#computation-graph

Question 1: Select all that apply

A deep learning framework can run each tensor operation as soon as Python reaches it (eager, imperative or define-by-run execution), or first build the whole computation as a graph and then compile and run it (symbolic or define-and-run execution). Which statements are true? Select all that apply.

In eager mode you can print any intermediate tensor or inspect it in a Python debugger1
A compiled graph can be optimized as a whole and run without the Python interpreter2
Only a static graph can be differentiated automatically; eager code needs finite differences3
Eager execution is always faster, since it skips the compilation step entirely4

Timed test