PyTorch

Question 1

What are the three great things about PyTorch?

Answer 1

1. Python-Native Debugging: Since the graph in PyTorch is characterized at runtine you can use native Python debuggers like pdb, ipdb, or with print statements! e.g., can use pdb.set_trace()
2. Imperative Programming: how to achieve our goal! Not symbolic or declarative. Imperative operates line by line. Non impertaive usually has a compile step.
3. Dynamic Graph: define by run; a graph is created on the fly; computes as line is processed; can add dynamic layers while training, etc.

Negatives: not really for prod. What can you do? BUild model with pytorch, fully trained model; convert to TensorFlow and use TensorFlow Serving (prod model serving architecture, to serve users; can scale to millions of people; Google uses it)