Master | Release |
---|---|
Pytorch - Py + Nim
A Nim frontend for pytorch, aiming to be mostly auto-generated and internally using ATen.
Because Nim compiles to C++, this is not a wrapper or binding library. It generates 1-to-1 native ATen code.
The only requirement from pytorch is ATen's core tensor library. Because of this, nimtorch is extremely versatile and can compile on any kind of device.
Current status
Early stage
- Automatically generated, from
Declarations.yaml
, the full ATen API - Cuda support ( add -d:cuda when compiling with nim )
- WASM support ( add -d:wasm when compiling with nim )
- Automatically generated, from
derivatives.yaml
, gradient procs - Autograd
- Add missing derivatives
- More high level pytorch API (Module, Models etc)
- ...
The final aim is to be as compatible as possible with the pytorch API.
Why
Ease of use of the python language while keeping fully bare metal native C++ performance
Python code
# GRUCell
gi = x.matmul(w_input.t()) + b_input
gh = hidden.matmul(w_recur.t()) + b_recur
i_r, i_i, i_n = gi.chunk(3, 1)
h_r, h_i, h_n = gh.chunk(3, 1)
resetgate = (i_r + h_r).sigmoid()
inputgate = torch.sigmoid(i_i + h_i)
newgate = (i_n + resetgate * h_n).tanh()
hy = newgate + inputgate * (hidden - newgate)
Nim code
# GRUCell
let
gi = x.matmul(w_input.t()) + b_input
gh = hidden.matmul(w_recur.t()) + b_recur
(i_r, i_i, i_nn) = gi.chunk(3, 1)
(h_r, h_i, h_n) = gh.chunk(3, 1)
resetgate = (i_r + h_r).sigmoid()
inputgate = torch.sigmoid(i_i + h_i)
newgate = (i_nn + resetgate * h_n).tanh()
hy = newgate + inputgate * (hidden - newgate)
Getting started
Requirements
Linux: A recent distribution on par with ubuntu 18.04 in terms of libc and basic libraries, gcc compiler
macOS: We compile with 10.13 min version flags but might work even on lower versions, XCode for the compilers
Windows: Windows 10, Visual Studio Runtime 2017 and Visual Studio 2017 (any edition)
WASM: Latest Emscripten compiler and tools
Super easy, using conda
Linux, macOS and Windows
conda create -n nimtorch -c fragcolor nimtorch
(add cuda10.0
for cuda 10 linux only or add wasm
for wasm version)
source activate nimtorch
or on windows: conda activate nimtorch
This will install: nim and ATen binaries, fragments and nimtorch all in one command, nothing else needed.
Make sure you use a recent version of conda and have a compiler installed in your system, on windows you have to add --cc:vcc
and be on a developer prompt.
Make sure your system is recent (ubuntu 18.04 reference / macOS High Sierra / Windows 10) and you have cuda 9.2 installed (if you need cuda, linux only, more cuda versions coming, please open a issue if you need a specific version).
Test with with something like:
nim cpp -o:test -r $ATEN/dist/pkgs/nimtorch-\#head/tests/test_xor.nim
or on windows... (because dlls need to be side by side)
nim cpp -o:%ATEN%/lib/test.exe -r %ATEN%/dist/pkgs/nimtorch-#head/tests/test_xor.nim
Semi manual way
Linux, macOS and Windows
Check what version of ATen/PyTorch we need in conda/nimtorch/meta.yaml
- should be something like aten ==2018.10.10.1089
Note the version as you will need it in the next step
conda create -n aten -c fragcolor aten={version}
or
WASM
conda create -n aten -c fragcolor aten={version} wasm
or Cuda 10.0 (linux only)
conda create -n aten -c fragcolor aten={version} cuda10.0
activate aten environment
source activate aten
or on windows: conda activate aten
- Make sure you have a recent Nim and Nimble version in your path
- Easy option: install nim with choosenim
choosenim devel
- clone the release branch
git clone -b release https://github.com/fragcolor-xyz/nimtorch.git
cd nimtorch
nimble develop
finally
run self test nim cpp -o:test -r torch.nim
(use -o:%ATEN%/lib/test.exe
instead on windows because of dll location)
in the case of WASM:
run self test nim cpp -d:wasm -o:test.js torch.nim && node test.js
(needs node.js)
Manual way without requiring conda
Build ATEN
pip2 install pyyaml typing
git clone -b fragcolor-devel https://github.com/fragcolor-xyz/pytorch.git
cd pytorch
git reset --hard <commit hash> # from torch/commit.txt
git submodule update --init
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release -DUSE_CUDA=OFF -DBUILD_ATEN_ONLY=ON -DCMAKE_INSTALL_PREFIX=`pwd`/output ../
make -j4
make install
# also copy derivatives if we want to run generator.nim or nimble test
# notice generator.nim might need python3 and pyyaml
cp ../tools/autograd/derivatives.yaml `pwd`/output/share/
Test the build
cd
ATEN=
nim cpp -r -f -o:/tmp/z01 torch.nim # for eg: ATEN=pathto/pytorch/build/output/
Notes
- We suggest setting
OMP_WAIT_POLICY
environment variable toPASSIVE
when running on CPU.