As I know, pytorch ver. 1.1 has a bug(I don't know it is really bug) that it doesn't track the mean and variance from training set at evaluation step.
And it makes results over repeated evaluation different.
At first evaluation, futhermore, the evaluation result become too poor because there is no tracked mean & var.
I also tried to replicate your result by your own code, but it couldn't generate 88.5(avenue) or 97.0(ped2).
Did you notice the pytorch bug?
How could you acheive the result even with this bug?
I 'll appreciate your reply. Thank you.
As I know, pytorch ver. 1.1 has a bug(I don't know it is really bug) that it doesn't track the mean and variance from training set at evaluation step. And it makes results over repeated evaluation different. At first evaluation, futhermore, the evaluation result become too poor because there is no tracked mean & var. I also tried to replicate your result by your own code, but it couldn't generate 88.5(avenue) or 97.0(ped2). Did you notice the pytorch bug? How could you acheive the result even with this bug? I 'll appreciate your reply. Thank you.