Faster R-CNN faster_rcnn source code file folder contains three faster_rcnn.py, resnet.py , vgg16.py .
1.faster_rcnn.py comments
. 1 class _fasterRCNN (nn.Module): 2 "" " Faster RCNN " "" . 3 DEF the __init__ (Self, classes, class_agnostic): # class-agnostic was fittingly regression-based method is only 2 bounding box, i.e. the foreground and background . 4 Super (_fasterRCNN, Self). the __init__ () . 5 self.classes classes = # type . 6 self.n_classes = len (classes) # number of categories . 7 self.class_agnostic = class_agnostic # foreground background class . 8 # Loss two kinds Loss . 9 self.RCNN_loss_cls = 0 10 self.RCNN_loss_bbox = 0 11 12 # define rpn 定义RPN网络 13 self.RCNN_rpn = _RPN(self.dout_base_model) 14 self.RCNN_proposal_target = _ProposalTargetLayer(self.n_classes)#候选区域对应gt 15 self.RCNN_roi_pool = _RoIPooling(cfg.POOLING_SIZE, cfg.POOLING_SIZE, 1.0/16.0)#POOLING 16 self.RCNN_roi_align = RoIAlignAvg(cfg.POOLING_SIZE, cfg.POOLING_SIZE, 1.0/16.0) 17 18 self.grid_size = cfg.POOLING_SIZE * 2 if cfg.CROP_RESIZE_WITH_MAX_POOL else cfg.POOLING_SIZE . 19 self.RCNN_roi_crop = _RoICrop () 20 is 21 is DEF Forward (Self, im_data, im_info, gt_boxes, num_boxes): # image annotation information block number information 22 is the batch_size = im_data.size (0) 23 is 24 im_info = im_info.data 25 gt_boxes = gt_boxes.data 26 is num_boxes = num_boxes.data 27 28 # feed to image data base model base to obtain feature Map 29 # fed to the image data base model to obtain the basic features of FIG. 30 base_feat = self.RCNN_base (im_data) 31 is 32 # Feed Base to Obtain Feature Map TP RPN ROIs 33 is # feature back to FIG RPN obtained ROIS 34 is ROIs, rpn_loss_cls, rpn_loss_bbox = self.RCNN_rpn (base_feat, im_info, gt_boxes, num_boxes) 35 36 # IF IT IS Training phrase, the then use Refining trubut bboxes for Ground 37 [ # if ground truth is trained using regression 38 is IF self.training: 39 roi_data = self.RCNN_proposal_target (ROIs, gt_boxes, num_boxes) 40 ROIs, rois_label, rois_target, rois_inside_ws, rois_outside_ws = roi_data 41 is 42 is rois_label = Variable(rois_label.view(-1).long()) 43 rois_target = Variable(rois_target.view(-1, rois_target.size(2))) 44 rois_inside_ws = Variable(rois_inside_ws.view(-1, rois_inside_ws.size(2))) 45 rois_outside_ws = Variable(rois_outside_ws.view(-1, rois_outside_ws.size(2))) 46 else: 47 rois_label = None 48 rois_target = None 49 rois_inside_ws = None 50 rois_outside_ws = None 51 rpn_loss_cls = 0 52 rpn_loss_bbox = 0 53 54 rois = Variable(rois) 55 # do roi pooling based on predicted rois 56 #进行ROI POOLING，下面pooling方式 57 58 if cfg.POOLING_MODE == 'crop': 59 # pdb.set_trace() 60 # pooled_feat_anchor = _crop_pool_layer(base_feat, rois.view(-1, 5)) 61 grid_xy = _affine_grid_gen(rois.view(-1, 5), base_feat.size()[2:], self.grid_size) 62 grid_yx = torch.stack([grid_xy.data[:,:,:,1], grid_xy.data[:,:,:,0]], 3).contiguous() 63 pooled_feat = self.RCNN_roi_crop(base_feat, Variable(grid_yx).detach()) 64 if cfg.CROP_RESIZE_WITH_MAX_POOL: 65 pooled_feat = F.max_pool2d(pooled_feat, 2, 2) 66 elif cfg.POOLING_MODE == 'align': 67 pooled_feat = self.RCNN_roi_align(base_feat, rois.view(-1, 5)) 68 elif cfg.POOLING_MODE == 'pool': 69 = self.RCNN_roi_pool pooled_feat (base_feat, rois.view (-1,5 )) 70 71 is # Feed to the pooled Features Top Model 72 # characteristics after pooling to the last feedback model 73 is pooled_feat = self._head_to_tail (pooled_feat) 74 75 # BBOX offset Compute 76 # calculates offset bounding box 77 bbox_pred = self.RCNN_bbox_pred (pooled_feat) 78 IF self.training and Not self.class_agnostic: 79 # SELECT the ROI to the Corresponding Columns ACCORDING Labels 80 #Selecting the appropriate column according roi tag 81 bbox_pred_view = bbox_pred.view (bbox_pred.size (0), int (bbox_pred.size (. 1) /. 4),. 4 ) 82 bbox_pred_select = torch.gather (bbox_pred_view,. 1, rois_label.view ( rois_label.size (0),. 1,. 1) .expand (rois_label.size (0),. 1,. 4 )) 83 bbox_pred bbox_pred_select.squeeze = (. 1 ) 84 85 # Compute Object classification probability 86 # computing objects classification probability 87 cls_score = self.RCNN_cls_score (pooled_feat) 88 cls_prob = F.softmax (cls_score,. 1 ) 89 90 RCNN_loss_cls = 0 91 RCNN_loss_bbox = 0 92 93 if self.training: 94 # classification loss 95 #分类损失 96 RCNN_loss_cls = F.cross_entropy(cls_score, rois_label) 97 98 # bounding box regression L1 loss 99 #回归损失 100 RCNN_loss_bbox = _smooth_l1_loss(bbox_pred, rois_target, rois_inside_ws, rois_outside_ws) 101 102 103 cls_prob = cls_prob.view(batch_size, rois.size(1), -1) 104 = bbox_pred.view bbox_pred (the batch_size, rois.size (. 1), -1 ) 105 106 return ROIs, cls_prob, bbox_pred, rpn_loss_cls, rpn_loss_bbox, RCNN_loss_cls, RCNN_loss_bbox, rois_label 107 108 # initialization weights 109 DEF _init_weights (Self): 110 DEF normal_init (m, mean, STDDEV, truncated = False): # sd 111 # truncated Normal random normal 112 "" " 113 weight initalizer:. truncated Normal and random Normal 114 " "" 115 # X IS A Parameter 1 16 IF truncated: 117 m.weight.data.normal_().fmod_(2).mul_(stddev).add_(mean) # not a perfect approximation 118 else: 119 m.weight.data.normal_(mean, stddev) 120 m.bias.data.zero_() 121 122 normal_init(self.RCNN_rpn.RPN_Conv, 0, 0.01, cfg.TRAIN.TRUNCATED) 123 normal_init(self.RCNN_rpn.RPN_cls_score, 0, 0.01, cfg.TRAIN.TRUNCATED) 124 normal_init(self.RCNN_rpn.RPN_bbox_pred, 0, 0.01, cfg.TRAIN.TRUNCATED) 125 normal_init(self.RCNN_cls_score, 0, 0.01, cfg.TRAIN.TRUNCATED) 126 normal_init(self.RCNN_bbox_pred, 0, 0.001, cfg.TRAIN.TRUNCATED) 127 128 def create_architecture(self): 129 self._init_modules() 130 self._init_weights()
ref:https://blog.csdn.net/weixin_43872578/article/details/87930953