Graphs in Deep Learning

Jurassic · Post by **Jurassic** » Mon Dec 16, 2019 12:34 pm

1) Why does TF and PyTorch need graphs?

2) Why do we care if they are static or dynamic?

Simple high level explanations please

sharpe_machine · Post by **sharpe_machine** » Mon Dec 16, 2019 4:44 pm

2) C++ vs Python? C++ allows very effective optimizations but requires static declarations of literally everything (well, modern C++ has type inference, but still) - you should declare tensor shapes, tensor types, etc.. On the other hand, Python interprets the code and can successfully deal with the wrong data type passed in or wrong shape of the tensor.

1) Well, I do not know the answer to "why" question. But computational graphs have some nice features such as easy derivative computation which is essential for backprop.