Chapter 4

1. Which TensorFlow Hub module can be used to instantiate an Inception classifier for ImageNet?

The model at https://tfhub.dev/google/tf2-preview/inception_v3/classification/2 can be directly used to classify ImageNet-like images, as this classification model was pretrained over this dataset.

2. How can the first three residual macro-blocks of a ResNet-50 model from Keras Applications be frozen?

The code is as follows:

freeze_num = 3
# Looking at `resnet50.summary()`, we could observe that the 1st layer of the 4th macro-block is named "res5[...]":
break_layer_name = 'res{}'.format(freeze_num + 2)
for layer in resnet50_finetune.layers:
   if break_layer_name in layer.name:
        break
    if isinstance(layer, tf.keras.layers.Conv2D):
        # If the layer is a convolution, and isn't after 
        # the 1st layer not to train:
        layer.trainable = False

3. When is transfer learning discouraged?

Transfer learning may not be beneficial when the domains are...