Course Recommendation — PyTorch Fundamentals: A Challenging but Worthwhile Course

I finally finished the PyTorch course!

In the context of my overall learning plan, this is really just one small step (if you're curious, take a look at my TensorRT-LLM Learning Path first).

Throughout the course I could genuinely feel the instructor's care and dedication. Borrowing a phrase Professor Hung-yi Lee often uses — this is the kind of patient, methodical teaching you only find in "ancient times."

What do I mean? In this AI era everyone moves fast, but this course essentially prohibits you from using AI freely. Each module's graded assignments pack four or five handwritten problems into a single Jupyter Notebook, preceded by a lengthy context setup. Each assignment submission takes at least an hour. Honestly, it's not easy — without the copy-paste shortcut, the learning curve gets even steeper and takes considerably longer.

On top of that, you're expected to memorize the names of PyTorch modules layer by layer. That's genuinely a grind. In practice I still had to reference previous examples most of the time.

My compromise was to write down my previous assignments — which module names I had used — and then use the assignment prompts to recall and locate the corresponding functions and adapt the parameters.

But because of that constraint, I shifted my focus toward understanding the core templates and formulas, why things are done this way, and what the critical points to watch for are.

The syntax details of individual functions? I'm confident I can ask AI to handle that anytime. As a human in this era, what we should actually understand is the essence and logic — for example: why do you need to call zero_grad before computing gradients and stepping the optimizer? What is the underlying mechanism of different layers (why does an image need Max Pooling? Why do you Flatten before feeding into a fully connected neural network)? How should input and output layer dimensions be defined?

These are the core concepts you absolutely cannot get wrong when describing a neural network architecture to AI. Otherwise, we're just being led by AI rather than directing it.

All in all, if you're interested in CNNs or want to get started with PyTorch, I highly recommend this Coursera PyTorch Fundamentals course.