Function _LSTM

caffe2/python/rnn_cell.py:1467–1576

Adds a standard LSTM recurrent network operator to a model. cell_class: LSTMCell or compatible subclass model: ModelHelper object new operators would be added to input_blob: the input sequence in a format T x N x D where T is sequence size, N - batch size and D - inpu

(
    cell_class,
    model,
    input_blob,
    seq_lengths,
    initial_states,
    dim_in,
    dim_out,
    scope=None,
    outputs_with_grads=(0,),
    return_params=False,
    memory_optimization=False,
    forget_bias=0.0,
    forward_only=False,
    drop_states=False,
    return_last_layer_only=True,
    static_rnn_unroll_size=None,
    **cell_kwargs
)

def _LSTM(
    cell_class,
    model,
    input_blob,
    seq_lengths,
    initial_states,
    dim_in,
    dim_out,
    scope=None,
    outputs_with_grads=(0,),
    return_params=False,
    memory_optimization=False,
    forget_bias=0.0,
    forward_only=False,
    drop_states=False,
    return_last_layer_only=True,
    static_rnn_unroll_size=None,
    **cell_kwargs
):
    '''
    Adds a standard LSTM recurrent network operator to a model.

    cell_class: LSTMCell or compatible subclass

    model: ModelHelper object new operators would be added to

    input_blob: the input sequence in a format T x N x D
            where T is sequence size, N - batch size and D - input dimension

    seq_lengths: blob containing sequence lengths which would be passed to
            LSTMUnit operator

    initial_states: a list of (2 * num_layers) blobs representing the initial
            hidden and cell states of each layer. If this argument is None,
            these states will be added to the model as network parameters.

    dim_in: input dimension

    dim_out: number of units per LSTM layer
            (use int for single-layer LSTM, list of ints for multi-layer)

    outputs_with_grads : position indices of output blobs for LAST LAYER which
            will receive external error gradient during backpropagation.
            These outputs are: (h_all, h_last, c_all, c_last)

    return_params: if True, will return a dictionary of parameters of the LSTM

    memory_optimization: if enabled, the LSTM step is recomputed on backward
            step so that we don't need to store forward activations for each
            timestep. Saves memory with cost of computation.

    forget_bias: forget gate bias (default 0.0)

    forward_only: whether to create a backward pass

    drop_states: drop invalid states, passed through to LSTMUnit operator

    return_last_layer_only: only return outputs from final layer