mmpose
mmpose copied to clipboard
open-mmlab
When I use my own dataset，the loss is 0
Reimplement a custom model but all the components are implemented in MMPose. My picture‘s information like this：
Used the config like this:
but the loss is 0
When i use the dataset like this, it is successful.
**
**
i want to know how to change the config or change something else. Thank you.
@zhzh1234 Hi, according to your screenshots, I think you want to change the size of input images, which is not likely to result in 0 loss. Could you please check if your annotations align with 1024x576 shaped images?
Thanks for your reply，my annotations like this：
I used “labelme” to label my images，
I can't locate the error baesd on information you share. I suggest printing out your model predictions and targets when calculating loss to check why loss become 0
Thank you，And I want to know how to printing out my model predictions and targets when calculating loss.
It depends on what kind of algorithm and loss you are using. For example, if you're using JointMSELoss, you can just print output and target here:
https://github.com/open-mmlab/mmpose/blob/efd9674e2a7ecc8f6b94047468709d681cc85479/mmpose/models/losses/mse_loss.py#L25
to check the prediction and training target.
If you're using some IDE like Pycharm, it's more convenient to use debugger to check the values of output and target
