Base Model training = False

[AdarshVijayakumar99]

In the classification problem , the model created using Functional API, we have base-model (EfficientNetB0). We set it to base_model.trainable = False, So that the model don't update the weights while training. Then while creating the functional model, we have set ( base_model(input, training = False). Why are we giving training = False, if we have already set the layers in the base_model as trainable = False. I am not able to grasp the concept here. Please feel free to answer this question. Thanks in advance. :)

[harshitmadaan3504]

Hi, I also faced the same problem and when I did some reasearch on it I found out that When you set base_model.trainable = False, you are freezing the layers of the base model, which means that their weights will not be updated during training. However, this does not affect the behavior of layers like BatchNormalization and Dropout within the base model. These layers have different behaviors depending on whether the model is in training or inference mode, regardless of whether the layers are trainable or not.
In short this is important for consistent and stable training of your top layers.This prevents the randomness and varying statistics that could destabilize the training of the top layers of your model.