How to rewrite Keras Stateful LSTM in pure Tensorflow?

Asked May 26 '19 at 17:52

Active May 26 '19 at 18:06

Viewed 143 times

Anyone could share an idea/blog/code snipet on how to convert a Keras Stateful LSTM into pure Tensorflow model? And then train it on batch..

Tensorflow doesn't support Keras Stateful LSTM on TPUs. Their devs refused to fix it.

I have tons of TPU time reserved and no way to use it for now. Any help is appreciated.

Model example and code to train:

model = Sequential()
model.add(LSTM(neurons, batch_input_shape=(window_size, n_steps, inputs_n), stateful=True)) 
model.add(Dense(outputs_n, activation='sigmoid'))
…
H = model.train_on_batch(X, y)

GitHub issue: https://github.com/tensorflow/tensorflow/issues/28837

edited May 26 '19 at 18:06

asked May 26 '19 at 17:52

Boppity Bop

9,613
13
72
151

Can you share the link to github issue that you raised for this? What is your TF version? – dgumo May 26 '19 at 18:03
version doesn't matter. this doesn't work in any TF version. – Boppity Bop May 26 '19 at 18:06
1

Stateful models work on TPUs but they are atrocious for performance. Batches has to be processed in order making it impossible to parallelize the calculations. – Björn Lindqvist May 31 '20 at 17:06

How to rewrite Keras Stateful LSTM in pure Tensorflow?

0 Answers0