Documents are available on the ReNom.jp web site.
ReNom version 2.6.1
Changes from 2.5
Please refer to
changes at renom.jp.
- Improved performance.
- Add function
curand_set_seedto determine the curand random number generator's seed.
- Add argument
ignore_biasto all parametrized class.
- Add argument
reduce_sumto all loss function class.
Convlution 3d, nd.
Max, Avg Pooling 3d, nd.
Changes from 2.6.0
- Bug fix of the class LayerNormalize. Not to raise index error when gpu calculation.
- Add an example of LayerNormalize.
- python2.7, 3.4
- numpy 1.13.0, 1.12.1
- pytest 3.0.7
- cython 0.24
- cuda-toolkit 8.0, 9.1
- cudnn 5.1, 6.0, 7.1
- matplotlib 2.0.2
- pandas 0.20.3
- scikit-learn 0.18.2
- scipy 0.19.0
- tqdm 4.19.4
First clone the ReNom repository.
git clone https://github.com/ReNom-dev-team/ReNom.git
Then move to the ReNom folder, install the module using pip.
cd ReNom pip install -e .
To activate CUDA, you have to build cuda modules before
pip install -e .
using following command.
python setup.py build_ext -if
Please be sure that the environment variable CUDA_HOME is set correctly.
$ echo $CUDA_HOME /usr/local/cuda-9.1
If you set an environment variable RENOM_PRECISION=64, calculations are performed with float64.
Default case, the precision is float32.
Limit of tensor dimension size.
In ReNom version >= 2.4, only tensors that have less than 6 dimension size can be operated.
“ReNom” is provided by GRID inc., as subscribed software. By downloading ReNom, you are agreeing to be bound by our ReNom Subscription agreement between you and GRID inc. To use ReNom for commercial purposes, you must first obtain a paid license. Please contact us or one of our resellers. If you are an individual wishing to use ReNom for academic, educational and/or product evaluation purposes, you may use ReNom royalty-free. The ReNom Subscription agreements are subject to change without notice. You agree to be bound by any such revisions. You are responsible for visiting www.renom.jp to determine the latest terms to which you are bound.