关于使用自己的数据集进行训练的问题

[linklllllllll]

作者您好，首先非常感谢您的开源工作，我是使用culane数据集的格式进行训练，但是训练结束后在log_path中没有权重文件，且在训练过程中evaluating the results总是nearly no results，尝试了一些方法但仍然无果，故向您讨教，希望您能帮忙解答。 1.没有权重文件的情况具体是怎样才会产生的呢？ 2.在evaluating过程中在log_path/culane_eval_tmp/test_image中产成的图像为空，请问这是为什么呢？ 以下为训练过程中的部分显示; 100%|███████████| 20/20 [00:01<00:00, 11.24it/s, col_top1=0.003, col_top2=0.008, col_top3=0.009, ext_col=0.912, ext_row=0.825, loss=12.214, top1=0.007, top2=0.012, top3=0.021] 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 2.80it/s] ....... iou_threshold: 0.5 im_width: 1640 im_height: 590 x_factor: 1 y_factor: 1

Evaluating the results... tp: 0 fp: 0 fn: 0 no positive detection no ground truth positive finished process file precision: -1 recall: -1 Fmeasure: -1

res_normal -1.0 res_crowd -1.0 res_night -1.0 res_noline -1.0 res_shadow -1.0 res_arrow -1.0 res_hlight -1.0 res_curve -1.0 res_cross -1.0 nearly no results! nearly no results! 0 本人小白，训练模型仍是一知半解的状态，希望自己的疑惑能够得到解答

[cfzd]

@linklllllllll 你的学习率设置的太大了，导致网络根本没有收敛，所以测试结果为0。在保存模型时，我们会保存最好的那一个，因为结果为0，所以不会有模型被保存。可能需要减少学习率直至网络开始收敛。

[Hymilex]

@cfzd Hey,brother！I met the same problem.It seems like no effect after increasing lr to 0.0005.It makes me confused yet.

[cfzd]

@Hymilex It seems the img width and img height are modified. Incorrect config would make the evaluation fail.

[Hymilex]

@cfzd my custom dataset width and height is 416. I have used original value and modified tried,but both got failed.

[cfzd]

@Hymilex You can have a very simple way to verify the problem. For both prediction and annotation, there are stored in many xxx.lines.txt files. The annotation is like something anno_dir/xxx/xxx.lines.txt, and the prediciton is something like detect_dir/xxx/xxx.lines.txt. You can directly read some corresponding predictions and annotations, draw them on a canvas using python, and show them to compare them visually.