
