François CholletVerified account

@egrefen @TwitterEng I hardly see how you could do that more with original nn in Torch...

@hugo_larochelle @TwitterEng Want to write a nonstandard module? Gotta figure out the backward pass dynamics.

@egrefen @TwitterEng ah I see, well you can still wrap new modules as its own autograd operation, and benefit from autograd outside of it

@hugo_larochelle @TwitterEng Fair enough. Just think I (and probably others) would miss this sort of thing if not forced to look at it

@egrefen @hugo_larochelle @TwitterEng early optimization is the root of all evil, right? We should look at grads like a profiling step.

@edersantana @dwf @hugo_larochelle @twittereng Ignore point about optim. Still think analysis and understanding of backwards pass is crucial

@egrefen @edersantana @hugo_larochelle For implementation, though, it's two more ways my clumsy hands can fail me. :)

@dwf @edersantana @hugo_larochelle Would you crank through algebra if you didn't *need* to, e.g. with autograd? If so, props. I might not :)