Fabian Pedregosa

Thanks Matteo. We're almost there, but there's a test that has been commented out while I think its important to keep it, as it's testing a valid implementation of the...

yeah, I think you're right. At least the x = prox(...) like should be re-evaluated if the step-size has changed. Would you be willing to submit a pull request with...

Could be. But never? I'm surprised. I'll take a closer look into it

No worries about the history, it's fine if there are multiple commits

another option for the dataset is to use something more standard that already includes mnist like tf-datasets, https://www.tensorflow.org/datasets/overview

thanks @gideonite l it all looks good to me. @GeoffNN do you mind taking a look a this ? Thanks

that would be great @matteofrigo !

backtracking does work (in theory and practice) on non-convex problems

Yes On Tue, Nov 24, 2020, 21:14 Geoffrey Negiar wrote: > I meant does it work for accelerating convergence in practice on > non-convex problems? > > — > You...

(as long as we're talking about deterministic algorithms, the top comment makes it confusing)