Week 2: Transfer learning - input layer size

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

Hello,

the course focuses on replacing the last layer for another dense classification layer suitable for our application. For example we replaced the 1000 size dense softmax layer in the MobileNet coursework by a 1 size linear classification layer to classify the Alpacas.

My question is, what about the input layer? The original MobileNet weights were trained on imagenet images, but these had to be of some particular size, right? Does this mean that our Alpaca dataset has to be of the same size, or is this arbitrary? What if I have images of different resolution, can I still use Transfer learning?
(I tried to specify different IMG_SHAPE in the assignment when calling

base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,
                                               include_top=True,
                                               weights='imagenet')

, but keras wasn’t happy.

2

I think your images have to be the same size as the training set.
If necessary, you can use an image processing tool to resize your images before you use them.