We have released Neural Network Libraries v1.33.0!
Please see “Spotlight” section for important changes.
BugFix: Attribute error in numpy dtype aliases
In the latest numpy version, the deprecated dtype aliases including
numpy.bool exipired. Any of the previous versions of nnabla no longer work (even just importing nnabla) unless you downgrade the numpy to a proper version. We fixed the problem in this version.
Distributed execution requires NCCL 2.10.2 or later
Now reduce operations over devices or nodes such as all-reduce in nnabla-ext-cuda depends on the newer versions of NVIDIA’s NCCL APIs (2.10.2 or later). Please make sure that you have an appropriate version. You don’t have to care about it if you use wheel binaries which bundle the required CUDA libraries.
We add a new example of CLIPort.
CLIPort is a multi-task manipulation policy learning algorithm based on CLIP and Transporter!. Please try it!
Demo: CLIP zero-shot image classification in browser
You can quickly try zero-shot image classification demo using CLIP in browser. There are two options to run demos:
|Click this badge to run it on Colab. You don’t need to setup anything if you have Google account 😀|
|Web app||You can quickly setup a web app (Gradio) in a Docker container just with a few steps on your machine.|
- Fix a potential bug in memory-release behavior in auto forward mode
- dynamic numpy version
- set windows all-in-one path automatically when import
- fix cuda_array_copy runtime error on Windows
- fix segmentation fault issue on docker container
- Patch for directory traversal vulnerability
- Improve CenterNet
- Improve CenterNet
- Update default python version to 3.8 in documents and makefiles (CPU / GPU)
- fix namespace packages for wheel to pass twine check (CPU / GPU)
- build python with pyenv (CPU / GPU)
- add support for android-ndk-r25b
- adjust save/load function for optimizer state.
- support larger than 2GB file by refining the use of hdf5