3.3 Answers#

Should you avoid using a pre-trained network that was trained on vastly different data that the one you want to train on?

Using a pre-trained network can be useful as this network might have learned something well, possibly on a lot more data than you have. By re-training the model, also called fine-tuning, you want to make sure than the model is trained to perform well on your type of data.

The notebook states that "it is advisable to also load the learning rate that was used when the training ended". Why is that?

If you are training on the same images, or on very similar images, then part of the training as already been done. By fine-tuning, starting with the initial learning rate, the network might take longer to converge.

Why go all the way back to the parking lot while you are already close to the mountain top?