Interest in support of LSTM functionalities

Feature requests
#1

Hello,

I am sending a quick message to signal my interest on the subject of improving LSTM support on microcontrollers.

As part of our work, we encountered errors when executing our Machine Learning algorithm while working with TensorFlow Lite due to unsupported operations (tied to Unidirectional Sequence LSTM & LSTM in general)

In order to solve our problem, we are are working on implementing a LSTM neural network on a sensor and would be interested to receive input & feedback on the subject.

I started this work using EdgeImpulse but had to switch to operating directly on the TensorFlow Lite library for this implementations, but should our implementation work it should end up as a contribution to TensorFlow Lite and would probably enable you to test this functionality with EdgeImpulse.

We will provide feedbacks on these implementation as well as a tutorial on how we developped our algorithms in order to test possible uses of LSTM in constrained networks.

Thanks all for any feedback you might provide, we will reach out to you with news & feedbacks regarding our experiments.
And thank you all for the work you’re doing with EdgeImpulse, I couldn’t have achieve the comprehension I actually have on the work I’m doing without your input and your platform.

Regards
Antoine BERNARD

1 Like
#2

Hi @Spiderweak, that is awesome news and something we’ve been tracking for a while too, great to see some work here. I assume you’re basing this off the LSTM kernels that are already in TensorFlow Lite?

Would be happy to support in any way we can, and perhaps get the help from Arm here as well to do CMSIS-NN versions of these kernels.

1 Like
#3

I just want to chime in that I think having LSTM capabilities in Edge impulse would be AWESOME and open up many new possibilities.

1 Like