Question on fluid.layers.dynamic_lstm

[jczaja]

Hi,

I have question of returned values of fluid.layers.dynamic_lstm. According to :
http://paddlepaddle.org/documentation/api/en/0.14.0/layers.html#dynamic-lstm

dynamic_lstm does return hidden state as well as cell state. My questions are:

1. what is the intention of returning cell state by dynamic_lstm?
2. In what situation it can be useful for a model to have an access to cell state when using dynamic_lstm?
3. Can you point me fluid model that is actively(not ignoring) using cell state as returned by dynamic_lstm ?

Best Regards,

[luotao1]

Discussed with @qingqing01 :

what is the intention of returning cell state by dynamic_lstm

• Desired by the tensorflow API, we return cell state as well.
  

  Paddle/benchmark/tensorflow/rnn/rnn.py

  Line 63 in decda73

  output, _cell_state = rnn.rnn(cell, incoming, dtype=tf.float32)

• Backward of lstm use cell as well.
  

  Paddle/paddle/fluid/operators/lstmp_op.h

  Line 240 in decda73

  auto* cell_out = ctx.Input<LoDTensor>("Cell");

2. In what situation it can be useful for a model to have an access to cell state when using dynamic_lstm?
3. Can you point me fluid model that is actively(not ignoring) using cell state as returned by dynamic_lstm?

There is no fluid model using cell state now.

[jczaja]

Thanks @luotao1,

Meantime I found an additional argument after having cell state returned via python interface which is
that if sequence is to long to have it processed at once via dynamic_lstm then it could be split into multiple execution of dynamic_lstm's and returning cell_state as well as providing input c0 will allow this scenario to happen.