JavierAntoran

Thanks! I encountered the same issue and this fix worked for me!

The method you mention computes the Hessian of the softmax-cross entropy layer. You can find the expression in appendix A.2 of this paper: https://arxiv.org/pdf/1905.12558.pdf.

Hi @sujanme25, The training scripts should autodetect whether you are on CPU or GPU. The code should work on both out of the box. If you encounter that this is...

Thanks for flagging this! It looks to be a pandas version issue. We are looking into it.

Hi @danielkelshaw, Thanks for your question. Similarly to the regular reparametrisation trick, the [local reparametrisation trick](https://arxiv.org/abs/1506.02557) is used to estimate gradients with respect to parameters of a distribution. However, the...

For general ideas about re-casting learning as inference, I would check out chapter 41 of David MacKay's [Information Theory, Inference, and Learning Algorithms](https://www.inference.org.uk/itprnn/book.pdf). Yarin Gal's [thesis](http://mlg.eng.cam.ac.uk/yarin/thesis/thesis.pdf) is also a good...

Hi @alwaysuu, Yes, changing the batch size to 1 will result in different weights being used for each input. However, it could make training very slow due to large variance...

`log_gaussian_loss` is just the log of the [Gaussian probability density](https://en.wikipedia.org/wiki/Normal_distribution). You will see that this looks like a squared error loss with a regularisation term which is just the log...

Hi @EvannaB 1. I believe that all models should train in minibatches by default. 2. You can remove the MC dropout layer, but you will remove the uncertainty estimation capabilities...

Hi, The short answer is no. With this code you can generate some signals which given some additional encoding software (which is not in this repo) you could potentially transmit...