fix lbfgs variable names (#7037)

Switches the step/direction variable names (steps and directions are flipped
in the current implementation of the two loop-recursion). This change does
not change the numerical output of the program, but should make it easier
to follow.

1 files changed, 7 insertions, 7 deletions

diff --git a/torch/optim/lbfgs.py b/torch/optim/lbfgs.py
index 1ccf1ef658..ac096262a8 100644
--- a/torch/optim/lbfgs.py
+++ b/torch/optim/lbfgs.py
@@ -148,8 +148,8 @@ class LBFGS(Optimizer):
                         old_stps.pop(0)
 
                     # store new direction/step
-                    old_dirs.append(s)
-                    old_stps.append(y)
+                    old_dirs.append(y)
+                    old_stps.append(s)
 
                     # update scale of initial Hessian approximation
                     H_diag = ys / y.dot(y)  # (y*y)
@@ -165,20 +165,20 @@ class LBFGS(Optimizer):
                 al = state['al']
 
                 for i in range(num_old):
-                    ro[i] = 1. / old_stps[i].dot(old_dirs[i])
+                    ro[i] = 1. / old_dirs[i].dot(old_stps[i])
 
                 # iteration in L-BFGS loop collapsed to use just one buffer
                 q = flat_grad.neg()
                 for i in range(num_old - 1, -1, -1):
-                    al[i] = old_dirs[i].dot(q) * ro[i]
-                    q.add_(-al[i], old_stps[i])
+                    al[i] = old_stps[i].dot(q) * ro[i]
+                    q.add_(-al[i], old_dirs[i])
 
                 # multiply by initial Hessian
                 # r/d is the final direction
                 d = r = torch.mul(q, H_diag)
                 for i in range(num_old):
-                    be_i = old_stps[i].dot(r) * ro[i]
-                    r.add_(al[i] - be_i, old_dirs[i])
+                    be_i = old_dirs[i].dot(r) * ro[i]
+                    r.add_(al[i] - be_i, old_stps[i])
 
             if prev_flat_grad is None:
                 prev_flat_grad = flat_grad.clone()