Table Of Contents
Table Of Contents

RNN

mxnet.ndarray.RNN(data=None, parameters=None, state=None, state_cell=None, state_size=_Null, num_layers=_Null, bidirectional=_Null, mode=_Null, p=_Null, state_outputs=_Null, out=None, name=None, **kwargs)

Applies recurrent layers to input data. Currently, vanilla RNN, LSTM and GRU are implemented, with both multi-layer and bidirectional support.

Vanilla RNN

Applies a single-gate recurrent layer to input X. Two kinds of activation function are supported: ReLU and Tanh.

With ReLU activation function:

\[h_t = relu(W_{ih} * x_t + b_{ih} + W_{hh} * h_{(t-1)} + b_{hh})\]

With Tanh activtion function:

\[h_t = \tanh(W_{ih} * x_t + b_{ih} + W_{hh} * h_{(t-1)} + b_{hh})\]

Reference paper: Finding structure in time - Elman, 1988. https://crl.ucsd.edu/~elman/Papers/fsit.pdf

LSTM

Long Short-Term Memory - Hochreiter, 1997. http://www.bioinf.jku.at/publications/older/2604.pdf

\[\begin{split}\begin{array}{ll} i_t = \mathrm{sigmoid}(W_{ii} x_t + b_{ii} + W_{hi} h_{(t-1)} + b_{hi}) \\ f_t = \mathrm{sigmoid}(W_{if} x_t + b_{if} + W_{hf} h_{(t-1)} + b_{hf}) \\ g_t = \tanh(W_{ig} x_t + b_{ig} + W_{hc} h_{(t-1)} + b_{hg}) \\ o_t = \mathrm{sigmoid}(W_{io} x_t + b_{io} + W_{ho} h_{(t-1)} + b_{ho}) \\ c_t = f_t * c_{(t-1)} + i_t * g_t \\ h_t = o_t * \tanh(c_t) \end{array}\end{split}\]

GRU

Gated Recurrent Unit - Cho et al. 2014. http://arxiv.org/abs/1406.1078

The definition of GRU here is slightly different from paper but compatible with CUDNN.

\[\begin{split}\begin{array}{ll} r_t = \mathrm{sigmoid}(W_{ir} x_t + b_{ir} + W_{hr} h_{(t-1)} + b_{hr}) \\ z_t = \mathrm{sigmoid}(W_{iz} x_t + b_{iz} + W_{hz} h_{(t-1)} + b_{hz}) \\ n_t = \tanh(W_{in} x_t + b_{in} + r_t * (W_{hn} h_{(t-1)}+ b_{hn})) \\ h_t = (1 - z_t) * n_t + z_t * h_{(t-1)} \\ \end{array}\end{split}\]
Parameters:
  • data (NDArray) – Input data to RNN
  • parameters (NDArray) – Vector of all RNN trainable parameters concatenated
  • state (NDArray) – initial hidden state of the RNN
  • state_cell (NDArray) – initial cell state for LSTM networks (only for LSTM)
  • state_size (int (non-negative), required) – size of the state for each layer
  • num_layers (int (non-negative), required) – number of stacked layers
  • bidirectional (boolean, optional, default=0) – whether to use bidirectional recurrent layers
  • mode ({'gru', 'lstm', 'rnn_relu', 'rnn_tanh'}, required) – the type of RNN to compute
  • p (float, optional, default=0) – Dropout probability, fraction of the input that gets dropped out at training time
  • state_outputs (boolean, optional, default=0) – Whether to have the states as symbol outputs.
  • out (NDArray, optional) – The output NDArray to hold the result.
Returns:

out – The output of this function.

Return type:

NDArray or list of NDArrays