Tensorflow| EfficientNetB0

[MkRakesh]

Hi, Daniel .. I have been watching your tensorflow course and practcing for a while now. In transfer learning with EfficientnetB0 as base_line model we are skipping normalize/resizing since the base_line model already has these layers but the sametime we are freezing the layers first and tuning last 5 or 10 layers. So since these resizing/normalisation layers are among the first few layers and they are freezed while we use them , in effect we are not passing the image without resizing/normalising right.. I am in some chaos regardingthis.. Also i am getting a validation accuracy of 75% and while predicting my model is biased to a single class out of 101 classes.. Hope you will clear these concers. regards

[mrdbourke]

Hi @MkRakesh ,

Good question!

It's actually the reverse for the unfreezing - we unfreeze the last ~10 or so layers, the layers closest to the output of the model (so the initial layers stay frozen).

Also, since the resizing/normalization layers often do not have any learnable parameters, these layers won't change even if they are frozen/unfrozen.

Does this make sense?

[MkRakesh]

Okay alright ..got it . I thought we are completely discarding the layers..Thanks for the reply Daniel..

Also Daniel , I have deployed the Deployed the N-BEATS forecasting/BitPredict and works fine.. Really appreciate your crazy training.. I am really waiting for GANs/Diffusion course from you.. Thank you.

https://nbeats-bitpredict.streamlit.app/

[mrdbourke]

Woah! That's really cool! Massive effort on the deployment!