What does training = False actually do for Tensorflow Transfer Learning?

Question

I have this code right here:

base_model = tf.keras.applications.resnet_v2.ResNet50V2(input_shape=input_shape, include_top=False, weights='imagenet')

base_model.trainable = False

inputs = tf.keras.Input(shape=input_shape)

x = data_augmentation(inputs)

x = tf.keras.applications.resnet_v2.preprocess_input(x)

x = base_model(x, training = False)

What does training = False actually do when we use it for base_model? I know that training is a boolean value to specify we want to run during training on inference mode, but following the Transfer Learning guide on Tensorflow, I can't figure out what it actually does.

We set base_model.trainable = False, this means that the layers won't learn and we are just going to use what they learnt from imagenet. But what does base_model(x, training = False) does? I know that this won't run during training, does so when I am calling the fit() method, what is happening to base_model since training is set to False?

I've read that it has something to do with Fine Tuning and batch norm layers but I am a bit lost.

Also should I use fine-tuning? If I am planning not use it because the model is performing well anyway should I set trainining = True? Or not set that value at all?

Answer 1

In general, that depends on your layers. For example, the dropout layer only sets values to 0, when training=True . Another example is the BatchNormalization layer, which works different during training and inference. For other layers, like the classical dense layer, it does not make a difference. If you really want to know all the details, you will have to read all the used layers and their specific behavior.