If you find your machine learning models to be a bit too large, TensorFlow’s Model Optimization Toolkit now comes with an implementation for post-training float16 quantisation. The new addition quantises model constants like weights to a reduced precision floating point data type, leading to smaller models with only little accuracy lost, according to the TF team.
To enable this algorithm, you’ll have to use the default optimisation settings on a trained float32 model with the supported types of the target spec set to float16. When using a GPU, the reduced precision parameter can be run directly, without the conversion step.
Istio dishes out bug fix releases
The team behind service mesh Istio has released versions 1.2.3 and 1.1.12 of its project. While both releases are meant to fix an issue that could lead to a sidecar infinitely forwarding requests to itself, v1.2.3 also gets rids of Redis connectivity issues and a bug that led to the Stackdriver adapter shutting down after telemetry started.
Other additions should mitigate case-sensitivity in regex-based HTTP URI matching for virtual service and help skipping Kubernetes JWT authentication when SDS isn’t in use.
GNU C library hits 2.30
A couple of days ago, “the” C library in the GNU/Linux systems glibc saw the release of version 2.30. Major new features include support for unicode 12.1 and a twalk_r function, which is similar to regular twalk but passes an additional caller-supplied argument to the callback function.
Other than that support for the Minguo calendar for locales zh_TW, cmn_TW, hak_TW, nan_TW, and lzh_TW, as well as an entry for the new Japanese era has been added. The new release also tackles undefined behaviour, by letting a variety of memory allocation functions fail when the total object size is larger than PTRDIFF_MAX. More information can be found in the official announcement.
Cloudera Altus Director ups Java game
Cloudera’s tool to help with the management of Cloudera clusters on the big three cloud platforms AWS, Microsoft Azure, and Google Cloud Platform is now available in v6.3. The release is the first to run on the OpenJDK 11 and comes with a number of updates to the cloud provider plugins. Improvements included in those are for example support for smaller VM types for Azure, as well as shorter instance names and shared VPCs for GCE.
And good news for floppy lovers
A couple of days after Linus Torvalds declared Linux’s floppy driver as orphaned, developer Denis Efremov has taken over maintainership of the poor little thing. Efremov is no stranger to the code, having described floppy ioctls in Google’s syzkaller. He also has hardware for testing changes available, so let’s hope those readers will stay intact for the next couple of years.