I would recommend this roadmap, if you haven’t dwelled into pytorch much. (docs in pytorch.org are good enough.) Since there are a lot of resources it might be hard to structure your learning material. This is the one I’m following.
I can confirm that the free Udacity pytroch course is very useful. When you dive into the depth of pytorch, you also start to understand why some decisions were made in fastai, which are very unclear if you come from tensorflow background. So thanks for the link!
The udacity course seems like a pretty well versed course with good material on the fundamentals of deep learning and image processing as well. There is a lot of concept explanation and less code in many of the videos. For those looking to only code, I think a fair amount can be skipped.
@SHAR1 Point 3 links to videos made 2 years ago, are they still valid for the current version of torch? Also the new dev_nb’s are out so they may be better to go through instead of the older.
Personally I think the docs are the a pretty good place for the basics - read about the basics operations, autograd, etc , but their examples can lead one to copy paste code without getting it if they don’t know those.
Excuse me for the late reply, last couple of days have been pretty busy. Nevertheless, I managed to go through intro to deep learning with pytorch from udacity: all lectures, MCQs, code snippet quiz & final project . I still have to complete some of the practice notebook assignments. Mainly because, it looked promising and I like to look at things at different perspectives.
This course complements fastai so well. For beginners who don’t have much time to understand the math, taking both of the courses parallely would be ideal. After completing both the courses you can start with deeplearning.ai and dig into the math (if you are curious), till fastai v3 part 2 starts. This might be a decent roadmap for you.
Here is my takeaways from the intro to deep learning with pytorch (which are unique to this course) …
The high budget animations are soothing to watch, which makes it easier to complete the course material.
A proper structure is always easier to follow.
Assignments are also pretty thought out. I wish they put more questions.
I am truly happy that we are using discourse and not slack. The noise in slack is very hard to ignore.
Intro to neural networks (Lesson 2) : Explanation of Perceptrons as logical operators, perceptron algorithm, meaning of error, loss function.
Intro to pytorch (Lesson 4) : Notebooks are thoroughly documented.
CNN (Lesson 5) : Best explanation of filters. Why multi layered perceptron vs CNN, assignment on opencv and cnn (pytorch).
Style Transfer (Lesson 6) : I like the way they make(forcibly) you read the style transfer paper to complete an assignment.
RNN&LSTM (Lesson 7) : Dissecting an LSTM and explaining what each gate does. I do feel the examples takes are too high level. Writing test data for your models. Sequence Batching. I invested some time in understanding the tensor sizes (shapes of the i/o tensors) of elements in the whole pipeline. This lesson helped me to get a hold of all the intricacies of pytorch one has to keep in mind while implementing a RNN model.
Sentiment analysis (Lesson 8) : A Simple practical RNN problem statement. I have to spend more time here.
Deploying pytorch models : I was a bit disappointed here. Same content as in pytorch docs. An alternative example would have been useful.
Challenge Project : Get highest accuracy on a dataset. A sense of competition, some may like it.
I may have missed out on other important aspects, but the above is what all I can remember. Finally, for beginners who are sparse on time, it is a great resource.
For college students, if its not exam week. Please, take up any popular deep learning mooc, like the one @ecdrid mentioned above.