Skip to content

Latest commit

 

History

History
85 lines (57 loc) · 2.99 KB

README.md

File metadata and controls

85 lines (57 loc) · 2.99 KB
Raw

Implementation of recurrent LSTM neural networks

This is a very single-purpose library that implements AVX-optimized LSTM neural networks for C for Unix platforms. Works at least on Linux.

The neural network architecture currently supported by this library is:

  • An input layer.

  • At least one hidden layer; each hidden layer must contain a multiple of 4 LSTM nodes.

  • An output layer.

All layers are fully connected to their respective next layer. Additionally, the last hidden layer is fully connected to first hidden layer. If there is only one hidden layer (this is common architecture), then that layer is connected to itself.

LSTM nodes have forget gate, output gate and input gate and peepholes from memory cells to all of these. All gates also have bias. The trainable parameters are weights (including all weights to gates and feedback gates from last layer to first), peepholes and biases.

I'm using this library with non-gradient based optimization which is why this library does not compute gradients.

License

License is LGPL 2.1.

Compile

You can use the Makefile:

$ make

This should result in a liblstm.so file.

You can use make install but the Makefile is hard-coded to install to /usr/local. This library is just 3 files so you can also just include the files in your project or compile by hand.

API

API is simplistic. You might see the type real in these functions, at the moment, real is always double.

lstm* allocate_lstm( size_t num_inputs, size_t num_outputs, const size_t* num_hiddens, size_t num_hidden_layers );
void free_lstm( lstm* lstm );

allocate_lstm allocates an LSTM network. Returns NULL if out of memory or configuration is not supported. There must be at least one hidden layer and each hidden layer must have a multiple of 4 neurons. Use free_lstm to free up your resources.

void propagate(lstm* restrict lstm, const real* restrict input, real* restrict output);

propagate runs one iteration of the network. Inputs are taken from input vector and outputs are written to your output. Feedback activations are saved so your network has state.

void reset_lstm( lstm* lstm );

reset_lstm resets all memory cells and feedback activations.

size_t serialized_lstm_size( const lstm* lstm );
void serialize_lstm( const lstm* lstm, real* serialized );
void unserialize_lstm( lstm* lstm, const real* serialized );

These functions are used to serialize and unserialize parameters from an lstm. serialized_lstm_size returns the number of parameters and this is the number of reals you need for serialize_lstm and unserialize_lstm.

serialize_lstm takes an lstm and writes all parameters to one-dimensional array. unserialize_lstm does the reverse and updates parameters in lstm according to given one-dimensional array.

lstm networks created with the same size parameters in allocate_lstm have identical size and you can serialize and unserialize between them.