The PyTorch implementation of L-BFGS in torch.optim.LBFGS is fast, however, it does not seem to be as reliable as the one in SciPy.
https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.fmin_l_bfgs_b.html
What needs doing?
- Create a new Solver in
solver.py that includes scipy's L-BFGS.
- Compare reliability with PyTorch version on a diverse dataset of textures.
What's the challenge?
- Ensure that
scipy module is not required, only imported locally when necessary.
- Ensure the solver works with both CUDA and CPU tensors.
The PyTorch implementation of L-BFGS in
torch.optim.LBFGSis fast, however, it does not seem to be as reliable as the one in SciPy.https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.fmin_l_bfgs_b.html
What needs doing?
solver.pythat includes scipy's L-BFGS.What's the challenge?
scipymodule is not required, only imported locally when necessary.