Hrmm, probably better to actually look at the root issue. Seems like on the mallocs we should check layer state. Although this patch will workaround the issue, we should probably have a more general debate on what the behavior should even be training with no learnable params.

Although this patch will workaround the issue, we should probably have a more general debate on what the behavior should even be training with no learnable params.

I agree. It doesn't make sense to me to train without learnable params (maybe someone just want to see the loss of a network? still doesn't quite make sense to me...). Perhaps train without learnable params should be banned in the future?

Since it is a simple 3 line workaround, I'll first merge this, and debates can follow in new issues.