kylematoba
kylematoba
A good self-contained example would be to reconcile against the PyTorch `fft` module. E.g. check that interpolated points are close to the exact value computed over a finer grid.
Same here, would appreciate a copy of the code to replicate the benchmarks to [email protected]. Thanks!
This is a pretty serious bug. E.g. https://github.com/cleverhans-lab/cleverhans/blob/master/tutorials/torch/cifar10_tutorial.py does not work with norm == 2 and adversarial training turned on. The `loss.backward()` die with something like ``` RuntimeError: one of...
this also breaks for `torch.float16`.
This seems great. It'd help me if you could also make the same change in this call: https://github.com/cleverhans-lab/cleverhans/blob/master/cleverhans/torch/attacks/projected_gradient_descent.py#L123.
Sorry for not better recognising what was going on -- in both of the `prob` functions above, the problem is that the thing you are trying to canonicalise is R^n....
No, freeing everything with ``` dd_FreeMatrix(h); set_free(impl_linset); set_free(redset); free(newpos); dd_free_global_constants(); ``` does not segfault for me. I'm on a recent mac, FWIW, using `/Library/Developer/CommandLineTools/usr/bin/cc`.
`--no-build-isolation` works for me with 1.0.8, torch==2.0.0+cu117.
seeing this as well on M2, 13.4 host with Win 11.