Standalone impulses: now supports TensorFlow Lite

We’ve updated the example program to compile impulses on macOS and Linux to also support impulses that depend on TensorFlow Lite: example-standalone-inferencing. This should make it trivial to deploy your full algorithm (incl. DSP, neural networks and anomaly detection code) on any target that has a C++ compiler.

To get the latest and the greatest, just follow the instructions in Running your impulse locally.

(Note: The Mbed OS example already supported this, so no need to update there.)

SDK update

This requires a patch in the SDK that will be deployed on Monday, so if you try this out earlier (or on an impulse that you’ve created earlier), edit edge-impulse-sdk/porting/posix/ and add:

#include <stdio.h>


As advised in the instructions ,when I try to copy the raw features from test data from edgeImpulse project, the number of copied elements is not matching with the the number of elements expected. I am getting the following error.

Edge Impulse standalone inferencing (Mbed)
The size of your 'features' array is not correct. Expected 660 items, but had 657

Please advise
/cc @Hardik

Hi @paulphilip this is accelerometer data? In that case could you pad it with 0, 0, 0 for now? This might be related to the frequency issues we saw earlier on your project, but can investigate only tomorrow.

HI @janjongboom
Yes, This is accelerometer data. I will pad 0, 0, 0 and run the test. We saw variability between the features generated by the embedded target and online Model Testing module for same data. I can create another topic if you want.
/cc @Hardik

@paulphilip which project ID is this for (there are a few under your account)? I’ll take a look at it tomorrow.

@janjongboom This is related to Project 5392.