Getting GRU working

Report bugs
#1

Hello,

I would love to run my keyword spotting example from the Coursera course using a GRU network instead of the 1D CNN and compare the two. I edit the keras notebook to implement a GRU network which trains well but the job fails when it has to convert it to a tflite model. Is there a work around for this?

#2

Hi @mnyamukuru, note that you can currently not deploy this on device because the underlying inferencing engine does not support recurring neural networks.

We should not fail when converting though, unless there’s ops in the network that are unsupported even by full TensorFlow Lite. What’s your project ID? I’ll take a look.

#3

Thanks @janjongboom, my project ID is 21946. Do I need to make it public for you to take a look?

#4

@mnyamukuru Sorry for the late reply - no need to make it public, will check in the morning.

#5

@janjongboom No worries at all. Thank you for checking this out.

#6

@mnyamukuru We’ll be rolling out a fix this week. Note that the on-device performance metrics will not show proper memory info, as we use TensorFlow Lite for Microcontrollers underneath (which misses these ops), but at least training will succeed. :slight_smile: :rocket:

edit This is now released.