Struct LSTMCell 

pub struct LSTMCell {
    pub weight_ih: Tensor,
    pub weight_hh: Tensor,
    pub bias_ih: Tensor,
    pub bias_hh: Tensor,
    /* private fields */
}

LSTM cell with PyTorch’s [i, f, g, o] gate order.

weight_ih shape: [4*hidden, input]; weight_hh shape: [4*hidden, hidden]. bias_ih and bias_hh both [4*hidden] — summed in Self::step to match nn.LSTM’s packing, so PyTorch checkpoints load without remapping.

Not a Layer — cells take (x, h, c), not a single tensor.

Fields

weight_ih: Tensor

weight_hh: Tensor

bias_ih: Tensor

bias_hh: Tensor

Implementations

impl LSTMCell

pub fn new( weight_ih: Tensor, weight_hh: Tensor, bias_ih: Tensor, bias_hh: Tensor, ) -> Self

Create an LSTM cell from existing weight/bias tensors. hidden_size is derived from weight_ih.shape()[0] / 4.

pub fn with_dims(input_size: usize, hidden_size: usize, dtype: DType) -> Self

Create an LSTM cell with deterministic sin() initialization, zero biases.

pub fn hidden_size(&self) -> usize

pub fn step( &self, x: &Tensor, h: &Tensor, c: &Tensor, ) -> Result<(Tensor, Tensor)>

One LSTM step. Returns (h_next, c_next).

Shapes: x: [B, input], h, c: [B, hidden].

Trait Implementations

impl Clone for LSTMCell

fn clone(&self) -> LSTMCell

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.