Flambeau provides Nim bindings to libtorch C++ API and torchvision.
The library is not ready for general use at the moment.
The library installation including:
- auto-downloading and deploying of libtorch
- optional auto-compiling of torchvision into a shared library (if requested)
Initialize submodule :
git submodule update --init --recursive
On Ubuntu :
sudo apt-get install ffmpeg libpng-dev libjpeg-dev libzip-dev
On OpenSuse :
sudo zypper install libpng16-devel libjpeg62-devel libzip-devel ffmpeg
git clone https://github.com/SciNim/flambeau
cd flambeau
nimble install or nimble develop
Note that install and develop will download and build libtorch so it may take a while.
Torchvision can now be built if desired:
nimble build_torchvision
As this library is still in heavy developments, some constructs are a bit brittle.
If you wish to return a RawTensor (the wrapped Torch Tensor type)
from a procedure, you need to annotate the procedure with the
{.noInit.} pragma, like so:
proc foo(x: float): RawTensor {.noInit.} =
...Otherwise, you will get a segmentation fault due to the implicit
initialization of the RawTensor object.
Note: you can use the {.push.} and {.pop.} pragmas at top level in
your code, if you wish to write multiple procedures returning
RawTensor without having to add this pragma to each procedure.
Note 2: In theory the {.requiresInit.} pragma should mean that the
RawTensor type is not implicitly initialized. However, this pragma
does not solve the issue at this time.
By default the Torch installation downloaded by this package contains
CUDA support. However, by default Nim packages using Flambeau compile
without CUDA support. This is controlled by the cuda compile time
option. Compile with:
nim cpp -d:cuda <foo>to compile with CUDA support.
Compared to Numpy and Arraymancer, Flambeau inherits the following PyTorch limitations:
- No string tensors.
- No support for negative step in tensor slice (
a[::-1])
Licensed and distributed under either of
- MIT license: LICENSE-MIT or http://opensource.org/licenses/MIT
or
- Apache License, Version 2.0, (LICENSE-APACHEv2 or http://www.apache.org/licenses/LICENSE-2.0)
at your option. This file may not be copied, modified, or distributed except according to those terms.