Use nnom project with LSTM Network about nnom HOT 5 OPEN

hi, @kouroshD thanks for your interest,
The RNN types (inc. GRU/LSTM) are not currently supported because Keras does not support access to the intermediate output inside RNN cells. There is no straight forward method to quantize them. I am not sure I will/can implement it in the near future.

If you definitely need to use RNN layers, you may build the RNN cells using the supported layers (Dense, Conv, Sigmoid, TanH..) then concatenate the outputs. In this way, the intermediate output can be explicitly quantized thus it should be able to convert using the scripts.

Or, like many others just give up RNN types, and you can use Temporal Convolutional Networks (TCN) (e.g. WaveNet) to deal with time sequence data (speech/sensor data). TCN layers are also good at dealing time sequence data well. It required regular Convolution with Dilation. It is already implemented in NNoM.

from nnom.

Hi @kouroshD

I am not sure if you are still interested in RNN layers. I am now working on them
Just make the simple RNN cell tested. It provides no impact on accuracy comparing to Keras' RNN, and is available on this Develop branch. You could see the UCI-HAR example as the example for RNN at the moment.
LSTM will be followed up in this week if everything goes well.

Thanks,
Jianjia

from nnom.

Hi @kouroshD
LSTM and GRU are now working. Please check uci-har-rnn example

from nnom.

Hi @majianjia . Thank you for the updates and the support you have provided. In the following days, I will definitely look at them, and try to run them also my self.

Thanks,
Kourosh

from nnom.

This has been answered, and should be closed. CC @majianjia

from nnom.