Brandon Amos

> But this can be done on any function with `torch.autograd.grad`, right? I am not still sure I understand how the Theseus computation specifically matters here. Hmm, it would indeed...

> Thanks all for the discussion. @bamos How about starting this as a new short tutorial and then seeing if there is any overlap with others? Including visualization of gradients...

Hmm, for your problem here, how is the speed in comparison to cvxpylayers? I've been thinking about deprecating this repo and pointing everybody to cvxpylayers

> Cvxpylayers allows batches, however, not on the GPU but CPU only, correct? For the setting I ran, cvxpylayers was significantly slower, in fact it was slower than parallelizing the...

Thanks! I'd merge it in but it seems like I don't have write access to this repo anymore. @zkolter, can you add me back to it?

Hi @nicola-decao - for handling just linear programs you can use https://github.com/cvxgrp/cvxpylayers / https://github.com/cvxgrp/diffcp

Hmm, I don't remember why I implemented it like this and at a first glance it indeed sounds wrong to ignore the last state, especially if there is a terminal...

Hi, I just tried running the MNIST experiment and am hitting nans there too. It's been a while since I've ran that example and I've changed the qpth library since...

Hi, the nans were coming up in the backwards pass in qpth and I've pushed a fix to it here: https://github.com/locuslab/qpth/commit/e2cac495909159aae12461262d0ee540ddf9abd6 Here's the convergence of one of my new runs...

I just added an error message for this case so nobody accidentally starts getting silent errors from this.