Skip loading parameter hm.2.weight, required shapetorch.Size([11, 256, 1, 1]), loaded shapetorch.Size([80, 256, 1, 1]).

[yao-zheng-yi]

My num_classes = 11 and I use command to train my own dataset

python main.py ctdet --exp_id paper_dataset --batch_size 8 --lr 1.25e-4 --gpus 0 --load_model ../models/ctdet_coco_dla_2x.pth

How can I do to solve the parameter skipping probem?

[V63A]

I think you don't need to solve this problem. It just let you know that the weights of head model can't load to your head model, but it can still train start from random value.

I trained with two categories data and I can't load the weights thoroughly either, but I got good results at the end.

[Dhagash4]

@V63A can you specify the parameters that you gave while running the main.py because I am trying to train and then run test.py it's giving strange results or can you tell me how to evaluate trained model

[V63A]

@V63A can you specify the parameters that you gave while running the main.py because I am trying to train and then run test.py it's giving strange results or can you tell me how to evaluate trained model

Hi, are you train on your own dataset? If so, then I think you can refer to https://github.com/xingyizhou/CenterNet/blob/master/readme/DEVELOP.md or search some other sites, because there are lots of codes should be changed. This is the sites I followed before, but it's written in Chinese.(https://blog.csdn.net/weixin_42634342/article/details/97756458)

[JinYuannn]

I suggest you use the pre-trained weights of backbone, and then use the initialization function provided in the code to initialize the weights of the head weights, and then train your model.