Avoid zero division in _cubic_interpolate (#42093)
Summary:
I encountered a zero division problem when using LBFGS:
File "/home/yshen/anaconda3/lib/python3.7/site-packages/torch/optim/lbfgs.py", line 118, in _strong_wolfe
bracket[1], bracket_f[1], bracket_gtd[1])
File "/home/yshen/anaconda3/lib/python3.7/site-packages/torch/optim/lbfgs.py", line 21, in _cubic_interpolate
d1 = g1 + g2 - 3 * (f1 - f2) / (x1 - x2)
ZeroDivisionError: float division by zero
My solution is to determine whether "line-search bracket is so small" before calling _cubic_interpolate
Pull Request resolved: https://github.com/pytorch/pytorch/pull/42093
Reviewed By: pbelevich
Differential Revision: D22770667
Pulled By: mrshenli
fbshipit-source-id: f8fdfcbd3fd530235901d255208fef8005bf898c
