SensiML Cloud Release Notes

Current Release

2022.2.2 (5/18/2022)

What’s New

Major Features

Minor Features

  • Improved descriptions of supported platforms

Past Releases

2022.2.1 (4/26/2022)

What’s New

Major Features

  • Adds support for the Silicon Labs xG24 Dev Kit

  • Adds support for accelerating NN ops using the Matrix Vector Processor on the Silicon Labs xG24 Dev Kit

  • Adds support for specifying int8/uint8 inputs to TensorFlow models (moving forward we will prioritize int8 support as most accelerators do not support uint8 inputs)

Minor Features

  • The confusion matrix and accuracy in the Test Model tab of the Analytic Studio are now only compared against labeled ground truth data. Previously, unlabeled regions would be considered as the Unknown label.

  • Adds a backoff parameter to the Segment Filter Energy threshold allow for N segments to pass after the threshold is triggered

  • Adds support for storing a color value in against labels

  • Improved error messages responses

Bug Fixes

  • Fixes an issue where pipelines with downsampling filters report the wrong segment length

  • The Feature Cascade feature transform now correctly drops segments with noncontiguous sections

2022.2.0 (3/23/2022)

What’s New

Major Features

  • Adds power spectrum feature generator

  • Performance and stability improvements

  • Support for Arduino Nicla Sense ME Platform

Minor Features

  • Adds support for on device model profiling for Silicon Labs Thunderboard Sense 2

  • Improved performance of Knowledge Pack firmware generation

  • Improved support for tracking TensorFlow training

  • Incorrect requests for capture files now returns the name of the file that does not exist instead of only raising a does not exist exception

Removed Features

  • Removed support for the auto execution_type parameter from the pipeline API (/project/<uuid>/pipeline/<uuid>/)

  • Deprecated version 1 of the Generate Knowledge Pack API

    • /project/<uuid>/knowledge-pack/<uuid>/generate_lib/

    • /project/<uuid>/knowledge-pack/<uuid>/generate_source/

    • /project/<uuid>/knowledge-pack/<uuid>/generate_binary/

Bug Fixes

  • Fixes an issue with code generation for device profiling on some platforms

2022.1.1 (2/15/2022)

What’s New

Minor Features

  • Adds support for storing pipeline hyper_params to the SandBox API

  • Performance and scalability improvements

2022.1.0 (2/2/2022)

What’s New

Major Features

Minor Features

  • Improved documentation for supported DSP and ML library functions

  • Improved model registry support for TensorFlow models

  • Improved AutoML training metrics results

  • Adds support for removing unknown patterns from the PME after training

  • Adds a new segment filter Segment Energy Threshold Filter

  • Updated the Adaptive Windowing Segmenter algorithm to allow taking the absolute value of the signal

  • Improved support for TensorFlow Lite files generated from the most recent versions of TensorFlow

Bug Fixes

  • Fixes an issue with the logs for custom transforms

  • Fixes an issue with the MAX_VECTOR_SIZE not being generated for some projects

2021.2.9 (12/27/2021)

What’s New

Minor Features

  • Improved support for tracking pipeline CPU usage and runtime

  • Additional query optimization improvements

  • Improvements to query error messages on failure

2021.2.8 (11/12/2021)

What’s New

Minor Features

  • Adds API to check if the query cache is in sync with the projects training data

  • Adds support for the latest firmware for Microchip Technology SAMD21 ML Eval Kit (SAM-IoT WG)

Bug Fixes

  • Fixes an issue with a missing header file in the Nordic Thingy binary download

2021.2.7 (11/04/2021)

What’s New

Major Features

  • Adds support for automatic onsemi sensor configuration file generation onsemi RSL10.

  • Adds support for including the scratch buffers with custom feature generators, see documentation.

  • Adds support for custom feature generators that produce more than one feature, see documentation.

Minor Features

  • Adds better error handling and logging for custom feature generator upload

  • Adds better logging for bonsai decision trees

2021.2.6 (10/29/2021)

What’s New

Major Features

  • Adds support for training fully connected neural network

  • Adds anomaly detection for AutoML optimizations

Minor Features

  • Adds streaming decimation filter

  • Adds peak frequency feature generator

Bug Fixes

  • Fixes codegen issue for multiple sensor filters

  • Fixes query caching issue with overwriting cache

2021.2.5 (10/06/2021)

What’s New

Major Features

  • Queries are now cached when executed by a pipeline or by calling the cache query API. By caching the query, your dataset is versioned to the time the query is created. This will speed up model execution time and allow you to continue to update your dataset, yet still build models/test models against older versions of the dataset. Query caches can be updated by calling the cache query API.

  • Adds support for TensorFlow Lite for Microcontrollers inference on Raspberry Pi 3/4

2021.2.4 (9/22/2021)

What’s New

Major Features

  • Adds support for onsemi RSL10 Sense

Minor Features

  • Adds the ability to return partial segments for general threshold segmenter if the capture file ends before segmentation finishes

Bug Fixes

  • Fixes codegen issue for segment transforms which had magnitude transforms sensors

  • Fixes codegen for correlation cross column feature generator

  • Fixes sensor configuration for Microchip Knowledge Packs

  • Always include testdata.h with Arm GCC Knowledge Packs

  • Improved server stability

2021.2.3 (9/07/2021)

What’s New

Major Features

  • Adds support for Windows x86 Knowledge Pack libraries

Bug Fixes

  • Fixes issue with Thunderboard Sense 2 Knowledge Pack binary sensor configuration

  • Improved support for 8/16 bit microcontroller architectures

2021.2.2 (8/24/2021)

What’s New

Major Features

Minor Features

  • Adds additional information to model.json in the Knowledge Pack download including Knowledge Pack summary and the expected input sensors

Beta Feature

  • Add ability to add custom functions as part of the your DSP/ML pipelines, see documentation (here)

Bug Fixes

  • Fixes issue with password reset

  • Fixes a bug where capture configurations created for MQTT/SN would incorrectly configure sensors for Knolwedge Pack SensiML AI applications

2021.2.1 (7/29/2021)

What’s New

Minor Features

  • Adds ranges to all transform fields for better validation

  • Improved Knowledge Pack profiling information

Feature Preview

  • Add ability to add custom functions as part of the your DSP/ML pipelines (contact us for access)

Bug Fixes

  • Fixes an issue where team admin could not delete some of their users

  • Fixes some code generation issues for the MCHP and Android NDK builds

  • Recognition now return an empty list instead of raising an exception when no segments are found

  • Improved load balancing and server stability

2021.2.0 (6/30/2021)

What’s New

Major Features

  • Adds Knowledge Pack support for Microchip Technology SAMD21 ML Eval Kit (SAM-IoT WG)

  • Adds Knowledge Pack profiling option for cycle measurements of feature generators and classifier inference on platforms that support the DWT_CYCCNT register

  • Adds Knowledge Pack support for Android NDK

  • Adds support for custom pipelines as part of AutoML search

  • Adds support for using feature cascade as part of hierarchical model creation

Bug Fixes

  • Minor bug fixes and performance improvements

2021.1.2 (5/17/2021)

Bug Fixes

  • Fixes source code download file extension to be .zip instead of tar.bz2

  • Removes deprecated kb_print_model_result from sml_recognition_run.c

  • Fixes issue with using custom feature generators with hierarchical models

  • Fixes recognition mode failing for Bonsai decision tree models

2021.1.1 (4/19/2021)

What’s New

Major Features

  • Adds additional Knowledge Pack APIs which enable finer control of pipeline step execution

  • Improved tuning of the TensorFlow memory usage to reduce the overall memory footprint of TF Lite models

  • Adds a model_json.h file to the library download which contains information about the Knowledge Pack pipeline

Bug Fixes

  • Fixes a bug in input contracts that affected some sample feature generators

  • Fixes a compile issue for QuickAI Knowledge Packs

2021.1.0 (3/04/2021)

What’s New

Major Features

  • Adds Support for SparkFun Thing Plus - QuckLogic EOS S3

Bug Fixes

  • Fix Audio Recognition for QuickFeather Binary where the audio Flag was not being set correctly

2020.3.1 (12/04/2020)

What’s New

Major Features

2020.3.0 (11/10/2020)

What’s New

Major Features

  • Adds additional segment transforms to normalize segments

  • Allow source code download for enterprise and standard customers

  • Adds Cortex M7, Nano 33 library build option

  • Adds augmentation libraries for segment data

  • Adds ability to specify classifiers and training algorithms for AutoML search

  • Improved logging for AutoML pipelines

  • AutoML now returns the fitness score for all pipelines searched across

Bug Fixes

  • Fixes issue where scaling was not performed to the full width before MFCC feature extraction

2020.2.2 (09/08/2020)

What’s New

Minor Features

  • Adds additional Feature Extractors (linear regression stats, zero crossings, positive zero crossings, negative zero crossings, shape median difference, shape absolute median difference)

Bug Fixes

  • Performance improvements and minor bug fixes

2020.2.1 (08/19/2020)

What’s New

Major features

Minor Features

  • Adds support for multi-channel DTW

  • Adds additional feature generators

  • Performance improvements

Bug Fixes

  • Improved validation for custom sensors

2020.2.0 (07/17/2020)

What’s New

Major features

  • Adds ability to test multiple captures in a single calls.

  • Adds ability to generate confusion matrix for a target label when running test data.

  • Adds ability to download source file for enterprise level accounts.

  • Allow user to specify which classifier algorithms will be used as part of AutoML optimization.

Minor Features

  • Increased the number of decision trees available in random forest and boosted tree ensembles during inference.

Bug Fixes

  • Fixes issue where sensor columns could be generated in different order than sensor configuration specified

2020.2.0 (07/17/2020)

What’s New

Major features

  • Adds ability to test multiple captures in a single calls.

  • Adds ability to generate confusion matrix for a target label when running test data.

  • Adds ability to download source file for enterprise level accounts.

  • Allow user to specify which classifier algorithms will be used as part of AutoML optimization.

Minor Features

  • Increased the number of decision trees available in random forest and boosted tree ensembles during inference.

Bug Fixes

  • Fixes issue where sensor columns could be generated in different order than sensor configuration specified

2020.1.6 (05/04/2020)

What’s New

Major features

  • Adds ability to specify decision tree of strong classifiers to optimize against

Minor Features

  • Adds Interleave feature generator for combining sensors channels

  • Performance improvements and bug fixes

2020.1.5 (04/14/2020)

What’s New

Minor Features

  • Adds a threshold setting to tflite post processing to return unknown below threshold value

  • Improvements to database query performance

  • caching optimizations for increased performance

2020.1.4 (04/02/2020)

What’s New

Minor Features

  • Improvements to tensorflow-lite micro support

  • Improvements to database query performance

  • Adds API’s to Knowledge Pack for setting feature vector directly as well as recognizing feature vector

2020.1.3 (03/23/2020)

What’s New

Minor Features

  • Improvements to tensorflow-lite micro support

  • Additional Bulk API’s for faster egress

  • Minor bug fixes

2020.1.2 (03/03/2020)

What’s New

Minor Features

  • Adds new bulk API’s to improve performance of uploading/deleting/updating multiple segments at the same time

  • Improved performance for a number of feature transforms and extractors

  • adds beta support for tensorflow lite micro

  • Adds a more detailed query statistics endpoint for richer information

  • Adds ability to include the segment_uuid in the query

2020.1.1 (02/04/2020)

What’s New

Minor Features

  • Adds new segment filter threshold algorithm

2020.1.0 (01/20/2020)

What’s New

Major Features

  • Adds Profiling and better debug logs to Knowledge Packs

  • Increased Max Feature vector size to 2048

Minor Features

  • Adds option to use less than comparison as part of the windowing threshold algorithm

  • Adds DTW to Hierarchical Clustering Training Algorithm

Bug Fixes

  • Fixes issue where auto segmentation heuristics would generate invalid parameter settings

  • Fixes issue where DTW distances larger than uint16 were not being truncated

  • Fixes issue where Mayhew board could be configured incorrectly when generating a Knowledge Pack

  • Minor bug fixes and performance improvements

2019.3.6 (11/05/2020)

What’s New

Major Features

  • Support for sqlite optimized DCL

Minor Features

  • Computed distances for PME are stored as part of the knowledge pack

  • Model Size is stored as part of the Knowledge Pack

2019.3.5 (10/21/2020)

What’s New

Major Features

  • Support for AD7476 Sensor at up to 1Mhz

Minor Features

  • Additional API for flushing model ring buffer to clean state

  • Stability Improvements

2019.3.4 (10/08/2020)

What’s New

Major Features

  • Boosted tree classifiers now part of autosense optimization routine

Bug Fixes

  • Stability improvements to pipeline performance scheduling

2019.3.3 (09/23/2020)

Bug Fixes

  • Fixed issue where feature validation was to strict for some validation methods

  • Improved error message reporting for Knowledge Pack downloads

2019.3.2 (09/19/2020)

What’s New

Major Features

  • Implementation of bonsai decision tree classifier which combines dimensionality reduction with an efficient tree classifier structure

  • Store full results from train, validation and test in the Knowledge Pack

  • Performance and stability improvements

Bug Fixes

  • Fixed issue where some column name characters weren’t being correctly sanitized during firmware generation

2019.3.1 (08/22/2020)

What’s New

Major Features

  • Addition of Dynamic Time Warping as a distance metric for the PME classifier

  • Added two new model selection methods (metadata k-fold, and stratified metadata-kfold)

2019.3.0 (07/30/2020)

What’s New

Major Features

  • Support for SensorTile 1.0 Knowledge Pack Binary and Library Builds

2019.2.0 (06/27/2020)

What’s New

Major Features

  • Adding for under sampling the majority class in order to balance a data set

  • Adding support as part of auto sense pipeline for balancing data sets

  • Adding support for supplying a user specified validation method to the auto sense pipeline

  • Adding support for specifying capture uuid as part of the metadata in a query

Bug Fixes

  • Fixes issue with some queries failing due to the names of the metadata

2019.1.4 (06/11/2020)

What’s New

Major Features

  • Adding support for Chilkat Hardware Knowledge Pack creation

Minor Features

  • Improvements to accuracy calculations of AutoSense pipeline

2019.1.3 (06/04/2020)

What’s New

Minor Features

  • Speed optimizations for recognize signal

  • Project statistics now returns information about all captures and segments

Bug Fixes

  • Fixed issue where having a decision tree ensemble and gradient boost classifier in the same model would fail to compile

  • Fixed issue where terminating a pipeline wasn’t always removing it from the active pipeline queue

2019.1.2 (05/22/2020)

What’s New

Minor Features

  • Speed improvements to AutoSense pipeline and underlying training algorithms

  • QuickAI SDK 1.2.1 release.

Bug Fixes

  • Fixed issue in QuickAI 1.2 SDK when recording using ADC with 3 channels

2019.1.1 (05/14/2020)

What’s New

Minor Features

  • Improved server performance to increase number of batch jobs executed in parallel during pipeline execution

Bug Fixes

  • Fixed issue where uploading a large feature vector file could being split up before being sent to the TVO or selector set steps

2019.1.0 (05/05/2020)

What’s New

QuickLogic S3 AI Recognition updates

  • Support for recognition from 1-4 Channel ADC Mayhew at 16khz

  • Support for recognition for 1 Channel ADC Mayhew at 100khz

  • Support for recognition of Audio at 16khz

  • Support for IMU recognition from 25-1600Hz

AutoSense Pipeline

  • Includes random forest algorithm as part of the search over the classifier space

  • Now allows users to select whether or not to use a classifier that will return unknown when it is unsure of the result

  • Allows users to build a submodel using autogrouping of classes

Other Major Features

  • Added a boosted tree ensemble classifier that performs binary classification

  • Captures can now be associated with the capture configuration that created them

  • Improvements to upload speed of metadata labels

  • Status messages now return more information about running pipelines

Notes: The minimum SensiML client version is 2019.1.0

Bug Fixes

  • Fixed issue where pipeline would appear to be in the queue but actually be running

2.5.1 (02/28/2020)

What’s New

  • General performance improvements

  • General Security improvements

Notes: The minimum SensiML client version is 2.5.3

2.5.0 (01/15/2020)

What’s New

  • Additional Board Support for ChilKat platform

  • Feature generators automatically iterate through input columns and specify their correct input

  • Major Server Stability Improvements for handling larger data sizes

  • AutoGenerated Knowledge Packs now support knowledge rehydration, previously only pipeline rehydration was supported

  • New segmentation, feature generation and sampling algorithms added

  • Any segmenter can be used as input to cascade feature

  • Ability to specify multiple datafiles as input to a pipeline

  • Better error messages returned for many endpoints

  • Naming convention for classifier “PVP” has been deprecated, all pipelines are required to use the name “PME” for this classifier

Notes: The minimum SensiML client version is 2.5

Bug Fixes

  • Knowledge Pack rehydration now accounts for feature family generators

2.4.0 (11/09/2018)

What’s New

  • Additional Feature generator

    • Convolution Max

  • Additional Streaming Filter

    • Downsample

    • High Pass

  • High Frequency Data Collection using the Quick AI Module

  • HW acceleration support for QuickAI Hard Neurons

  • DSP optimizations for Knowledge Packs built targeting arm m3/m4 processors

2.3.3 (10/31/2018)

Bug Fixes

  • Support for segments up to length 8192

  • Server Stability Improvements

  • Improvements to error messages

  • Improvements to QuickAI FFE data capture

2.3.2 (10/24/2018)

What’s New

  • Adding Support for QuickAI low power ffe for pre-processing sensor data

  • Increase number of classes supported by PME reinforcement learning

  • Adding model.json to Knowledge Pack download that has information about the contained model

  • Stability and speed improvements to Auto Sense pipelines

Bug Fixes

  • Fix Hierarchical Clustering bug where Nan was being returned and causing a crash

2.3.1

What’s New

  • Custom validation method can be used by the automation engine

  • Additional API’s for Knowledge Pack to enable loading/saving models to/from flash

    • flush_model

    • get_model_header

    • get_model_pattern

  • Additional API’s for Knowledge Pack to support cascade windowing with reset

2.3.0

What’s New

Major Features

  • QuickAI board now supports capture and SensiML recognition without re-flashing

  • Adding support for reinforcement learning to PME algorithm on the device

  • Adding API’s to the c Knowledge Pack to retrieve information about the model such as the class map, model patterns, model map etc. (see kb.h for full list of API’s)

Minor Features

  • Pipeline status is returned during pipeline execution

  • General stability improvements and bug fixes

  • Return a model.json file with all Knowledge Packs that describes the model

  • Bug Fix where terminating a pipeline didn’t terminate correctly all the time

2.2.2

What’s New

Major Features

  • Adding software support for QuickAI board (hardware accelerated classification and FPGA feature generation acceleration support will be added in future releases)

  • Adding ability to use the emulator for hierarchical models via recognize_signal

  • Added Knowledge Pack support for decision tree ensemble trained via random forest training algorithm

  • Adding new class of feature generators (cross column) for use in comparing features across sensor columns

Minor Features

  • Hierarchical models now generate their calls to arbitrary depth

  • Added two tail t-test based feature selector

  • Min max scale now accepts partial parameters and will scale the rest

  • General stability improvements and bug fixes

2.2.1

What’s New

Major Features

  • Library Code generation for RPI, Arm and Ubuntu with gcc version 7.2

  • Support for constructing a feature vector from multiple sliding windows

Minor Features

  • Added Moving Average Sensor Transform

  • SensiML Labs (Experimental Features)

  • Random Forest Classifier (Important: This feature is in an early concept stage, it cannot be used on a device)

  • Adding an API to add new patterns to the device while it is running

Bug Fixes

  • Schema error on upload now returns message for which fields are incorrect

  • Minor bug fixes and stability improvements

2.2.0

What’s New

New Platforms

  • FreeRTOS

Major Features

  • Automation now has a cross validation option to prevent under/overfitting

  • Adding double peak segmentation algorithm (a key based segmentation algorithm)

Minor Features

  • KP download now includes option to explicitly define source (Audio, Motion, Custom)

  • Combine labels allows renaming labels and creating new groups

  • Auto Combine label, automatically splits many events into two groups

SensiML Labs (Experimental Features)

  • Adding Cascade Windowing Segmenter (Important: This feature is in an early concept stage, it cannot be used on a device)

  • Adding Bonsai Decision Tree Classifier (Important: This feature is in an early concept stage, it cannot be used on a device)

Bug Fixes

  • Fixes to pipeline seeds for automation

  • Fix overflow bug for raw data

2.1.3

What’s New

Major Features

  • Metadata Separator for choosing the best class split in Hierarchical models

2.1.2

What’s New

Major Features

  • Adds outlier removal samplers for improving model accuracy

Bug Fixes

  • Fixes bug with Hierarchical Models not returning correct results

2.1.1

What’s New

Major Features

  • Support for building Knowledge Packs using audio data for Nordic Thingy

Minor Features

  • Added ability to specify a capture file as the test data in recognize signal

  • Add new transform for grouping labels into subgroups. See Combine Labels in docs

  • Added ability to use entire segment from parent model for submodels

  • Adds a padding option to min max scale which can improve classification accuracy in some cases

Bug Fixes

  • Fixed issues with some feature generators (conv avg, min percentile, sum)

  • Fix some issues with correlation feature selectors

2.1.0

What’s New

Major Features

  • Knowledge Pack compatibility with Nordic Thingy

Minor Features

  • Server side optimizations for faster query performance

Bug Fixes

  • Queries not correctly selecting segments when labels have been created by more than one autosegmenter

  • Fix integer overflow in magnitude sensor transform when more than 2 axis are used

2.0.0

What’s New

Major Features

  • Pipeline Automation - Automated pipelines reduce the amount of code you have to write to find good features and pipeline parameters. Use pre-defined Pipeline Seeds (“Basic Features”, “Advanced Features”, “Downsample Features”, “Histogram Features”) - or define your own pipeline and let the automation API fine-tune the parameters with its genetic algorithm.

  • Convolution/Submodels - A segment captured and classified by one model, can be fed and used by other models which can use the entire segment or perform their own segmentation

  • Segmenter Discovery/Optimization - Given a labeled data set, the server will optimize the parameters for detecting those segments from the signal

Minor Features

  • Optimization for core function for latency and memory usage

  • SensiML Python SDK list functionality - for most types of objects on the server is now supported. client.list_* to allow easy information discovery

  • Knowledge Packs can now be saved and retrieved by name. models from multiple pipelines/projects can now be combined into a single binary file

  • Grid Search can now be performed over the validation, classifier and training algorithm of the tvo step. The option to replace Hierarchical Clustering with Neuron Optimization in the grid search has been removed. To use Neuron Optimization, use it in the pipeline tvo step like all other training algorithms

  • Addition of new feature generators and transforms

    • Transform: Pre-emphasis Filter

    • Feature Generator: MFCC

    • Feature Generator: FFT

    • Feature Selector: Custom Feature Selection

    • Validation Method: Set Sample Validation