We are pleased to announce the release of NxSDK 0.9.0. The installation files are located at /nfs/ncl/releases. Installation steps stay the same as of 0.9.0. Please refer to getting started guides within Startup Materials[also available in nxsdk-apps-0.9.0.tar.gz/docs]

The release is available for all participants who have signed the INRC participation agreement and have access to the INRC Cloud.

Highlights

Improved DNN Integration via NxTF Interface, Slayer and SNN Toolbox

Run your pre-trained Keras or Slayer models now on Loihi. SNN Toolbox now supports Loihi as a backend configuration. Or, use NxTF layers with Keras interface to compile models directly.

Integrate & Pipeline Models, Processes & Input Generators using Composability

Create pipelines which can compose of models and features developed by others and use the composability framework to compile it down to Loihi. Promotes feature reuse and faster collaboration using a common interface

Compact packaging using Embedded Execution Engine for running networks and models on devices with memory constraints

Improved support for running models on devices such as ODROID or UP Squared where you just want to ship pre-packaged network with snips and get it running using host snips for external interaction

Improved Performance - Compilation, DVS, Spike Injection and more

0.9.0 comes with many performance improvements. Serializing/Deserializing of networks have been improved. Spike Injection now uses a compact encoding making it 10x faster

New Features/Improvements

General

NxTF : Users should be able to create (or port from existing) deep neural networks on Loihi. NxTF implements the Keras interface. Users can also use SNN Toolbox to port exist DNNs to Loihi.
Composability : Provides an interface to enable modules, applications and frameworks to be stitched together and compiled down to the NxBoard Graph.
Slayer : Slayer integration has been added within nxsdk_modules. It allows to run models pre-trained using slayer on Loihi
**Embedded Execution Engine ** : A new execution engine to run on embedded platform with minimum memory requirements
Concurrent Host Snips are now available for users to run concurrent execution on host while neurocore execution is in progress
DVS :
- Performance improvement - Access to the snip for users to modify the spike injection
Basic Spike Generator : Optimized the spike gen (5-10x improvement)
- Users can specify which LMT it will run on (removing the requirement that they are on the same LMT in same phase)
Dependencies for nxsdk_modules have been moved out of nxsdk into its own requirements.txt. Hence, you might need to optionally run pip install -r nxsdk-apps/nxsdk_modules/requirements.txt to fetch dependencies for some of these modules.

NxNet

NxSDK compiler now supports user specified custom partitioners and mappers. A custom partitioner and mapper can be individually specified for each network in a network tree.
Updated documentation for APIs to use and query resource maps

Jupyter Tutorials	Feature Covered	Description	Category
s_multicx_neuron_self_reward.ipynb	Mutlicompartment neurons, reward channels	This tutorial illustrates learning a synaptic weight to synchronise spiking of soma and dendrite compartments	NxNet
t_vMinExp_and_vM axExp.ipynb	Compartment settings	This tutorial demonstrates the usage of vMaxExp and vMinExp in compartments	NxNet

Python Tutorials	Feature Covered	Description	Category
tutorial_26_conc urrent_host_snip s_with_yarp_int egration	Host Snips	This tutorial demonstrates using yarp ports with concurrent host snips.	Snips

NxSDK Modules

Jupyter Tutorials	Feature Covered	Description	Category
nxsdk_modules/dnn/ tutorials/a_nxtf_ partitioning_mappi ng	DNN mapping	This tutorial demonstrates how to use the NxTF framework to partition a DNN for Loihi.	NxTF
nxsdk_modules/dnn/ tutorials/b_nxtf_ mnist	DNN mapping	This tutorial demonstrates how to use the NxTF framework to run a DNN on Loihi.	NxTF
nxsdk_modules/dnn/ tutorials/c_snntb_mnist	DNN mapping	This tutorial demonstrates how to use the SNN Toolbox to run a DNN on Loihi.	NxTF
nxsdk_modules/dnn/ tutorials/d_mnist_using_composables	Composables	This tutorial demonstrates how to use the composability interface to compose DNN with Input Generator and run the MNIST network on Loihi.	NxTF
nxsdk_modules/slay er/tutorials/nmnist /NMNIST	SLAYER fully connected	This tutorial demonstrates how to run a SLAYER-trained fully connected model on Loihi.	SLAYER
nxsdk_modules/slay er/tutorials/gestur e/gesture	SLAYER convnet	This tutorial demonstrates how to run a SLAYER-trained convolutional model on Loihi.	SLAYER

NxCore

Embedded Execution: Use board.generateEEEArtifacts() to package and distribute execution artifact to be run in an embedded environment. This is primarily useful when the only interaction is via host snips (All communication and execution happen between host and neurocores/lmt)
Serialization/Deserialization performance has been improved by using binary format. This makes board.dumpNeuroCores(...) and board.dumpNeuroCores(...) faster

Python Tutorials	Feature Covered	Description	Category
tutorial_22_host _snips_with_yar p_integration	Host Snips	This tutorial demonstrates using yarp ports with host snips.	Snips
tutorial_23_host _snips_with_ros _integration	Host Snips	This tutorial demonstrates using ros with host snips.	Snips
tutorial_24_cont rol_loop_using_ rospy	Snips	This tutorial demonstrates using ros with channels from superhost useful for proto-typing with ROS.	Snips

API Changes/Deprecations

Custom partitioners that partitioned an entire network tree is no longer supported. Custom partitioners are now specified on a per NxNet object level.
In createChannel : elementType would be deprecated in 0.9 in favor of messageSize, which provides more flexibility of specifying size of a message.

Major bug fixes for release

Fixed the slowness issue while loading neurocores (board.loadNeuroCores) on KapohoBay
Fixed a bug related to logical id spilling over core boundary

NxSDK 0.9.0 is now available