Method _apply

caffe2/python/rnn_cell.py:605–700

(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    )

class MILSTMCell(LSTMCell):

    def _apply(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    ):
        hidden_t_prev, cell_t_prev = states

        fc_input = hidden_t_prev
        fc_input_dim = self.hidden_size

        if extra_inputs is not None:
            extra_input_blobs, extra_input_sizes = zip(*extra_inputs)
            fc_input = brew.concat(
                model,
                [hidden_t_prev] + list(extra_input_blobs),
                self.scope('gates_concatenated_input_t'),
                axis=2,
            )
            fc_input_dim += sum(extra_input_sizes)

        prev_t = brew.fc(
            model,
            fc_input,
            self.scope('prev_t'),
            dim_in=fc_input_dim,
            dim_out=self.gates_size,
            axis=2,
        )

        # defining initializers for MI parameters
        alpha = model.create_param(
            self.scope('alpha'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=1.0),
        )
        beta_h = model.create_param(
            self.scope('beta1'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=1.0),
        )
        beta_i = model.create_param(
            self.scope('beta2'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=1.0),
        )
        b = model.create_param(
            self.scope('b'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=0.0),
        )

        # alpha * input_t + beta_h
        # Shape: [1, batch_size, 4 * hidden_size]
        alpha_by_input_t_plus_beta_h = model.net.ElementwiseLinear(