bn momentum is different

[FrancescoSaverioZuppichini]

Dear all,

Even if navigate into this code is harder than the 10 years journey of Odisseo, I have noticed that in _global_params you set batch_norm_momentum to 0.99, if you print the model you can see that

EfficientNet(
  (_conv_stem): Conv2dStaticSamePadding(
    ...
  )
  (_bn0): BatchNorm2d(32, eps=0.001, momentum=0.010000000000000009, affine=True, track_running_stats=True)
  (_blocks): ModuleList(
    ....

is set to 0.010.

Now, since it is a google paper basically nothing is really explained. Can somebody tell me what is the correct value?

Stay safe!

Best Regards

[ooodragon94]

hi mate did you try a tensorflow version of efficientnet? I'm not a tensorflow guy... but I think they have an official repo for efficientnet you could probably make some guesses there.

hope it helped and if it does, could you share us your result? I'm also interested

[lukemelas]

https://stackoverflow.com/a/48950072

[FrancescoSaverioZuppichini]

Thank you!