I just finished Part I and thought it was fantastic. I cannot believe this course is free, I am so grateful. Admittedly, I am now replaying the videos a few more times as 5-7 were difficult for me.
However, I wanted to take a quick moment to ask about transfer learning. As we saw with convnets and the fastai library it took only a few lines of code to use an existing model on our own small dataset and train it for our own personal projects (assuming the data was similar). We saw this throughout and including lesson 4 with nlp and imdb reviews.
My question is can we do just transfer learning using the fastai library without having to train models from scratch or dig deep into pytorch or fundamentals in theory? Is there a way to find a model that is pertinent to our cause (i.e. https://modelzoo.co/) and use fastai to add some layers to it and experiment?
For example, in ModelZoo they break model categories into: Computer Vision, NLP, Geverative Models, Reinforcement Learning, Unsupervised Learning, and Audio&Speech. Is it possible to use any one of these models with fastai to accomplish transfer learning as we saw in Part I? If so, is there any literature delving deeper into this somewhere? Is this even making sense? I apologize in advance here for my naivete maybe Part II goes into this, Im hoping below makes more sense than my words…
arch=<insert_downloaded_model_here> data = Foo.from_paths(PATH, tfms=tfms_from_model(arch, sz)) learn = Bar_Learner.pretrained(arch, data, precompute=True) learn.fit(0.01, 2)