Google Git
Sign in
android / platform / external / pytorch / 6ee6aedf20
commit6ee6aedf20f6e8d89aff8fa9c37d4230cc03af6f[log] [tgz]
authorZeming Lin <misterabc@devgpu029.prn2.facebook.com>Tue Feb 23 09:31:19 2016 -0800
committerZeming Lin <misterabc@devgpu029.prn2.facebook.com>Tue Feb 23 09:31:19 2016 -0800
treeae0067233f6281fbb0e1c2a575d24862c9243a44
parentbae607590676d7eb0c0eec8be976caa4cb91f639 [diff]
Adding SparseLinear with CUDA, requires buffer variable
2 files changed
tree: ae0067233f6281fbb0e1c2a575d24862c9243a44
  1. doc/
  2. generic/
  3. CMakeLists.txt
  4. init.c
  5. README.md
  6. THNN.h
README.md

THNN

THNN is a library that gathers nn‘s C implementations of neural network modules. It’s entirely free of Lua dependency and therefore can be used in any application that has a C FFI. Please note that it only contains quite low level functions, and an object oriented C/C++ wrapper will be created soon as another library.

There is also a CUDA counterpart of THNN (CUTHNN) in the cunn repository.

Links

Motivation

Torch‘s nn module provided many optimized C implementations of modules, but the source files contained Lua specific code and headers so they couldn’t be easily compiled and included anywhere else.

THNN is based on the same code, but is written in pure C, so it can be easily included in other code. Future C implementations should be committed to THNN.

API

THNN is a purely functional library. It provides 2-3 functions for each module, that perform the most important operations:

  • updateOutput - applies the module to an input
  • updateGradInput - accepts gradient w.r.t. output and previous module input, and computes a gradient w.r.t. that input
  • accGradParameters - (optional, only modules with parameters) accepts gradient w.r.t. output and previous module input, and computes gradient w.r.t. the parameters

For information on argument types please check the API reference.

This is all THNN library provides. An object oriented implementation similar to nn will be provided in a separate library. This one is just a set of CPU kernels.

Developer docs

This section will be expanded when FFI refactoring will be finished.