Version 1.21.0 Release

release-note-break-compat

• Make inplace option in most of Function operations obsolete
• Inplace methods reverts to using inplace=False

release-note-bugfix

• fix status 9 error
• fix race condition for data iterator with cache dataset
• Fix duplicate calls in forward_all
• Properly close file handlers at exception
• Bugfix: fix nn.grad for floating variables
• fix parameter error for graph viewer
• fix indexing leaf variable problem
• Add unitest for graph/parameters issue

release-note-build

• Fix APT_OPTS for runtime dockerfiles
• Skip some tests on 32bit system.
• CI: Restructure dockerfile directory
• Fix build without http_proxy
• temporarily comment out a test which cause randomly crash.
• change image_utils imread behavior for pillow backend compatibility

release-note-core

• Support variadic inputs of nbla::functions::concatenate()

release-note-doc

• Fix doc error
• Add missing function modifiers to doc
• improve extension document
• modify contents of Math section and Array Manipulation
• Improve dropout explanation and add searchsorted to document
• update document for randint
• Fix generating tensorboard writer related document problem

release-note-format-converter

• fix: update onnx_opset_tensorrt, use opset11 version
• support ceil_mode for onnx exporter
• support weightnormalization exporter
• support nnabla to export weightstandardization & instancenormalization & layernormalization in onnx
• Add a quantized tflite converter to export nnp to int8 tflite
• support nnabla to export spectralnorm in onnx

release-note-op-layer

• Double backward for cumsum
• Modify cumprod backward to deal with zero input
• Support p=0 for dropout function
• Ignore label less than 0 in cross entropy
• add dot & matmul operator
• Boolean Indexing Functions
• Remove ReLUCudaCudnn
• Add roi_align function and tests
• Higher-order gradients for normalization functions

release-note-utility

• Reduce create_cache log for s3 dataset
• feat: reduce create cache log and print every 5%
• Support multiple callback
• Implement custom module nvml to replace pynvml
• refactor:remove redundant image backend switch

Install the latest nnabla by:

pip install nnabla
pip install nnabla-ext-cuda102 # For CUDA version 10.2 users

You can also install the cuda extension with specific versions from one of the following. See also FAQ

• nnabla-ext-cuda100 (CUDA 10.0 x cuDNN 7.6)
• nnabla-ext-cuda102 (CUDA 10.2 x cuDNN 8.0)
• nnabla-ext-cuda110 (CUDA 11.0 x cuDNN 8.0)

pip install nnabla
pip install nnabla_ext_cuda102  # For CUDA 10.2 x cuDNN 8.0 users

For distributed training, You need to install the correct package that
matches the version of MPI installed on your system.

We prepared following packages.

• nnabla-ext-cuda100-nccl2-mpi2-1-1 (Ubuntu18.04 default)
• nnabla-ext-cuda100-nccl2-mpi3-1-6
• nnabla-ext-cuda102-nccl2-mpi2-1-1 (Ubuntu18.04 default)
• nnabla-ext-cuda102-nccl2-mpi3-1-6
• nnabla-ext-cuda110-nccl2-mpi2-1-1 (Ubuntu18.04 default)
• nnabla-ext-cuda110-nccl2-mpi3-1-6

If you want to use a version of MPI not listed above, you need to build it from the source.

Additional setup may be required depending on your OS or environment. Please check Python Package Installation Guide for details.

Install Guide
To use C++ inference feature, follow the demonstration on MNIST inference in C++.

Demonstration

Build Guide
The "nnabla-ext-cuda" package is temporarily unavailable. Use of this package is not recommended. Please install nnabla-ext-cuda110, nnabla-ext-cuda102, nnabla-ext-cuda100 instead.
The following nnabla CUDA extension packages have been deprecated and the PyPi repository has been (or going to be) closed.

• nnabla-ubuntu16
• nnabla-ubuntu18
• nnabla-ext-cuda
• nnabla-ext-cuda80
• nnabla-ext-cuda91
• nnabla-ext-cuda92
• nnabla-ext-cuda101
• nnabla-ext-cuda90-nccl2-ubuntu16
• nnabla-ext-cuda100-nccl2-ubuntu16
• nnabla-ext-cuda100-nccl2-ubuntu18
• nnabla-ext-cuda101-nccl2-ubuntu16
• nnabla-ext-cuda101-nccl2-ubuntu18

The following "nnabla-ext-cuda" docker images have been deprecated.

• py27-cuda92
• py36-cuda92
• py37-cuda92
• py27-cuda92-v1.0.xx
• py36-cuda92-v1.0.xx
• py37-cuda92-v1.0.xx

We've decided to change nnabla's versioning policy to semantic versioning.
This change has been applied from version 1.1.0.

Python 2 is no longer be supported from v1.5.0.
CUDA8.0 is no longer be supported from v1.6.0.
Python3.5 is no longer be supported from v1.14.0.
CUDA9.0 is no longer be supported from v1.14.0.