Function cudnn_LSTM

caffe2/python/rnn_cell.py:1706–1837

CuDNN version of LSTM for GPUs. input_blob Blob containing the input. Will need to be available when param_init_net is run, because the sequence lengths and batch sizes will be inferred from the size of this bl

(model, input_blob, initial_states, dim_in, dim_out,
               scope, recurrent_params=None, input_params=None,
               num_layers=1, return_params=False)

def cudnn_LSTM(model, input_blob, initial_states, dim_in, dim_out,
               scope, recurrent_params=None, input_params=None,
               num_layers=1, return_params=False):
    '''
    CuDNN version of LSTM for GPUs.
    input_blob          Blob containing the input. Will need to be available
                        when param_init_net is run, because the sequence lengths
                        and batch sizes will be inferred from the size of this
                        blob.
    initial_states      tuple of (hidden_init, cell_init) blobs
    dim_in              input dimensions
    dim_out             output/hidden dimension
    scope               namescope to apply
    recurrent_params    dict of blobs containing values for recurrent
                        gate weights, biases (if None, use random init values)
                        See GetLSTMParamNames() for format.
    input_params        dict of blobs containing values for input
                        gate weights, biases (if None, use random init values)
                        See GetLSTMParamNames() for format.
    num_layers          number of LSTM layers
    return_params       if True, returns (param_extract_net, param_mapping)
                        where param_extract_net is a net that when run, will
                        populate the blobs specified in param_mapping with the
                        current gate weights and biases (input/recurrent).
                        Useful for assigning the values back to non-cuDNN
                        LSTM.
    '''
    with core.NameScope(scope):
        weight_params = GetLSTMParamNames()['weights']
        bias_params = GetLSTMParamNames()['biases']

        input_weight_size = dim_out * dim_in
        upper_layer_input_weight_size = dim_out * dim_out
        recurrent_weight_size = dim_out * dim_out
        input_bias_size = dim_out
        recurrent_bias_size = dim_out

        def init(layer, pname, input_type):
            input_weight_size_for_layer = input_weight_size if layer == 0 else \
                upper_layer_input_weight_size
            if pname in weight_params:
                sz = input_weight_size_for_layer if input_type == 'input' \
                    else recurrent_weight_size
            elif pname in bias_params:
                sz = input_bias_size if input_type == 'input' \
                    else recurrent_bias_size
            else:
                assert False, "unknown parameter type {}".format(pname)
            return model.param_init_net.UniformFill(
                [],
                "lstm_init_{}_{}_{}".format(input_type, pname, layer),
                shape=[sz])

        # Multiply by 4 since we have 4 gates per LSTM unit
        first_layer_sz = input_weight_size + recurrent_weight_size + \
                         input_bias_size + recurrent_bias_size
        upper_layer_sz = upper_layer_input_weight_size + \
                         recurrent_weight_size + input_bias_size + \