bobo0810 / pytorch-yolov3-master

YOLOV3

Hi there 👋

任职

• BOSS直聘 视觉算法工程师

博客

研究领域

2023

• 大模型
• 多模态
• 预训练

2022

• 人脸相关
  • 人脸检测
  • 人像比对
• 图像内容安全
  • 涉黄涉政
  • 涉暴涉恐
  • 图像检索

pytorch-yolov3-master's Issues

RuntimeError: cannot perform reduction function max on tensor with no elements because the operation does not have an identity

当我调试时，这里不能被执行
tcls[mask] = tensor([], device='cuda:0', size=(0, 80), dtype=torch.int64).
在models.py第202行
loss_cls = (1 / nB) * self.ce_loss(pred_cls[mask], torch.argmax(tcls[mask], 1))
请问这里是怎么回事呢，你遇到过这个问题吗？

为啥box中心位置回归用交叉熵loss呢

PyTorch-YOLOv3-master/models.py

Line 238 in 8bc36be

loss_x = self.lambda_coord * self.bce_loss(x * mask, tx * mask)

这一行是计算中心坐标x的loss把，不是应该用mse loss吗？

Why use BCEloss compute loss of x y

this code
loss_x = self.lambda_coord * self.bce_loss(x * mask, tx * mask)
loss_y = self.lambda_coord * self.bce_loss(y * mask, ty * mask)
loss_w = self.lambda_coord * self.mse_loss(w * mask, tw * mask) / 2
loss_h = self.lambda_coord * self.mse_loss(h * mask, th * mask) / 2
Should it be
loss_x = self.lambda_coord * self.mse_loss(x * mask, tx * mask)
loss_y = self.lambda_coord * self.mse_loss(y * mask, ty * mask)
loss_w = self.lambda_coord * self.mse_loss(w * mask, tw * mask) / 2
loss_h = self.lambda_coord * self.mse_loss(h * mask, th * mask) / 2

关于网络的训练问题

你好，我在服务器上训练这个模型，但是没有训练完就中断了，请问怎么在上一次中断的地方接着训练？运行train.py文件都是使用的默认参数，可能是因为网络原因训练到第54个epoch时就中断了。望回答，谢谢

---- [Epoch 54/100, Batch 10156/14658] ----
+------------+--------------+--------------+--------------+
| Metrics    | YOLO Layer 0 | YOLO Layer 1 | YOLO Layer 2 |
+------------+--------------+--------------+--------------+
| grid_size  | 10           | 20           | 40           |
| loss       | 2.769155     | 2.100723     | 2.016298     |
| x          | 0.068778     | 0.048943     | 0.051337     |
| y          | 0.029317     | 0.046944     | 0.065335     |
| w          | 0.874106     | 0.218711     | 0.167731     |
| h          | 0.290770     | 0.107057     | 0.083462     |
| conf       | 1.464406     | 1.641079     | 1.612683     |
| cls        | 0.041778     | 0.037989     | 0.035749     |
| cls_acc    | 46.34%       | 41.38%       | 39.13%       |
| recall50   | 0.268293     | 0.241379     | 0.246377     |
| recall75   | 0.146341     | 0.103448     | 0.173913     |
| precision  | 0.366667     | 0.297872     | 0.171717     |
| conf_obj   | 0.598845     | 0.541579     | 0.518112     |
| conf_noobj | 0.005811     | 0.005290     | 0.004984     |
+------------+--------------+--------------+--------------+
Total loss 6.886176109313965
---- ETA 0:37:04.094678

Socket error Event: 32 Error: 10053.
Connection closing...Socket close.

Connection closed by foreign host.

Disconnected from remote host(center) at 10:18:00.

iou_thres和nms_thres有什么区别呢