Mathieu Blondel
Mathieu Blondel
Thanks for the bug report. I don't see any mistake in your code a priori. I tried to run your code and got a linear transpose rule error: ```python Traceback...
As a temporary workaround, you can set `implicit_diff=False` in `ProjectedGradient`. This way, there is only one implicit differentiation involved, the one of `projection_polyhedron`.
I don't think that's possible with `ScipyRootFinding`. So your root finding problem has a non-unique solution and you know that there exists a solution in `[min_x, max_x]`? One way would...
I don't think so, but a patch would be welcome.
Is it specific to LBFGS or does it happen with any solver?
Can you also check if the `jit` and `unroll` options to `LBFGS` have any impact on this? Depending on these options, a different loop implementation is used. Normally, solver objects...
It's nice to have a workaround but shouldn't garbage collection be able to do this automatically?
When you use ```python @implicit_diff.custom_root(optimality_fun) def fun(...): ``` `fun` and `optimality_fun` should have the same number of arguments, which is not the case here. Your function `euclidean_weighted_mean` includes non-differentiable arguments...
I think we can keep `solve_inv` and create a new `solve_pinv` in addition.