#frozen-base

Question 1: Single choice

You attach a new, randomly initialized dense head to a convolutional base pretrained on ImageNet, to classify cars and trucks. Why set pretrained_base.trainable = False before training?

The random head's large gradients would wreck the pretrained features1
A pretrained base cannot be trained any further once its weights have been loaded2
Freezing the base makes it learn new features suited to the cars and trucks3
The sigmoid output needs frozen layers below it to produce valid probabilities4

Timed test