3.3 passthrough操作
regorg layer分析：这里ReorgLayer层就是将26∗26∗512的张量中26∗26切割成4个13∗13，然后连接起来，使得原来的512通道变成了2048。
1 #darknet.py 2 self.reorg = ReorgLayer(stride=2) # stride*stride times the channels of conv1s
1 #reorg_layer.py 2 def forward(self, x): 3 stride = self.stride 4 5 bsize, c, h, w = x.size() 6 out_w, out_h, out_c = int(w / stride), int(h / stride), c * (stride * stride) 7 out = torch.FloatTensor(bsize, out_c, out_h, out_w) 8 9 if x.is_cuda: 10 out = out.cuda() 11 reorg_layer.reorg_cuda(x, out_w, out_h, out_c, bsize, stride, 0, out) 12 else: 13 reorg_layer.reorg_cpu(x, out_w, out_h, out_c, bsize, stride, 0, out) 14 15 return out
1 //reorg_cpu.c 2 int reorg_cpu(THFloatTensor *x_tensor, int w, int h, int c, int batch, int stride, int forward, THFloatTensor *out_tensor) 3 { 4 // Grab the tensor 5 float * x = THFloatTensor_data(x_tensor); 6 float * out = THFloatTensor_data(out_tensor); 7 8 // https://github.com/pjreddie/darknet/blob/master/src/blas.c 9 int b,i,j,k; 10 int out_c = c/(stride*stride); 11 12 for(b = 0; b < batch; ++b){ 13 //batch_size 14 for(k = 0; k < c; ++k){ 15 //channel 16 for(j = 0; j < h; ++j){ 17 //height 18 for(i = 0; i < w; ++i){ 19 //width 20 int in_index = i + w*(j + h*(k + c*b)); 21 int c2 = k % out_c; 22 int offset = k / out_c; 23 int w2 = i*stride + offset % stride; 24 int h2 = j*stride + offset / stride; 25 int out_index = w2 + w*stride*(h2 + h*stride*(c2 + out_c*b)); 26 if(forward) out[out_index] = x[in_index]; // 压缩channel 27 else out[in_index] = x[out_index]; // 扩展channel 28 } 29 } 30 } 31 } 32 33 return 1; 34 }
图片有错误，待改，输入的1,3点分布在输出的第1个feature map上，输入的2,4点分布在输出的第2个feature map上，idx2后面+w2
下图从右到左为forward计算方向，从左到右为backward求导方向
3.4 目标函数计算
1 #darknet.py 2 def loss(self): 3 #可以看出，损失值也是基于预测框bbox，预测的iou，分类三个不同的误差和 4 return self.bbox_loss + self.iou_loss + self.cls_loss 5 6 def forward(self, im_data, gt_boxes=None, gt_classes=None, dontcare=None): 7 conv1s = self.conv1s(im_data) 8 conv2 = self.conv2(conv1s) 9 conv3 = self.conv3(conv2) 10 conv1s_reorg = self.reorg(conv1s) 11 cat_1_3 = torch.cat([conv1s_reorg, conv3], 1) 12 conv4 = self.conv4(cat_1_3) 13 conv5 = self.conv5(conv4) # batch_size, out_channels, h, w 14 …… 15 …… 16 # tx, ty, tw, th, to -> sig(tx), sig(ty), exp(tw), exp(th), sig(to) 17 '''预测tx ty''' 18 xy_pred = F.sigmoid(conv5_reshaped[:, :, :, 0:2]) 19 '''预测tw th ''' 20 wh_pred = torch.exp(conv5_reshaped[:, :, :, 2:4]) 21 bbox_pred = torch.cat([xy_pred, wh_pred], 3) 22 '''预测置信度to ''' 23 iou_pred = F.sigmoid(conv5_reshaped[:, :, :, 4:5]) 24 '''预测分类class ''' 25 score_pred = conv5_reshaped[:, :, :, 5:].contiguous() 26 prob_pred = F.softmax(score_pred.view(-1, score_pred.size()[-1])).view_as(score_pred) 27 28 # for training 29 if self.training: 30 bbox_pred_np = bbox_pred.data.cpu().numpy() 31 iou_pred_np = iou_pred.data.cpu().numpy() 32 _boxes, _ious, _classes, _box_mask, _iou_mask, _class_mask = self._build_target( 33 bbox_pred_np, gt_boxes, gt_classes, dontcare, iou_pred_np) 34 _boxes = net_utils.np_to_variable(_boxes) 35 _ious = net_utils.np_to_variable(_ious) 36 _classes = net_utils.np_to_variable(_classes) 37 box_mask = net_utils.np_to_variable(_box_mask, dtype=torch.FloatTensor) 38 iou_mask = net_utils.np_to_variable(_iou_mask, dtype=torch.FloatTensor) 39 class_mask = net_utils.np_to_variable(_class_mask, dtype=torch.FloatTensor) 40 41 num_boxes = sum((len(boxes) for boxes in gt_boxes)) 42 43 # _boxes[:, :, :, 2:4] = torch.log(_boxes[:, :, :, 2:4]) 44 box_mask = box_mask.expand_as(_boxes) 45 #计算预测的平均bbox损失值 46 self.bbox_loss = nn.MSELoss(size_average=False)(bbox_pred * box_mask, _boxes * box_mask) / num_boxes 47 #计算预测的平均iou损失值 48 self.iou_loss = nn.MSELoss(size_average=False)(iou_pred * iou_mask, _ious * iou_mask) / num_boxes 49 #计算预测的平均分类损失值 50 class_mask = class_mask.expand_as(prob_pred) 51 self.cls_loss = nn.MSELoss(size_average=False)(prob_pred * class_mask, _classes * class_mask) / num_boxes 52 53 return bbox_pred, iou_pred, prob_pred
参考自：仙守