Going through the Lesson 2 between the 48th - 50 minute, Jeremy talks about unfreezing all the layers and then creating an array of 3 different learning rates and applying them to the NN, some clarifications
- Why divide the entire NN into only 3 groups why not more or less?
- There is a mention about adding our own layers where are we doing this in the course. Are we adding more layers to the pre-trained NN? If yes where are we doing this in the code?
Those concepts will be analyzed more in depth in lesson 3, you might want to watch that and then search for more detailed discussions in the forum, e.g. Question on how differential learning rates are applied to model
for your Q2 - I hope you’re reviewing the code as you watch. You can see that the model variable is set as ResNet34, then used, then unfreeze is called in the first bits of code. Lesson 1 was similar. So no added layers.
Jeremy explains first changing the weights near the end of the model, then going back through all of the model as a later step is a good method when the pre-trained model is close to the problem being solved.
Changing the architecture of the model makes sense where you have intuition that the data problem uses a different … ‘domain’ (not sure the right word), rather than changing the weights to fit a different ‘look’. I hope that makes sense (fellow learner)
Sorry for the late response thank you @Brad_S @blueseb I revert back to these after watching lesson 4