EloquentTinyML icon indicating copy to clipboard operation
EloquentTinyML copied to clipboard

Support for RNN (LSTM)

Open expeon07 opened this issue 2 years ago • 1 comments

Hi, would you know if there is a way to implement RNN in TFLite micro? I use LSTM for time series prediction but I have this error.

Only 1 subgraph is currently supported.

expeon07 avatar Mar 25 '22 10:03 expeon07

I actually had the same question when I tried to run a LSTM model last summer for a paper submission: https://github.com/dustinjoe/TinyML-with-Cloud-for-PHM At that time, no official support for RNNs, I remember someone from Tensorflow team mentioned some memory issue. But this time point, official LSTM support seems to be on the way right now: https://github.com/tensorflow/tflite-micro/pull/1044

dustinjoe avatar Apr 16 '22 15:04 dustinjoe

LSTM is indeed supported as many users reported. Be sure to use Tf==2.2.3

eloquentarduino avatar Sep 02 '22 06:09 eloquentarduino

@expeon07 @dustinjoe @eloquentarduino There is an operation called UnidirectionalSequenceLSTM in the official library description. But I am not sure whether that's LSTM you want or not.

whubaichuan avatar Nov 04 '22 18:11 whubaichuan