What does affect the number of trainable values for fine-tuning when you unfreeze some top layers?

[emrecicekyurt]

Hello everybody,

I've just experienced an interesting thing: while setting a model's top layers as trainable, there were two ways to do that; both gave different results.

While setting the top layers as trainable = True, the first option was setting the base_model.trainable =False then assigning the top layers to True in a for a loop. This method gives only 2 trainable variables after compiling the model which causes a lower accuracy.

The second option is assigning the base_model.trainable= True then setting up all layers untrainable excluding the top 10 layers in a for a loop. It gives 12 numbers of trainable variables (of course after compiling) that cause better results after fitting.

Trainable layers for both methods (only top 10 layer is unfrozen)

[muhammadanas0716]

This article may help

https://keras.io/guides/transfer_learning/

[emrecicekyurt]

Thank you for sharing the documentation but the point that I didn't understand is for both methods we have the same trainable layers with different trainable variables. Both making top layers frozen but resulting in different variables doesn't make sense.

[muhammadanas0716]

I really can't say much on this. But try to go through Daniel's notebook and compare it with yours. It will surely help you resolve this issue!