Thomas Unterthiner

Ah, I see! The question is if there's a need for it (if someone has 3+ dimensions, they likely won't use skcuda anyhow)? In any case, I also think that...

Are you sure you aren't working on an outdated version? Using the current master branch, both of the following lines throw a ValueError on my machine: ``` print(skcuda.misc.add_matvec(x23, a2)) print(x23.get()...

This probably is the only way to make gradient checking work correctly, though. If we didn't add the constant 1/2, we'd have to multiply errors by 2 during backprop to...

In my opinion full blown logging (á la `logging` module in the stdlib) adds more complexity that we need for now

I think a basic implementation could just add a `destination`/`sink`/`file` (which one do you prefer?) parameter to the constructor, which would store the corresponding file-object, and then `net._add_log` could just...

1 and 3 aren't options because they don't allow you to monitor things while training happens (e.g. it's frustrating to wait for 2 days for a run to finish only...

The problem with having this option is that I might want to have some output displayed on screen (e.g. per-epoch accuracy or even per-epoch parameter statistics) and other output on...

That solution seems a bit hackish to me, I don't think it's very intuitive for new users. It also makes it impossible to log different hooks into different files.

Why not implement the pickle serialization methods (`__getstate__` and `__setstate__`) instead? It would have the advantage that you can then use a network whenever pickling is needed (e.g. when using...

Is pickle insecure?