We have released Neural Network Libraries v1.37.0!
Along with this release, nnabla-rl v0.14.0 and nnabla-nas v0.15.0 have also been released!
Please see “Spotlight” for important changes.
Spotlight
nnabla: Support CUDA 12.0.1
We have started supporting the CUDA 12.0 extension.
To use the CUDA 12.0 extension, specify nnabla-ext-cuda120 and install it by:
pip install nnabla-ext-cuda120
You can use the all-in-one wheel or the docker image if you do not have the CUDA environment on your PC.
Both require nvidia-driver to be installed but include the CUDA 12.0 runtime, so you will be able to start using CUDA easily.
pip install https://nnabla.org/whl/nnabla_ext_cuda_alllib/nnabla_ext_cuda_alllib-1.37.0-cp310-cp310-manylinux_2_17_x86_64.whl
Here is the instruction to use docker:
docker pull nnabla/nnabla-ext-cuda-multi-gpu:py310-cuda120-mpi4.1.3-v1.37.0
docker run --rm -it --gpus all nnabla/nnabla-ext-cuda-multi-gpu:py310-cuda120-mpi4.1.3-v1.37.0
Please check DockerHub(nnabla-ext-cuda-multi-gpu) or DockerHub(nnabla-ext-cuda) for other tags.
In both cases, you need to install a nvidia driver.
For using CUDA 12.0, you need R525.85.12 or later for Linux, and R528.33 or later for Windows.
If NVIDIA’s Data Center GPU with R470 or later driver is used, you will also be able to run the CUDA 12.0 application by installing cuda-compat-12-0 and setting environment variables. (The above docker image already contains cuda-compat.)
For more information, see NVIDIA’s Forward Compatibility page.
sudo mkdir /usr/local/share/keyrings
$ curl -fsSL https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/3bf863cc.pub | sudo gpg --dearmor -o /usr/local/share/keyrings/nvidia-cuda-keyring.gpg
$ sudo echo "deb [signed-by=/usr/local/share/keyrings/nvidia-cuda-keyring.gpg] https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64 /" > /etc/apt/sources.list.d/nvidia-cuda.list
apt-get update
apt-get install cuda-compat-12-0
nnabla-rl: Add sample rl project template
We have prepared a template for those who want to start a reinforcement learning project but are unsure of what to write and how to learn.
The template uses nnabla-rl and provides a guide on how to start a reinforcement learning project. You can copy the template and easily understand how to modify the code according to your needs by referring to it.
Please feel free to make use of it!
nnabla
Build
- Replace centos7 to rockylinux8 (CPU / GPU)
- fix numpy error
- pycodestyle update auto format (CPU / GPU / example)
- Remove docker inspect in makefile (CPU / GPU)
- fix readthedocs and cython dependency
- upgrade docker base ubuntu2004 (CPU / GPU / example / c-runtime)
- update protobuf
- Cover hpcx environment for CI
- temporally limit cython
Utility
Format Converter
- Update tf version and drop python3.7
- fixed Importer and test of Unique
- Add onnx importer functions: ScatterND, EyeLike, Mod2, BitShift, and Unique
- fix onnx range inputs
Bugfix
- All executor properties should be saved
- Fix Slice, MatMul, and ConstantOfShape in ONNX importer
- adjust certifi dependency
- pack zlib dll for windows alllib wheel (CPU / GPU)
- Make dockerfiles noninteractive while building (CPU / GPU / c-runtime)
- fix TopK error when largest is false and k equals to sample size
- Remove cpplib compiler warnings
- fix top_k problem for duplicated integer values
- Type size correction.
- fix to the example failure
Document
nnabla-rl
Bugfix
- Fixing testing code errors
- Fix Deprecation error when using PendulumEnv
- Pass id as positional arg to avoid unexpected error on old gym
- fix evaluation script
Algorithm
Utility
Document
nnabla-nas
Bugfix
- Bugfix for loading dataloder of OFA reset_bn_statistics
- Fix ImageNet data file directory
- Fix Dynamic BN for inefficient memory use
- Fix bz parameter for schedulers
