What are pros and cons of Caffe2 comparing to PyTorch, TensorFlow?


It seems that PyTorch heart is Caffe2. I am doing tutorials in Caffe2. From the user-friendly point, I can’t say too much. It feels like everything was put into the performance. Is Caffe2 faster than TensorFlow? What are the pros and cons of Caffe2? Should I move to Caffe2 completely?

UPDATE: If I have time I would test all of them. TensorFlow is really fast but if Caffe2 is faster…

UPDATE: I have found this https://github.com/u39kun/deep-learning-benchmark/blob/master/README.md.


I wouldn’t say PyTorch’s heart is Caffe2, but rather both projects are merging as described here.

I’m not convinced by the benchmark, as it’s comparing PyTorch 0.3.0 to other frameworks in older versions. That being said, I’m skeptical most of the time if the benchmarks were performed by “outside” developers.

Do you have any performance issues using PyTorch?

Dear @ptrblck!

Performance on the second place. Extensiveness and easiness to use is a prime target.
Why I have chosen PyTorch? Because of it logical modular structure and easy to understand API, well thought extension capabilities, perfect documentation and it sounds not obvious but lack of functions that only few need.
Why I am raising this topic? Mainly because I want to see how PyTorch breathe, where is the source code for Conv2d. And I found a clue in Caffe2. I might be wrong but I looked at the source code after installation of PyTorch and there is nothing there about layers logic in Python code except for Linear but there is a large * caffe2 *.so where I guess the whole logic is. It’s not that PyTorch is closed and we should blame it. In TensorFlow situation is the same. But I want to click on documentation and see how the thing is working. For now it requires a patience to find it out.

Thank you very much for you answer @ptrblck without you people would stuck forever with their code.