def top_feature_net(input, anchors, inds_inside, num_bases):
stride=8
with tf.variable_scope("top_base") as sc:
arg_scope = resnet_v1.resnet_arg_scope(weight_decay=0.0)
with slim.arg_scope(arg_scope) :
net, end_points = resnet_v1.resnet_v1_50(input, None, global_pool=False, output_stride=8)
#pdb.set_trace()
block=end_points['top_base/resnet_v1_50/block4']
# block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='2')
tf.summary.histogram('rpn_top_block', block)
# tf.summary.histogram('rpn_top_block_weights', tf.get_collection('2/conv_weight')[0])
with tf.variable_scope('top') as scope:
#up = upsample2d(block, factor = 2, has_bias=True, trainable=True, name='1')
#up = block
up = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
scores = conv2d(up, num_kernels=2*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='score')
probs = tf.nn.softmax( tf.reshape(scores,[-1,2]), name='prob')
deltas = conv2d(up, num_kernels=4*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='delta')
#<todo> flip to train and test mode nms (e.g. different nms_pre_topn values): use tf.cond
with tf.variable_scope('top-nms') as scope: #non-max
batch_size, img_height, img_width, img_channel = input.get_shape().as_list()
img_scale = 1
# pdb.set_trace()
rois, roi_scores = tf_rpn_nms( probs, deltas, anchors, inds_inside,
stride, img_width, img_height, img_scale,
nms_thresh=0.7, min_size=stride, nms_pre_topn=300, nms_post_topn=50,
name ='nms')
#<todo> feature = upsample2d(block, factor = 4, ...)
feature = block
def top_feature_net(input, anchors, inds_inside, num_bases):
"""temporary net for debugging only. may not follow the paper exactly ....
:param input:
:param anchors:
:param inds_inside:
:param num_bases:
:return:
top_features, top_scores, top_probs, top_deltas, proposals, proposal_scores
"""
stride=1.
#with tf.variable_scope('top-preprocess') as scope:
# input = input
with tf.variable_scope('top-block-1') as scope:
block = conv2d_bn_relu(input, num_kernels=32, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='1')
block = conv2d_bn_relu(block, num_kernels=32, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
block = maxpool(block, kernel_size=(2,2), stride=[1,2,2,1], padding='SAME', name='4' )
stride *=2
with tf.variable_scope('top-block-2') as scope:
block = conv2d_bn_relu(block, num_kernels=64, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='1')
block = conv2d_bn_relu(block, num_kernels=64, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
block = maxpool(block, kernel_size=(2,2), stride=[1,2,2,1], padding='SAME', name='4' )
stride *=2
with tf.variable_scope('top-block-3') as scope:
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='1')
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='3')
block = maxpool(block, kernel_size=(2,2), stride=[1,2,2,1], padding='SAME', name='4' )
stride *=2
with tf.variable_scope('top-block-4') as scope:
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='1')
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='3')
with tf.variable_scope('top') as scope:
#up = upsample2d(block, factor = 2, has_bias=True, trainable=True, name='1')
#up = block
up = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
scores = conv2d(up, num_kernels=2*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='score')
probs = tf.nn.softmax( tf.reshape(scores,[-1,2]), name='prob')
deltas = conv2d(up, num_kernels=4*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='delta')
#<todo> flip to train and test mode nms (e.g. different nms_pre_topn values): use tf.cond
with tf.variable_scope('top-nms') as scope: #non-max
batch_size, img_height, img_width, img_channel = input.get_shape().as_list()
img_scale = 1
rois, roi_scores = tf_rpn_nms( probs, deltas, anchors, inds_inside,
stride, img_width, img_height, img_scale,
nms_thresh=0.7, min_size=stride, nms_pre_topn=500, nms_post_topn=100,
name ='nms')
#<todo> feature = upsample2d(block, factor = 4, ...)
feature = block
print ('top: scale=%f, stride=%d'%(1./stride, stride))
return feature, scores, probs, deltas, rois, roi_scores, stride
def top_feature_net(input, anchors, inds_inside, num_bases):
stride=8
with tf.variable_scope("top_base"):
arg_scope=mobilenet_arg_scope(weight_decay=0.0)
with slim.arg_scope(arg_scope):
end_points=mobilenet(input,num_classes=1000,is_training=True,width_multiplier=1,scope='MobileNet')
block=end_points['top_base/MobileNet/conv_ds_14/depthwise_conv']
with tf.variable_scope('top') as scope:
#up = upsample2d(block, factor = 2, has_bias=True, trainable=True, name='1')
#up = block
up = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
scores = conv2d(up, num_kernels=2*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='score')
probs = tf.nn.softmax( tf.reshape(scores,[-1,2]), name='prob')
deltas = conv2d(up, num_kernels=4*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='delta')
#<todo> flip to train and test mode nms (e.g. different nms_pre_topn values): use tf.cond
with tf.variable_scope('top-nms') as scope: #non-max
batch_size, img_height, img_width, img_channel = input.get_shape().as_list()
img_scale = 1
# pdb.set_trace()
rois, roi_scores = tf_rpn_nms( probs, deltas, anchors, inds_inside,
stride, img_width, img_height, img_scale,
nms_thresh=0.7, min_size=stride, nms_pre_topn=500, nms_post_topn=50,
name ='nms')
#<todo> feature = upsample2d(block, factor = 4, ...)
feature = block
# print ('top: scale=%f, stride=%d'%(1./stride, stride))
return feature, scores, probs, deltas, rois, roi_scores
def top_feature_net(input, anchors, inds_inside, num_bases):
"""temporary net for debugging only. may not follow the paper exactly ....
:param input:
:param anchors:
:param inds_inside:
:param num_bases:
:return:
top_features, top_scores, top_probs, top_deltas, proposals, proposal_scores
"""
stride=1.
#with tf.variable_scope('top-preprocess') as scope:
# input = input
with tf.variable_scope('top-block-1') as scope:
block = conv2d_bn_relu(input, num_kernels=32, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='1')
block = conv2d_bn_relu(block, num_kernels=32, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
block = maxpool(block, kernel_size=(2,2), stride=[1,2,2,1], padding='SAME', name='4' )
stride *=2
with tf.variable_scope('top-block-2') as scope:
block = conv2d_bn_relu(block, num_kernels=64, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='1')
block = conv2d_bn_relu(block, num_kernels=64, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
block = maxpool(block, kernel_size=(2,2), stride=[1,2,2,1], padding='SAME', name='4' )
stride *=2
with tf.variable_scope('top-block-3') as scope:
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='1')
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='3')
block = maxpool(block, kernel_size=(2,2), stride=[1,2,2,1], padding='SAME', name='4' )
stride *=2
with tf.variable_scope('top-block-4') as scope:
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='1')
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='3')
with tf.variable_scope('top') as scope:
#up = upsample2d(block, factor = 2, has_bias=True, trainable=True, name='1')
#up = block
up = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2')
scores = conv2d(up, num_kernels=2*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='score')
probs = tf.nn.softmax( tf.reshape(scores,[-1,2]), name='prob')
deltas = conv2d(up, num_kernels=4*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='delta')
#<todo> flip to train and test mode nms (e.g. different nms_pre_topn values): use tf.cond
with tf.variable_scope('top-nms') as scope: #non-max
batch_size, img_height, img_width, img_channel = input.get_shape().as_list()
img_scale = 1
rois, roi_scores = tf_rpn_nms( probs, deltas, anchors, inds_inside,
stride, img_width, img_height, img_scale,
nms_thresh=0.7, min_size=stride, nms_pre_topn=500, nms_post_topn=100,
name ='nms')
#<todo> feature = upsample2d(block, factor = 4, ...)
feature = block
print ('top: scale=%f, stride=%d'%(1./stride, stride))
return feature, scores, probs, deltas, rois, roi_scores, stride
def top_feature_net_r(input, anchors, inds_inside, num_bases):
"""
:param input:
:param anchors:
:param inds_inside:
:param num_bases:
:return:
top_features, top_scores, top_probs, top_deltas, proposals, proposal_scores
"""
stride=1.
#with tf.variable_scope('top-preprocess') as scope:
# input = input
batch_size, img_height, img_width, img_channel = input.get_shape().as_list()
with tf.variable_scope('feature-extract-resnet') as scope:
print('build_resnet')
block = ResnetBuilder.resnet_tiny(input)
# resnet_input = resnet.get_layer('input_1').input
# resnet_output = resnet.get_layer('add_7').output
# resnet_f = Model(inputs=resnet_input, outputs=resnet_output) # add_7
# # print(resnet_f.summary())
# block = resnet_f(input)
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='2')
stride = 8
feature = block
with tf.variable_scope('predict') as scope:
# up = upsample2d(block, factor = 2, has_bias=True, trainable=True, name='1')
# up = block
kernel_size = config.cfg.TOP_CONV_KERNEL_SIZE
print('\ntop_predict kernal_size: {}\n'.format(kernel_size) )
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(kernel_size, kernel_size),
stride=[1, 1, 1, 1], padding='SAME', name='1')
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(kernel_size, kernel_size),
stride=[1, 1, 1, 1], padding='SAME', name='2')
scores = conv2d(block, num_kernels=2 * num_bases, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME',name='score')
probs = tf.nn.softmax(tf.reshape(scores, [-1, 2]), name='prob')
deltas = conv2d(block, num_kernels=4 * num_bases, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME',name='delta')
#<todo> flip to train and test mode nms (e.g. different nms_pre_topn values): use tf.cond
with tf.variable_scope('NMS') as scope: #non-max
img_scale = 1
rois, roi_scores = tf_rpn_nms( probs, deltas, anchors, inds_inside,
stride, img_width, img_height, img_scale,
nms_thresh=0.7, min_size=stride, nms_pre_topn=500, nms_post_topn=100,
name ='nms')
print ('top: scale=%f, stride=%d'%(1./stride, stride))
return feature, scores, probs, deltas, rois, roi_scores, stride
def top_feature_net(input, anchors, inds_inside, num_bases):
stride = 4
# arg_scope = resnet_v1.resnet_arg_scope(weight_decay=0.0)
# with slim.arg_scope(arg_scope) :
with slim.arg_scope(vgg.vgg_arg_scope()):
# net, end_points = resnet_v1.resnet_v1_50(input, None, global_pool=False, output_stride=8)
block5, end_points = vgg.vgg_16(input)
# pdb.set_trace()
block3 = end_points['vgg_16/conv3/conv3_3']
block4 = end_points['vgg_16/conv4/conv4_3']
with tf.variable_scope("top_base") as sc:
block5_ = conv2d_bn_relu(block5,
num_kernels=256,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='5')
up5 = upsample2d(block5_,
factor=2,
has_bias=True,
trainable=True,
name='up1')
block4_ = conv2d_bn_relu(block4,
num_kernels=256,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='4')
up4 = upsample2d(block4_,
factor=2,
has_bias=True,
trainable=True,
name='up2')
up_ = tf.add(up4, up5, name="up_add")
block3_ = conv2d_bn_relu(block3,
num_kernels=256,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='3')
up = tf.add(up_, block3_, name='fpn_up_')
block = conv2d_bn_relu(up,
num_kernels=256,
kernel_size=(3, 3),
stride=[1, 1, 1, 1],
padding='SAME',
name='fpn_up')
tf.summary.histogram('rpn_top_block', block)
# tf.summary.histogram('rpn_top_block_weights', tf.get_collection('2/conv_weight')[0])
with tf.variable_scope('top') as scope:
#up = upsample2d(block, factor = 2, has_bias=True, trainable=True, name='1')
#up = block
up = conv2d_bn_relu(block,
num_kernels=128,
kernel_size=(3, 3),
stride=[1, 1, 1, 1],
padding='SAME',
name='2')
scores = conv2d(up,
num_kernels=2 * num_bases,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='score')
probs = tf.nn.softmax(tf.reshape(scores, [-1, 2]), name='prob')
deltas = conv2d(up,
num_kernels=4 * num_bases,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='delta')
#<todo> flip to train and test mode nms (e.g. different nms_pre_topn values): use tf.cond
with tf.variable_scope('top-nms') as scope: #non-max
batch_size, img_height, img_width, img_channel = input.get_shape(
).as_list()
img_scale = 1
# pdb.set_trace()
rois, roi_scores = tf_rpn_nms(probs,
deltas,
anchors,
inds_inside,
stride,
img_width,
img_height,
img_scale,
nms_thresh=0.7,
min_size=stride,
nms_pre_topn=nms_pre_topn_,
nms_post_topn=nms_post_topn_,
name='nms')
#<todo> feature = upsample2d(block, factor = 4, ...)
feature = block
# print ('top: scale=%f, stride=%d'%(1./stride, stride))
return feature, scores, probs, deltas, rois, roi_scores
def top_feature_net_r(input, anchors, inds_inside, num_bases):
"""
:param input:
:param anchors:
:param inds_inside:
:param num_bases:
:return:
top_features, top_scores, top_probs, top_deltas, proposals, proposal_scores
"""
batch_size, img_height, img_width, img_channel = input.get_shape().as_list(
)
with tf.variable_scope('feature-extract-resnet') as scope:
print('build_resnet')
block = ResnetBuilder.resnet_tiny_smaller_kernel(input)
feature = block
top_feature_stride = 4
# resnet_input = resnet.get_layer('input_1').input
# resnet_output = resnet.get_layer('add_7').output
# resnet_f = Model(inputs=resnet_input, outputs=resnet_output) # add_7
# # print(resnet_f.summary())
# block = resnet_f(input)
block = upsample2d(block,
factor=2,
has_bias=True,
trainable=True,
name='upsampling')
with tf.variable_scope('predict') as scope:
# block = upsample2d(block, factor=4, has_bias=True, trainable=True, name='1')
# up = block
# kernel_size = config.cfg.TOP_CONV_KERNEL_SIZE
top_anchors_stride = 2
block = conv2d_bn_relu(block,
num_kernels=128,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='2')
scores = conv2d(block,
num_kernels=2 * num_bases,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='score')
probs = tf.nn.softmax(tf.reshape(scores, [-1, 2]), name='prob')
deltas = conv2d(block,
num_kernels=4 * num_bases,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='delta')
with tf.variable_scope('RPN-NMS') as scope:
train_rois, train_roi_scores = tf_rpn_nms( probs, deltas, anchors, inds_inside, \
top_anchors_stride, img_width, img_height, img_scale=1, \
nms_thresh=0.7, min_size=top_anchors_stride, \
nms_pre_topn=CFG.TRAIN.RPN_NMS_PRE_TOPN, nms_post_topn=CFG.TRAIN.RPN_NMS_POST_TOPN, \
name ='train-rpn-nms')
infer_rois, infer_roi_scores = tf_rpn_nms( probs, deltas, anchors, inds_inside, \
top_anchors_stride, img_width, img_height, img_scale=1, \
nms_thresh=0.1, min_size=top_anchors_stride, \
nms_pre_topn=CFG.TRAIN.RPN_NMS_PRE_TOPN, nms_post_topn=CFG.TEST.RPN_NMS_POST_TOPN, \
name ='infer-rpn-nms')
print('top: anchor stride=%d, feature_stride=%d' %
(top_anchors_stride, top_feature_stride))
return feature, scores, probs, deltas, train_rois, train_roi_scores, top_anchors_stride,top_feature_stride, \
infer_rois, infer_roi_scores
def top_feature_net_r(input,
anchors,
inds_inside,
num_bases,
top_last_states=None):
"""
:param input:
:param anchors:
:param inds_inside:
:param num_bases:
:return:
top_features, top_scores, top_probs, top_deltas, proposals, proposal_scores
"""
stride = 1.
#with tf.variable_scope('top-preprocess') as scope:
# input = input
batch_size, img_height, img_width, img_channel = input.get_shape().as_list(
)
with tf.variable_scope('feature-extract-resnet') as scope:
print('build_resnet')
block = ResnetBuilder.resnet_tiny(input)
# resnet_input = resnet.get_layer('input_1').input
# resnet_output = resnet.get_layer('add_7').output
# resnet_f = Model(inputs=resnet_input, outputs=resnet_output) # add_7
# # print(resnet_f.summary())
# block = resnet_f(input)
block = conv2d_bn_relu(block,
num_kernels=128,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='2')
stride = 8
with tf.variable_scope('memory') as scope:
with tf.variable_scope('get_last_channel') as scope:
lstm_input = block[:, :, :, 127]
ori_shape = lstm_input.get_shape().as_list()
# [batch_size, sequence_length_max, vector_size]
lstm_input = tf.reshape(lstm_input,
[-1, 1, np.prod(ori_shape[1:])])
with tf.variable_scope('lstm') as scope:
lstm_cell = rnn.BasicLSTMCell(np.prod(ori_shape[1:]))
outputs, top_states = tf.nn.dynamic_rnn(
lstm_cell,
lstm_input,
initial_state=top_last_states,
dtype=tf.float32)
outputs = tf.reshape(outputs, [-1, ori_shape[1], ori_shape[2]],
'reshape')
with tf.variable_scope('merge') as scope:
block = tf.concat([
block[:, :, :, 0:127],
tf.reshape(outputs, [-1, ori_shape[1], ori_shape[2], 1])
], 3)
with tf.variable_scope('predict') as scope:
# up = upsample2d(block, factor = 2, has_bias=True, trainable=True, name='1')
# up = block
up = conv2d_bn_relu(block,
num_kernels=128,
kernel_size=(3, 3),
stride=[1, 1, 1, 1],
padding='SAME',
name='2')
scores = conv2d(up,
num_kernels=2 * num_bases,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='score')
probs = tf.nn.softmax(tf.reshape(scores, [-1, 2]), name='prob')
deltas = conv2d(up,
num_kernels=4 * num_bases,
kernel_size=(1, 1),
stride=[1, 1, 1, 1],
padding='SAME',
name='delta')
#<todo> flip to train and test mode nms (e.g. different nms_pre_topn values): use tf.cond
with tf.variable_scope('NMS') as scope: #non-max
img_scale = 1
rois, roi_scores = tf_rpn_nms(probs,
deltas,
anchors,
inds_inside,
stride,
img_width,
img_height,
img_scale,
nms_thresh=0.7,
min_size=stride,
nms_pre_topn=500,
nms_post_topn=100,
name='nms')
#<todo> feature = upsample2d(block, factor = 4, ...)
feature = block
print('top: scale=%f, stride=%d' % (1. / stride, stride))
return feature, scores, probs, deltas, rois, roi_scores, stride, top_states
def top_feature_net_r(input, anchors, inds_inside, num_bases, top_last_states=None):
"""
:param input:
:param anchors:
:param inds_inside:
:param num_bases:
:return:
top_features, top_scores, top_probs, top_deltas, proposals, proposal_scores
"""
stride=1.
#with tf.variable_scope('top-preprocess') as scope:
# input = input
batch_size, img_height, img_width, img_channel = input.get_shape().as_list()
with tf.variable_scope('feature-extract-resnet') as scope:
print('build_resnet')
block = ResnetBuilder.resnet_tiny(input)
# resnet_input = resnet.get_layer('input_1').input
# resnet_output = resnet.get_layer('add_7').output
# resnet_f = Model(inputs=resnet_input, outputs=resnet_output) # add_7
# # print(resnet_f.summary())
# block = resnet_f(input)
block = conv2d_bn_relu(block, num_kernels=128, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='2')
stride = 8
with tf.variable_scope('memory') as scope:
with tf.variable_scope('get_last_channel') as scope:
lstm_input = block[:,:,:,127]
ori_shape = lstm_input.get_shape().as_list()
# [batch_size, sequence_length_max, vector_size]
lstm_input = tf.reshape(lstm_input, [-1, 1, np.prod(ori_shape[1:])])
with tf.variable_scope('lstm') as scope:
lstm_cell = rnn.BasicLSTMCell(np.prod(ori_shape[1:]) )
outputs, top_states = tf.nn.dynamic_rnn(lstm_cell, lstm_input, initial_state=top_last_states, dtype=tf.float32)
outputs = tf.reshape(outputs, [-1, ori_shape[1],ori_shape[2]], 'reshape')
with tf.variable_scope('merge') as scope:
block = tf.concat([block[:, :, :, 0:127],tf.reshape(outputs,[-1,ori_shape[1],ori_shape[2],1])], 3)
with tf.variable_scope('predict') as scope:
# up = upsample2d(block, factor = 2, has_bias=True, trainable=True, name='1')
# up = block
up = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME', name='2')
scores = conv2d(up, num_kernels=2 * num_bases, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME',name='score')
probs = tf.nn.softmax(tf.reshape(scores, [-1, 2]), name='prob')
deltas = conv2d(up, num_kernels=4 * num_bases, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME',name='delta')
#<todo> flip to train and test mode nms (e.g. different nms_pre_topn values): use tf.cond
with tf.variable_scope('NMS') as scope: #non-max
img_scale = 1
rois, roi_scores = tf_rpn_nms( probs, deltas, anchors, inds_inside,
stride, img_width, img_height, img_scale,
nms_thresh=0.7, min_size=stride, nms_pre_topn=500, nms_post_topn=100,
name ='nms')
#<todo> feature = upsample2d(block, factor = 4, ...)
feature = block
print ('top: scale=%f, stride=%d'%(1./stride, stride))
return feature, scores, probs, deltas, rois, roi_scores, stride, top_states
