Negative GMM loss. How to interpret?

[Fairasname]

Hi,

I am currently training the MDR-RNN with VAEs of different latent vector sizes (LSIZE). I have noticed that the smaller the size, the smaller the GMM loss (and total loss) is. Specifically, by using an LSIZE of 4 (and default RSIZE of 256), the loss goes below zero.

On the other hand, on the code comments of trainmdrnn.py, I saw that: "The LSIZE + 2 factor is here to counteract the fact that the GMMLoss scales approximately linearily with LSIZE". So I suppose that the fact that the loss goes below zero is somewhat expected when using a small LSIZE.

Nonetheless, I wonder how we could interpret and compare the MDR-RNN losses in order to asses its performance. Also, do you know if the original World Models implementation also does the "LSIZE+2" scaling / has the loss below zero effect? I read their code but could not figure if that was the case.

Thanks again!

[ctallec]

One possibility would be that it is much easier for the MDN-RNN to model a low-dimensional latent space than a high dimensional one. The scaling is only designed to make losses comparable, but won't compensate an easier or a harder task. In the original paper, they did not have this scaling factor because they did not model the reward signal, which we did by default in a previous version of the code, but no longer do. When there is no reward modelling contribution, RMSprop should directly negate the constant multiplicative factor that differs between the two losses.

[Fairasname]

Alright, thanks for the clarification!