cost_op = rpn_bbox_loss_op + rpn_cls_loss_op
rpn_cls_train_op = optimizer(rpn_cls_loss_op)
rpn_bbox_train_op = optimizer(rpn_bbox_loss_op )
train_op = optimizer(rpn_bbox_loss_op )
sess=sess_start()
for i in range(2,55000 * 100):
src_img , src_gt_boxes =next_img_gtboxes(i)
h,w=np.shape(src_img)
src_im_dims = [(h,w)]
anchor_scales = [3,4,5]
rpn_cls_score=np.zeros([1,int(math.ceil(h/8.)),int(math.ceil(w/8.)),512])
rpn_labels, rpn_bbox_targets, rpn_bbox_inside_weights, rpn_bbox_outside_weights, bbox_targets, bbox_inside_weights, bbox_outside_weights = anchor_target(
rpn_cls_score=rpn_cls_score, gt_boxes=src_gt_boxes, im_dims=src_im_dims, _feat_stride=8, anchor_scales=anchor_scales)
"""
rpn_labels : (1, 1, h*9 , w)
rpn_bbox_targets : (1, 4*9, h, w)
rpn_bbox_inside_weights : (1, 4*9, h, w)
rpn_bbox_outside_weights: (1, 4*9, h, w)
"""
"""
# delete me
n, ch, h, w = np.shape(rpn_bbox_targets)
rpn_bbox_targets = np.reshape(rpn_bbox_targets , [1,4, h*9 , w])
rpn_bbox_targets = np.transpose(rpn_bbox_targets , [0,2,3,1])
rpn_bbox_targets = np.reshape(rpn_bbox_targets, [-1 , 4])
for rpn_bbox in rpn_bbox_targets:
def eval(self, img, src_gt_boxes, src_im_dims, _feat_stride,
anchor_scales):
assert len(np.shape(img)) == 4 and np.shape(img)[0] == 1
rpn_cls_score = np.zeros([
1,
int(math.ceil(h / float(_feat_stride))),
int(math.ceil(w / float(_feat_stride))), 512
])
rpn_labels, rpn_bbox_targets, rpn_bbox_inside_weights, rpn_bbox_outside_weights, bbox_targets, bbox_inside_weights, bbox_outside_weights = anchor_target(
rpn_cls_score=rpn_cls_score,
gt_boxes=src_gt_boxes,
im_dims=src_im_dims,
_feat_stride=_feat_stride,
anchor_scales=anchor_scales)
indices = np.where([np.reshape(rpn_labels, [-1]) > 0])[1]
feed_dict = {
self.x_: src_img,
self.im_dims: src_im_dims,
self.gt_boxes: src_gt_boxes,
self.phase_train: True,
self.rpn_labels_op: rpn_labels,
self.bbox_targets_op: bbox_targets,
self.bbox_inside_weights_op: bbox_inside_weights,
self.bbox_outside_weights_op: bbox_outside_weights,
self.indice_op: indices
}
roi_blobs, roi_scores = self.sess.run(
[self.roi_blobs_op, self.roi_scores_op], feed_dict=feed_dict)
# draw pos , neg rectangles and
roi_scores = np.reshape(roi_scores, [-1, 1])
keep = non_maximum_supression(dets=np.hstack([roi_blobs, roi_scores]),
thresh=0.7)
nms_roi_blobs = roi_blobs[keep]
nms_roi_scores = roi_scores[keep]
print 'NMS before {} ===> after {}'.format(len(roi_blobs),
len(nms_roi_blobs))
draw_rectangles(src_img,
nms_roi_blobs,
nms_roi_scores,
src_gt_boxes,
fname='{}.png'.format('tmp'))
src_img, src_gt_boxes = next_img_gtboxes(i)
h, w = np.shape(src_img)[:2]
src_im_dims = [(h, w)]
# rpn labels RPN Network 을 학습시킬 Label 이 있다 foreground 이면 1 , backgound = 0 , 학습 시키지 않을 box에 대해서는 -1 을 부여한다
# rpn_bbox_targets : anchor 들을 target 으로 변환했다
# src im dims : source image 의 원래 크기
# _feat stride : 얼마만큼 줄였는지
# _anchor scale : anchor scale
rpn_cls_score = np.zeros([
1,
int(math.ceil(h / float(_feat_stride))),
int(math.ceil(w / float(_feat_stride))), 512
])
rpn_labels, rpn_bbox_targets, rpn_bbox_inside_weights, rpn_bbox_outside_weights, bbox_targets, bbox_inside_weights, bbox_outside_weights = \
anchor_target(rpn_cls_score, src_gt_boxes, src_im_dims, _feat_stride, anchor_scales)
indices = np.where([np.reshape(rpn_labels, [-1]) > 0])[1]
src_img = src_img.reshape([1] + list(np.shape(src_img)) + [1])
feed_dict = {
x_: src_img,
im_dims: src_im_dims,
gt_boxes: src_gt_boxes,
phase_train: True,
rpn_labels_op: rpn_labels,
bbox_targets_op: bbox_targets,
bbox_inside_weights_op: bbox_inside_weights,
bbox_outside_weights_op: bbox_outside_weights,
indice_op: indices
}
def _prepare(self, img, src_gt_boxes, src_im_dims, _feat_stride,
anchor_scales):
# save topconv caches
assert len(np.shape(img)) == 4 and np.shape(img)[0] == 1
rpn_cls_score = np.zeros([
1,
int(math.ceil(h / float(_feat_stride))),
int(math.ceil(w / float(_feat_stride))), 512
])
rpn_labels, rpn_bbox_targets, rpn_bbox_inside_weights, rpn_bbox_outside_weights, bbox_targets, bbox_inside_weights, bbox_outside_weights = anchor_target(
rpn_cls_score=rpn_cls_score,
gt_boxes=src_gt_boxes,
im_dims=src_im_dims,
_feat_stride=_feat_stride,
anchor_scales=anchor_scales)
indices = np.where([np.reshape(rpn_labels, [-1]) > 0])[1]
feed_dict = {
self.x_: src_img,
self.im_dims: src_im_dims,
self.gt_boxes: src_gt_boxes,
self.phase_train: True,
self.rpn_labels_op: rpn_labels,
self.bbox_targets_op: bbox_targets,
self.bbox_inside_weights_op: bbox_inside_weights,
self.bbox_outside_weights_op: bbox_outside_weights,
self.indice_op: indices
}
return feed_dict