Lesson 12 Darknet Cifar with bad accuracy


(James_Ying) #1

There is the original Fastai accuracy.

And this is mine
Why mine accuracy start from a vary low number.
Also the loss is big
The final acc is not as good as original fastai records.

The only thing I changed is here. I don’t have multi-GPU.
But train on Windows10 with 1080Ti
Snipaste_2018-07-05_21-32-56


#2

I haven’t scrolled through that notebook much, but I’m guessing that you have to change the learning rate to go with the fact you’re note using multi-GPUs.
For a single GPU, you should use the other notebook (the wide resnet is also the network used to win the Dawnbench competition).


(James_Ying) #3

@sgugger
I run the lr_find(), and pick up the lr=2e-2(original is lr=1.3)
but, even original lr is much bigger than 2e-2, still works better.

Is the Multi-GPU affect the result so much?


(Fabio Graetz) #4

I find the same as you using one GPU. I wonder if the reason really is multiple GPUs. Did anyone else try this?