Train EfficientNetb0 Loss of 0.0

FrancescoSaverioZuppichini / glasses

High-quality Neural Networks for Computer Vision 😎

https://francescosaveriozuppichini.github.io/glasses-webapp/

MIT License

431 stars 37 forks source link

Train EfficientNetb0 Loss of 0.0 #289

Open Boraly opened 2 years ago

Boraly commented 2 years ago

I tried to train the efficientnetb0-model on my own dataset, but got a loss of 0.0. After some digging I found that the output of the forwad function was just a tensor of NaNs. I tried the code for efficientnetb1 too and it works flawless. But I can't get behind this bug. What could I try?

FrancescoSaverioZuppichini commented 2 years ago

That's interesting indeed, I don't know I'll have to do some digging. Maybe the stochastic depth is wrong