Naive question about flashlight HOT 6 CLOSED

Hi,
flashlight is a 1 year effort, we started the project before PyTorch C++ API existed and we wanted C++ first design for our project - wav2letter.

Here are a few reasons we chose Arrayfire

• C++11 fully-featured tensor library.
• Easy to use API.
• Good documentation and active community.
• Works seamlessly on multiple devices - supports CPU, CUDA, OPENCL.
• And more importantly, has a JIT compiler for efficiency.

from flashlight.

@vineelpratap Can you elaborate more on the reasons? Is PyTorch C++ the negations of the reasons you stated, i.e.

• Not a C++11 fully-featured tensor library
• Bad to use API
• Bad documentation and inactive community
• Works bad on multiple devices
• Has no JIT compiler

You make it sounds so bad.

from flashlight.

As mentioned, there was no C++ API for PyTorch when we started the project and we found Arrayfire to be good Tensor library to be used in our projects.

Please do not draw any further conclusions !

from flashlight.

Hi,
The notable difference is that flashlight use arrayfire as it's tensor library and pytorch uses Torch/ATen.
You can find the API for flashlight here - https://fl.readthedocs.io/en/latest/ .

from flashlight.

Can you expand a bit why ArrayFire was chosen? Why might a user prefer an ArrayFire-backed library over Torch/ATen or other alternatives?

from flashlight.

wasn't Caffe2 already there by that time ? :D
Anyways, awesome library !

from flashlight.