Recurrent Models

RNN and LSTM sequence models with a linear projection at every time step.

structure TorchLean.nn.models.RecurrentConfig : Type

Configuration for an RNN or LSTM followed by a time-distributed linear head.

Shapes follow the convention used by the runnable examples:

• input: (seqLen × inputSize)
• output: (seqLen × outputSize)

• seqLen : ℕ

  Number of time steps.

• inputSize : ℕ

  Number of features presented at each time step.

• hiddenSize : ℕ

  Width of the recurrent state.

• outputSize : ℕ

  Number of features produced at each time step.

@[implicit_reducible]

instance TorchLean.nn.models.instReprRecurrentConfig : Repr RecurrentConfig

def TorchLean.nn.models.instReprRecurrentConfig.repr : RecurrentConfig → ℕ → Std.Format

@[reducible, inline]

abbrev TorchLean.nn.models.recurrentInShape (cfg : RecurrentConfig) : Spec.Shape

Input shape (seqLen × inputSize).

@[reducible, inline]

abbrev TorchLean.nn.models.recurrentOutShape (cfg : RecurrentConfig) : Spec.Shape

Output shape (seqLen × outputSize).

def TorchLean.nn.models.rnnWithLinearHead (cfg : RecurrentConfig) : M (Sequential (recurrentInShape cfg) (recurrentOutShape cfg))

Vanilla RNN core plus time-distributed linear head:

rnn(seqLen, inputSize, hiddenSize) → linear(hiddenSize, outputSize).

def TorchLean.nn.models.lstmWithLinearHead (cfg : RecurrentConfig) : M (Sequential (recurrentInShape cfg) (recurrentOutShape cfg))

LSTM core plus time-distributed linear head:

lstm(seqLen, inputSize, hiddenSize) → linear(hiddenSize, outputSize).