> If this is just an input feature that does not need to be predicted and you are adding it similarly to the other input features such as node type, then it should work. If you are getting nans, it is probably because of a different reason (perhaps the feature has a very bad scale). Is the loss nan even for the very first iteration?

> If this is just an input feature that does not need to be predicted and you are adding it similarly to the other input features such as node type, then it should work. If you are getting nans, it is probably because of a different reason (perhaps the feature has a very bad scale). Is the loss nan even for the very first iteration?
>
> If it is not the scale, or something like that, then I am afraid it will be very hard for us to say what me be wrong.

Yeah, it happens in the first iteration.
When I debug it, this error is caused by the minus variance in the normalization of the solid concentration, leading to a Nan standard deviation. I guess the cause of this error is the python float number mechanism. For example, when I input 0.0414, the actual number obtained by the python is 0.04140000005.

_Originally posted by @yq60523 in https://github.com/deepmind/deepmind-research/issues/267#issuecomment-1061400898_