Method forward

python/singa/autograd.py:4838–4879

(self, x, hx, cx, w)

        self.seq_lengths = seq_lengths

    def forward(self, x, hx, cx, w):
        if training:
            if self.use_mask:
                (y, hy,
                 cy) = singa.GpuRNNForwardTrainingEx(x, hx, cx, w,
                                                     self.seq_lengths.data,
                                                     self.handle)
            else:
                (y, hy,
                 cy) = singa.GpuRNNForwardTraining(x, hx, cx, w, self.handle)
            self.inputs = {
                'x': x,
                'hx': hx,
                'cx': cx,
                'w': w,
                'y': y,
                'hy': hy,
                'cy': cy
            }
        else:
            if self.use_mask:
                (y, hy,
                 cy) = singa.GpuRNNForwardInferenceEx(x, hx, cx, w,
                                                      self.seq_lengths.data,
                                                      self.handle)
            else:
                (y, hy,
                 cy) = singa.GpuRNNForwardInference(x, hx, cx, w, self.handle)

        if self.return_sequences:
            # (seq, bs, data)
            return y
        else:
            # return last time step of y
            # (seq, bs, data)[-1] -> (bs, data)
            last_y_shape = (y.shape()[1], y.shape()[2])
            last_y = singa.Tensor(list(last_y_shape), x.device())

            src_offset = y.Size() - last_y.Size()
            # def copy_data_to_from(dst, src, size, dst_offset=0, src_offset=0):
            singa.CopyDataToFrom(last_y, y, last_y.Size(), 0, src_offset)
            return last_y

    def backward(self, grad):
        assert training is True and hasattr(