急！！！一个关于faster rcnn的代码指导

本人在写一个faster rcnn的代码改进，在进行多尺度融合时，出现bug，求大神在线指导。

改进：在faster rcnn上，进行简化的ION、FPN改进，

问题在进行多个rpn时，出现bug。

若代码通过，必谢

测试下回复是否有问题

？

请问怎么通过rpn获取多层的roi？

  with tf.variable_scope(self._resnet_scope, self._resnet_scope):

    conv_list = [net_conv3 , net_conv4 , net_conv5]

    rpn_cls_score_list = []

    rpn_cls_prob_list = []

    rpn_bbox_pred_lst = []

    for i in range(0 , 3):

      # build the anchors for the image

      self._anchor_component()      

      # rpn

      rpn = slim.conv2d(conv_list[i], 512, [3, 3], trainable=is_training, weights_initializer=initializer,scope="rpn_conv"+str(i) +"/3x3")

      self._act_summaries.append(rpn)

      rpn_cls_score = slim.conv2d(rpn, self._num_anchors * 2, [1, 1], trainable=is_training,

                    weights_initializer=initializer,

                    padding='VALID', activation_fn=None, scope='rpn_cls_score'+str(i))

      # change it so that the score has 2 as its channel size

      rpn_cls_score_reshape = self._reshape_layer(rpn_cls_score, 2, 'rpn_cls_score_reshape'+str(i))

      rpn_cls_prob_reshape = self._softmax_layer(rpn_cls_score_reshape, "rpn_cls_prob_reshape"+str(i))

      rpn_cls_prob = self._reshape_layer(rpn_cls_prob_reshape, self._num_anchors * 2, "rpn_cls_prob"+str(i))

      rpn_bbox_pred = slim.conv2d(rpn, self._num_anchors * 4, [1, 1], trainable=is_training,

                    weights_initializer=initializer,

                    padding='VALID', activation_fn=None, scope='rpn_bbox_pred'+str(i))

      print(rpn_cls_score.shape)

      print(rpn_cls_prob)

      print(rpn_bbox_pred)

      rpn_cls_score_list.append(rpn_cls_score)

      rpn_cls_prob_list.append(rpn_cls_prob)

      rpn_bbox_pred_lst.append(rpn_bbox_pred)

    rpn_cls_score = tf.concat(rpn_cls_score_list, axis=0)

    rpn_cls_prob = tf.concat(rpn_cls_prob_list, axis=0)

    rpn_bbox_pred = tf.concat(rpn_bbox_pred_lst, axis=0)

    if is_training:

     rois, roi_scores = self._proposal_layer(rpn_cls_prob, rpn_bbox_pred, "rois")

     rpn_labels = self._anchor_target_layer(rpn_cls_score, "anchor")

     # Try to have a determinestic order for the computing graph, for reproducibility

     with tf.control_dependencies([rpn_labels]):

      rois, _ = self._proposal_target_layer(rois, roi_scores, "rpn_rois")

    else:

     if cfg.TEST.MODE == 'nms':

      rois, _ = self._proposal_layer(rpn_cls_prob, rpn_bbox_pred, "rois")

     elif cfg.TEST.MODE == 'top':

      rois, _ = self._proposal_top_layer(rpn_cls_prob, rpn_bbox_pred, "rois")

     else:

      raise NotImplementedError

https://zhuanlan.zhihu.com/p/31401237

能不能就我的代码，帮我解答一下，万分感谢