Tom Breloff

Please do benchmark. In theory it should be very close in computation to a handcrafted version. I started ObjectiveFunctions last night with something similar to your code Alex, so we're...

Thinking out loud... if a Transformation has a `params(trans)` method defined which returns a Vector/view/CatView of parameters, and similar a `grad(trans)` for the vector of gradients, then we can avoid...

> Is grad!(objfunc::RegularizedObjective) missing data arguments? Sort of. I forgot to put it `

> How do tree based learners fit into this? Well, they may not be differentiable, so the `grad!` method may have to throw an error? The `transform!` method would just...

Yes, though I'm not sure its ready for prime time yet. I want to use it a bit and maybe make some changes. I was thinking about a second "trace_iter"...

I should note that this issue might be more appropriate in ObjectiveFunctions.jl, but we can have the discussion here

Thanks for the comment @mlubin. Do you think that's true for all types of functions? Dense and Sparse? What would you say are the biggest limitations of the package (aside...

Thanks Christof. Espresso is really similar to what I was building in Transformations. I have a lot to review!! On Monday, August 29, 2016, Christof Stocker [email protected] wrote: > also...

@jrevels "Toy problems" sounds worse than I intended. I mean more that I'd like a solution that can be scaled to complex deep learning models with many free parameters. One...

Just trying to make sense of the paper. It seems like they are proposing the gradient update for weight `w[i]` is: ``` θ[i] -= lr * (∇[i] - V) ```...