Unable to reproduce results from scratch

imtiazziko / LaplacianShot

Laplacian Regularized Few Shot Learning

81 stars 8 forks source link

Unable to reproduce results from scratch #6

Closed zgpio closed 3 years ago

zgpio commented 3 years ago

Hello, I follow the description in the README：

Alternatively to train the network on the base classes from scratch remove the "--evaluate " options in the following script.

But the result I got at the end is far from the result you gave:

Meta Test: BEST
feature            UN               L2N            CL2N
GVP 1Shot       0.3235(0.0010)  0.3469(0.0011)  0.3877(0.0010)
GVP_5Shot       0.3192(0.0010)  0.3486(0.0011)  0.3933(0.0011)

Is it something I haven't noticed. Hope to get your reply and guidance

imtiazziko commented 3 years ago

This is weird. The model is not trained correctly then. Please check your training log and check the loss. I think this can happen due to the pytorch version. This code was tested with Pytorch 1.2 and python 3.6. Please check with the other pytorch version and see.

Thanks