def roi_pool(self, input, pooled_height, pooled_width, spatial_scale, name):
"""
input:'conv5_3', 'roi-data'
roi-data:
(1) rois: (num of finally left proposal（根据每个roi和gtbox的overlaps大小来确定留下的fg和bg，多了的随机选择) blob[:,0]=0;
blob[:-2,1:5] = x1,y1,x2,y2(pred box in original image); blob[-2:,1:5] = x1,y1,x2,y2(gt_box)
[0:fg_rois_per_this_image（大约42）]: the left foregound; [fg_rois_per_this_image（128-42）:]:the left background
(2) labels: the final classes of the ground truth correspounding to per pred box [num of finally left proposal（128）,]
for ex: [9,15,15,15,9,9....]
(3): num of finally left proposals(128) * 4*21: [dx,dy,dw,dh](gt_boxes相对于rois) of 1 class
(4): num of finally left proposals * 4*21: [1, 1, 1, 1] of 1 class
(5): num of finally left proposals * 4*21: [true,true,true,true] of 1 class in 21; [false, false, false, false] in left of the classes
roi_pool(7, 7, 1.0/16, name='pool_5')
"""
# only use the first input
if isinstance(input[0], tuple):
input[0] = input[0][0]
if isinstance(input[1], tuple):
input[1] = input[1][0]
print input
return roi_pool_op.roi_pool(input[0], input[1],
pooled_height,
pooled_width,
spatial_scale,
name=name)[0]
def roi_pool(self, input, pooled_height, pooled_width, spatial_scale,
name):
# only use the first input
if isinstance(input[0], tuple):
input[0] = input[0][0]
""""# often called with: .roi_pool(7, 7, 1.0/16, name='pool_5')
print('................ roi pooling (network.py) .......................')
print(input[1][0]) # rois_wo
print(input[1][1]) # labels
print(input[1][2])
print(input[1][3])
print(input[1][4])
print(input[1][5]) # gt_orientations_new
print(input[1][6]) # orientations_pred_new
print(input[1][7]) # orientations_targets_new"""
if isinstance(input[1], tuple):
input[1] = input[1][0]
input_1a = input[1][:, 0:5] # rpn_rois as original, no orientations
input_1b = input[1][:, 5:6] # orientations as predicted
return roi_pool_op.roi_pool(input[0],
input_1a,
input_1b,
pooled_height,
pooled_width,
spatial_scale,
name=name)[0]
def _roi_pool_layer(self, bootom, rois, name):
print('ROI Pooling...')
with tf.variable_scope(name) as scope:
return roi_pool_op.roi_pool(bootom,
rois,
pooled_height=cfg.POOLING_SIZE,
pooled_width=cfg.POOLING_SIZE,
spatial_scale=1. / 16.,
name=name)[0]
def roi_pool(self, input, pooled_height, pooled_width, spatial_scale, pool_channel, name):
# only use the first input
if isinstance(input[0], tuple):
input[0] = input[0][0]
return roi_pool_op.roi_pool(input[0], input[1],
pooled_height,
pooled_width,
spatial_scale,
pool_channel,
name=name)
def load_detection_model(ws):
assert len(ws) == 8
data_blob = tf.placeholder(tf.float32)
rois_blob = tf.placeholder(tf.float32, shape=[None, 5])
pool5, _ = roi_pooling_op.roi_pool(data_blob, rois_blob, 7, 7, 0.0625)
flat_pool5 = tf.reshape(pool5, [-1, 25088])
fc6 = fc(flat_pool5, ws[0], ws[1])
fc7 = fc(fc6, ws[2], ws[3])
cls_prob = fc(fc7, ws[4], ws[5], 'softmax')
bbox_pred = fc(fc7, ws[6], ws[7], 'linear')
return data_blob, rois_blob, cls_prob, bbox_pred
def roi_pool(self, input, pooled_h, pooled_w, spatial_scale, name):
#only use the first input? what's the meaning
if isinstance(input[0], tuple):
input[0] = input[0][0]
if isinstance(input[1], tuple):
input[1] = input[1][0]
# print("roi_pool is :",input)
return roi_pool_op.roi_pool(input[0],
input[1],
pooled_h,
pooled_w,
spatial_scale,
name=name)[0]
def roi_pool(self, input, pooled_height, pooled_width, spatial_scale, name):
# only use the first input
if isinstance(input[0], tuple):
input[0] = input[0][0]
if isinstance(input[1], tuple):
input[1] = input[1][0]
print input
return roi_pool_op.roi_pool(input[0], input[1],
pooled_height,
pooled_width,
spatial_scale,
name=name)[0]
def roi_pool(self, input, pooled_height, pooled_width, spatial_scale, name):
"""
Rol pooling层有2个输入：
1. 原始的feature maps
2. RPN输出的proposal boxes（大小各不相同）
"""
# only use the first input
if isinstance(input[0], tuple):
input[0] = input[0][0]
if isinstance(input[1], tuple):
input[1] = input[1][1]
print input
return roi_pool_op.roi_pool(input[0], input[1], pooled_height, pooled_width, spatial_scale, name = name)[0]
def roi_pool(self, inputs, pooled_height, pooled_width, spatial_scale, name):
with tf.name_scope(name):
# only use the first input
if isinstance(inputs[0], tuple):
inputs[0] = inputs[0][0]
if isinstance(inputs[1], tuple):
inputs[1] = inputs[1][0]
print(inputs)
return roi_pool_op.roi_pool(inputs[0], inputs[1],
pooled_height,
pooled_width,
spatial_scale,
name=name)[0]
def frcnn_roi_pooling(self, features, rois, pooled_height, pooled_width, \
spatial_scale, name = "pool_5"):
'''
input:
features : Nbatch , h , w , C
rois : NROI x 5
output:
pooled_features: NROI x 7 x 7 x C
'''
return roi_pool_op.roi_pool(features,
rois,
pooled_height,
pooled_width,
spatial_scale,
name=name)[0]
def roi_pool(self, input, pooled_height, pooled_width, spatial_scale,
name):
# only use the first input
if isinstance(input[0], tuple):
input[0] = input[0][0]
if isinstance(input[1], tuple):
input[1] = input[1][0]
# Input[0] is the Feature map and Input[1] is the ROI data
print(input)
return roi_pool_op.roi_pool(input[0],
input[1],
pooled_height,
pooled_width,
spatial_scale,
name=name)[0]
def roi_pool(self, input, pooled_height, pooled_width, spatial_scale,
name):
# only use the first input ??? 因为 roi-data[0]才是rois,其他索引对应的是别的信息
# pooled_height=7；pooled_width=7；spatial_scale=1.0/16
if isinstance(input[0], tuple):
input[0] = input[0][0]
if isinstance(input[1], tuple):
input[1] = input[1][0]
print input
# input[0]是'conv5_3'[0]； input[1]是'roi-data'[0]
return roi_pool_op.roi_pool(input[0],
input[1],
pooled_height,
pooled_width,
spatial_scale,
name=name)[0]
def roi_pool(self, input, pooled_height, pooled_width, spatial_scale,
name):
# only use the first input
if isinstance(input[0], tuple):
input[0] = input[0][0]
if isinstance(input[1], tuple):
input[1] = input[1][0]
print "RoI input[0] = ", input[0], " input [1]= ", input[1],
output = roi_pool_op.roi_pool(input[0],
input[1],
pooled_height,
pooled_width,
spatial_scale,
name=name)[0]
print "RoI output =", output
return output