Function 'AddmmBackward0' returned nan values in its 1th output

reinforcement-learning
1

I have simple NN network which is consist of convolution layer followed by Linear layer to train RL agent using REINFORCE algorithm. I use

torch.optim.AdamW, for optimizer
nn.init.kaiming_normal_(m.weight, nonlinearity=‘relu’) for initialization

Although I set the torch.manual_seed(seed), numpy seed and random seed, I sometimes get the following error:

Function ‘AddmmBackward0’ returned nan values in its 1th output.

I have couple of question:
(1) nan value in its 1th output means there is problem in the weights, or the ipnput?
(2) Even when I have seed, Why do I sometimes get the error and sometimes my program successfully ends with reproducible output?
(3) How can I fix it?

I appreciate your help

2

In my case it was an exploding loss calculation.
Post your code entirely, otherwise it is almost impossible to help you