Although it provided satisfactory results in the CIFAR10 benchmark, the implementation of RPS LJE didn't actually follow the description in the paper. This is fixed in 55bffb191a10847943d43133c16bb93c1c778f99 . An abstraction is also implemented to simplify the shared code between RPS L2 and RPS LJE in c28abb6750443191f4d8f96b7877602da99fae77 .
The tests were adjusted to follow the current implementation.
I was forced to opt for a less optimized version in 727e5e4f0f406b499b399a5cbbeb9392d292513e without end-to-end lazy computations for the estimation of the hessian as it led to cuBLAS errors when applying the pinv operation.
21 is also fixed by removing the abstract method without the actual implementation in dev@f1a360f
More thorough documentation was added throughout and more notebooks are included in the README.
Fix RPS LJE and update documentation
Although it provided satisfactory results in the CIFAR10 benchmark, the implementation of RPS LJE didn't actually follow the description in the paper. This is fixed in 55bffb191a10847943d43133c16bb93c1c778f99 . An abstraction is also implemented to simplify the shared code between RPS L2 and RPS LJE in c28abb6750443191f4d8f96b7877602da99fae77 .
The tests were adjusted to follow the current implementation.
I was forced to opt for a less optimized version in 727e5e4f0f406b499b399a5cbbeb9392d292513e without end-to-end lazy computations for the estimation of the hessian as it led to cuBLAS errors when applying the pinv operation.
21 is also fixed by removing the abstract method without the actual implementation in dev@f1a360f
More thorough documentation was added throughout and more notebooks are included in the README.