is it possible to train an LSTM neural network with Ossian/Merlin?

[f-e-l-i-x]

I have specified in /Ossian/train/.../speakers/.../naive_01_nn/processors/acoustic_predictor/config.cfg:

`
hidden_layer_size : [1024, 1024, 1024, 1024, 512]
hidden_layer_type : ['TANH', 'TANH', 'TANH', 'TANH', 'LSTM']
...

sequential_training : True
`

run_merlin runs fine, but store_merlin produces an error complaining about
' assert len(param_vals) == len(layer_types) * 2 ## W and b for each layer' in store_merlin.py

Can Ossian be used with an LSTM neural network?

[rnomals]

Hi f-e-l-i-x,
I've got the same issue when storing Merlin models after trained with LSTM. Did you solved it or find any particular reason behind that issue.
Thanks..

[f-e-l-i-x]

I haven’t found out anything in this regard. I would very much like to know.

[RobinAlgayres]

Hi guys,
I was also facing the same issue. I forked Ossian on my github so that it can run LSTM on OSSIAN.
It boils down to two things :

• changing store_merlin_model.py so that it can store LSTM in addition to DNN.
• adding a forward pass for lstm in NN.py.

It is not completely tested yet but I have already trained Ossian with one lstm+tanh layer . If you face issues while training multiple layer lstm please let me know.
Note that any other models will not run (Bilstm included, I might add a patch to it later)
Best.