demo cannot work

[linchunmian]

各位好，我按教程配好环境，然后跑了下demo，可是运行了images文件下所有的图片，结果都是一样的，预测的矩形框位置都是在图片上方，这是什么原因呢？

[xingyizhou]

This is because the last layer of the model is not properly loaded. Have you seen any skip loading xxx information in the log? If so, please make sure your --arch matches the model you loaded. See readme for details.

[linchunmian]

This is because the last layer of the model is not properly loaded. Have you seen any skip loading xxx information in the log? If so, please make sure your --arch matches the model you loaded. See readme for details.

Thanks. I check the detailed information in readme, again.
(1) firstly, I use ctdet_coco_dla_2x for object detection, and run 'python demo.py ...'
(2) It seems not to occur any skip loading xxx information in the log file, which confuses me more.
I follow the installation and get_start process step by step, but stuck in it.
What need I do next, or how do I check this problem?

[xingyizhou]

Can you show the exact command you run?

[linchunmian]

Can you show the exact command you run?
It is
'python demo.py ctdet --demo ../images --load_model ../ctdet_coco_dla_2x.pth'

[xingyizhou]

That should be correct. Can you show the full log as well? Also, please make sure you haven't modified any code.

[linchunmian]

Thanks for reply!
When I rerun the demo.py file, the log information is as follow:

It indeed skips loading parameter, but I never modify any code.
So, How can I do to solve the parameter skipping probem?
Any help you provided would be appreciated!

[xingyizhou]

That's strange, can you try re-download and re-install the code? It shows your model uses a last later of 12 categories, which is never used in my version. An easier way is to add --num_classes 80 in the command, but I really have no idea why it is 12 categories by default.

[topdou96]

你好，在CenterNet/src/lib/opts.py 里把 hm参数改成和你对应的大小（也就是把12改成80）就可以了

[topdou96]

你好，在CenterNet/src/lib/opts.py 里把 hm参数改成和你对应的大小（也就是把12改成80）就可以了。
hm参数也就是num_classes 例如如果你用的是ctdet，就把ctdet那一栏的‘num_classes’从12改成80

[linchunmian]

@xingyizhou @topdou96
Sorry for delay, and thanks very much for your help!
I have run demo sucessfully!

[linchunmian]

It is great job. Though coco-pretrained model weight is used for my own image, it still presents excellent detection performance!

[hlq1025]

This is because the last layer of the model is not properly loaded. Have you seen any skip loading xxx information in the log? If so, please make sure your --arch matches the model you loaded. See readme for details.

I am new to centernet.Could you tell me the backbone of dddnet of the pretrained model?