An LSTM implementation that supports recurrent dropout. More...

Inheritance diagram for deeplima::nets::torch_modules::LSTMwRecDropoutImpl:

Public Member Functions
	LSTMwRecDropoutImpl ()=default

	LSTMwRecDropoutImpl (int64_t input_size, int64_t hidden_size, int64_t num_layers, bool bias=true, bool batch_first=false, float dropout=0, bool bidirectional=false, bool pad=false, float rec_dropout=0)

std::pair< torch::Tensor, std::pair< torch::Tensor, torch::Tensor > >	rnn_loop (torch::Tensor x, torch::Tensor batch_sizes, torch::nn::LSTMCellImpl *cell, std::vector< torch::Tensor > inits, bool reverse=false)
	RNN loop for one layer in one direction with recurrent dropout Assumes input is PackedSequence, returns PackedSequence as well.

std::tuple< torch::nn::utils::rnn::PackedSequence, std::tuple< torch::Tensor, torch::Tensor > >	forward (torch::nn::utils::rnn::PackedSequence input, torch::optional< std::tuple< torch::Tensor, torch::Tensor > > hx={})

Detailed Description

An LSTM implementation that supports recurrent dropout.

Constructor & Destructor Documentation

deeplima::nets::torch_modules::LSTMwRecDropoutImpl::LSTMwRecDropoutImpl ( )

default

deeplima::nets::torch_modules::LSTMwRecDropoutImpl::LSTMwRecDropoutImpl	(	int64_t	input_size,
		int64_t	hidden_size,
		int64_t	num_layers,
		bool	bias = `true`,
		bool	batch_first = `false`,
		float	dropout = `0`,
		bool	bidirectional = `false`,
		bool	pad = `false`,
		float	rec_dropout = `0`
	)

inline

std::tuple< torch::nn::utils::rnn::PackedSequence, std::tuple< torch::Tensor, torch::Tensor > > deeplima::nets::torch_modules::LSTMwRecDropoutImpl::forward	(	torch::nn::utils::rnn::PackedSequence	input,
		torch::optional< std::tuple< torch::Tensor, torch::Tensor > >	hx = `{}`
	)

inline

std::pair< torch::Tensor, std::pair< torch::Tensor, torch::Tensor > > deeplima::nets::torch_modules::LSTMwRecDropoutImpl::rnn_loop	(	torch::Tensor	x,
		torch::Tensor	batch_sizes,
		torch::nn::LSTMCellImpl *	cell,
		std::vector< torch::Tensor >	inits,
		bool	reverse = `false`
	)