ResLSTM

ResLSTM(input_size => hidden_size;
    return_state=false,
    kwargs...)

Residual long short-term memory network (Kim et al., 2017). See ResLSTMCell for a layer that processes a single sequence.

Arguments

• input_size => hidden_size: input and output dimension of the layer.

Keyword arguments

• memory_size: internal memory cell dimension. Default is hidden_size.
• init_kernel: initializer for the input to hidden weights. Default is glorot_uniform.
• init_recurrent_kernel: initializer for the hidden to hidden weights. Default is glorot_uniform.
• init_peephole_kernel: initializer for the peephole weights. Default is glorot_uniform.
• init_projection_kernel: initializer for the cell output projection weights. Default is glorot_uniform.
• init_residual_kernel: initializer for the residual projection weights. Default is glorot_uniform.
• bias: include input to recurrent bias or not. Default is true.
• recurrent_bias: include recurrent to recurrent bias or not. Default is true.
• peephole_bias: include peephole to recurrent bias or not. Default is true.
• independent_recurrence: flag to toggle independent recurrence. If true, the recurrent to recurrent weights are a vector instead of a matrix. Default false.
• integration_mode: determines how the input and hidden projections are combined. The options are :addition and :multiplicative_integration. Defaults to :addition.
• return_state: Option to return the last state together with the output. Default is false.

Equations

\[\begin{aligned} \mathbf{i}(t) &= \sigma\left( \mathbf{W}^{i}_{ih} \mathbf{x}(t) + \mathbf{W}^{i}_{hh} \mathbf{h}(t-1) + \mathbf{p}^{i} \odot \mathbf{c}(t-1) + \mathbf{b}^{i} \right), \\ \mathbf{f}(t) &= \sigma\left( \mathbf{W}^{f}_{ih} \mathbf{x}(t) + \mathbf{W}^{f}_{hh} \mathbf{h}(t-1) + \mathbf{p}^{f} \odot \mathbf{c}(t-1) + \mathbf{b}^{f} \right), \\ \mathbf{g}(t) &= \tanh\left( \mathbf{W}^{g}_{ih} \mathbf{x}(t) + \mathbf{W}^{g}_{hh} \mathbf{h}(t-1) + \mathbf{b}^{g} \right), \\ \mathbf{c}(t) &= \mathbf{f}(t) \odot \mathbf{c}(t-1) + \mathbf{i}(t) \odot \mathbf{g}(t), \\ \mathbf{o}(t) &= \sigma\left( \mathbf{W}^{o}_{ih} \mathbf{x}(t) + \mathbf{W}^{o}_{hh} \mathbf{h}(t-1) + \mathbf{W}^{o}_{ch} \mathbf{c}(t) + \mathbf{b}^{o} \right), \\ \mathbf{m}(t) &= \mathbf{W}_{p} \tanh\left( \mathbf{c}(t) \right), \\ \mathbf{h}(t) &= \mathbf{o}(t) \odot \left( \mathbf{m}(t) + \mathbf{W}_{r} \mathbf{x}(t) \right) \end{aligned}\]

Forward

reslstm(inp, (state, cstate))
reslstm(inp)

Arguments

• inp: The input to the reslstm. It should be a vector of size input_size x len or a matrix of size input_size x len x batch_size.
• (state, cstate): A tuple containing the hidden and cell states of the ResLSTM. state should be a vector of size hidden_size or a matrix of size hidden_size x batch_size; cstate should be a vector of size memory_size or a matrix of size memory_size x batch_size. If not provided, they are assumed to be vectors of zeros, initialized by Flux.initialstates.

Returns

• New hidden states new_states as an array of size hidden_size x len x batch_size. When return_state = true it returns a tuple of the hidden stats new_states and the last state of the iteration.