Deep Learning Tensorflow vs Keras vs PyTorch - Code in Python

I’m going through this tutorial on matrix decompositions from fast.ai and they use PyTorch in their examples. So I was curious as to what the differences were compared to other libraries I have heard of such as Keras and Tensorflow.

http://codeinpython.com/tutorials/deep-learning-tensorflow-keras-pytorch/

My Takeaways:

  • Keras is high-level API to manage experimentation of neural networks of other libraries e.g. Tensorflow
  • Tensorflow is Google’s general machine learning library that uses a “static computational graph” i.e. you define your model and then run it
  • PyTorch is a “dynamic computational graph”, where you can have any number of inputs throughout the model and the model is modular so that you can debug parts of it at a time, but only works on Linux and macOS (more on PyTorch).

Edit: Some more comparisons between PyTorch and TensorFlow.

1 Like

I am taking a new AI nanodegree from udacity, and they are using PyTorch. I think there are similarities between the usage of PyTorch and other popular python libraries that makes the learning curve easier to handle.
P.S congrats on getting the top contributor’s badge.

1 Like

Interesting continuation of the conversation with another blog post from someone else!

From article:

TL;DR:

Keras may be easier to get into and experiment with standard layers, in a plug & play spirit.

PyTorch offers a lower-level approach and more flexibility for the more mathematically-inclined users.

1 Like