dnn.py 文件源码

def L_op(self, inputs, outputs, output_grads):
        desc, w, x, hx = inputs[:4]
        cx = inputs[4] if len(inputs) == 5 else None
        reserve, y, hy = outputs[:3]
        _, dy, dhy = output_grads[:3]
        dcy = output_grads[3] if len(output_grads) == 4 else None
        # Since the op return two outputs which contain essentially
        # the same information, the user will most likely only use one
        # of them. This leads to the situation that the other is
        # considered "disconnected" by theano in the gradient.
        # However we know that this isn't really the case so we fix it
        # here.

        # If all the ys are disconnected, then you get a boring
        # gradient instead of an error.  But in that case you
        # shouldn't call this method anyway.
        if isinstance(dy.type, DisconnectedType):
            dy = as_gpuarray_variable(y.zeros_like(),
                                      context_name=y.type.context_name)
        if isinstance(dhy.type, DisconnectedType):
            dhy = None
        if dcy and isinstance(dcy.type, DisconnectedType):
            dcy = None
        dinputs = GpuDnnRNNGradInputs(rnn_mode=self.rnn_mode,
                                      grad_h=(dhy is not None),
                                      grad_c=(dcy is not None))(
            desc, x, y, dy, dhy, dcy, w, hx, cx, reserve, return_list=True)
        reserve2, dx, dhx = dinputs[:3]
        dw = GpuDnnRNNGradWeights()(
            desc, x, hx, y, reserve2, w)
        res = [DisconnectedType()(), dw, dx, dhx]
        if cx is not None:
            res.append(dinputs[3])  # dcx
        return res