brainpy.dyn.layers.LSTM#

class brainpy.dyn.layers.LSTM(num_in, num_out, Wi_initializer=XavierNormal(scale=1.0, mode=fan_avg, in_axis=- 2, out_axis=- 1, distribution=truncated_normal, seed=3417531), Wh_initializer=XavierNormal(scale=1.0, mode=fan_avg, in_axis=- 2, out_axis=- 1, distribution=truncated_normal, seed=3917797), b_initializer=ZeroInit, state_initializer=ZeroInit, activation='tanh', mode=TrainingMode, train_state=False, name=None)[source]#

Long short-term memory (LSTM) RNN core.

The implementation is based on (zaremba, et al., 2014) 1. Given \(x_t\) and the previous state \((h_{t-1}, c_{t-1})\) the core computes

\[\begin{split}\begin{array}{ll} i_t = \sigma(W_{ii} x_t + W_{hi} h_{t-1} + b_i) \\ f_t = \sigma(W_{if} x_t + W_{hf} h_{t-1} + b_f) \\ g_t = \tanh(W_{ig} x_t + W_{hg} h_{t-1} + b_g) \\ o_t = \sigma(W_{io} x_t + W_{ho} h_{t-1} + b_o) \\ c_t = f_t c_{t-1} + i_t g_t \\ h_t = o_t \tanh(c_t) \end{array}\end{split}\]

where \(i_t\), \(f_t\), \(o_t\) are input, forget and output gate activations, and \(g_t\) is a vector of cell updates.

The output is equal to the new hidden, \(h_t\).

Notes

Forget gate initialization: Following (Jozefowicz, et al., 2015) 2 we add 1.0 to \(b_f\) after initialization in order to reduce the scale of forgetting in the beginning of the training.

Parameters
  • num_out (int) – The number of hidden unit in the node.

  • state_initializer (callable, Initializer, bm.ndarray, jax.numpy.ndarray) – The state initializer.

  • Wi_initializer (callable, Initializer, bm.ndarray, jax.numpy.ndarray) – The input weight initializer.

  • Wh_initializer (callable, Initializer, bm.ndarray, jax.numpy.ndarray) – The hidden weight initializer.

  • b_initializer (optional, callable, Initializer, bm.ndarray, jax.numpy.ndarray) – The bias weight initializer.

  • activation (str, callable) – The activation function. It can be a string or a callable function. See brainpy.math.activations for more details.

  • trainable (bool) – Whether set the node is trainable.

References

1

Zaremba, Wojciech, Ilya Sutskever, and Oriol Vinyals. “Recurrent neural network regularization.” arXiv preprint arXiv:1409.2329 (2014).

2

Jozefowicz, Rafal, Wojciech Zaremba, and Ilya Sutskever. “An empirical exploration of recurrent network architectures.” In International conference on machine learning, pp. 2342-2350. PMLR, 2015.

__init__(num_in, num_out, Wi_initializer=XavierNormal(scale=1.0, mode=fan_avg, in_axis=- 2, out_axis=- 1, distribution=truncated_normal, seed=3417531), Wh_initializer=XavierNormal(scale=1.0, mode=fan_avg, in_axis=- 2, out_axis=- 1, distribution=truncated_normal, seed=3917797), b_initializer=ZeroInit, state_initializer=ZeroInit, activation='tanh', mode=TrainingMode, train_state=False, name=None)[source]#

Methods

__init__(num_in, num_out[, Wi_initializer, ...])

clear_input()

get_delay_data(identifier, delay_step, *indices)

Get delay data according to the provided delay steps.

load_states(filename[, verbose])

Load the model states.

nodes([method, level, include_self])

Collect all children nodes.

offline_fit(target, fit_record)

offline_init()

online_fit(target, fit_record)

online_init()

register_delay(identifier, delay_step, ...)

Register delay variable.

register_implicit_nodes(*nodes, **named_nodes)

register_implicit_vars(*variables, ...)

reset([batch_size])

Reset function which reset the whole variables in the model.

reset_local_delays([nodes])

Reset local delay variables.

reset_state([batch_size])

Reset function which reset the states in the model.

save_states(filename[, variables])

Save the model states.

train_vars([method, level, include_self])

The shortcut for retrieving all trainable variables.

unique_name([name, type_])

Get the unique name for this object.

update(sha, x)

The function to specify the updating rule.

update_local_delays([nodes])

Update local delay variables.

vars([method, level, include_self])

Collect all variables in this node and the children nodes.

Attributes

c

Memory cell.

global_delay_data

h

Hidden state.

mode

Mode of the model, which is useful to control the multiple behaviors of the model.

name

Name of the model.