Should a beginner start with PyTorch or TensorFlow for Data Science?

All answers to this question.

For a beginner in 2024, I highly recommend PyTorch. The reason is simple: when you get an error, the stack trace actually makes sense. Since it’s dynamic, the error happens exactly where the problem is in your code. TensorFlow (even with Keras) can sometimes give very cryptic errors related to its underlying C++ engine or graph construction. PyTorch’s documentation has also become world-class, with a "Recipes" section that covers almost every common use case. Plus, most of the "cool" AI projects on GitHub you'll want to play with are written in PyTorch, so you'll be able to learn by reading other people's code much more easily.

   Answered 2025-01-12 by Gwendolyn Price

But isn't Keras (TensorFlow's high-level API) much simpler for a basic image classifier? You can build a model in like 10 lines of code. Isn't PyTorch a bit more "boilerplatey" because you have to write your own training loop (the for loop over epochs)?

   Answered 2025-01-13 by Mitchell Young

• Mitchell, you're right that Keras is very concise. However, writing that "boilerplate" training loop in PyTorch is actually a huge learning advantage. It forces you to understand what is actually happening: the forward pass, the loss calculation, and the backpropagation. When your model doesn't work, you'll actually know why because you wrote the steps. Keras hides too much of the "magic," which can make troubleshooting harder for a student.

     Commented 2025-01-14 by Sabrina West

Start with PyTorch. It’s what 90% of the industry uses for new projects now. Learning TensorFlow today is like learning to drive a manual car when everyone is buying automatics.

   Answered 2025-01-15 by Randall Stevens

• Randall is right. Even if you eventually have to use TensorFlow at a future job, moving from PyTorch to TF is much easier than the other way around.

     Commented 2025-01-16 by Gwendolyn Price