mindspore情感分析时RNN网络搭建出错

问题描述：

在参考https://gitee.com/mindspore/docs/blob/r1.1/tutorials/training/source_zh_cn/advanced_use/nlp_sentimentnet.md#%E4%BD%BF%E7%94%A8sentimentnet%E5%AE%9E%E7%8E%B0%E6%83%85%E6%84%9F%E5%88%86%E7%B1%BB进行情感分类实战训练时，想要把lstm网络改为rnn网络

【操作步骤&问题现象】

在查看mindsprore文档后发现nn.RNN和nn.LSTM的区别在于LSTM比RNN多一个参数c，把代码中的所有c删除后训练时出现函数参数个数不匹配的原因，但本人水平有限，没有找到具体哪里导致的这个原因。

【截图信息】

修改后代码如下：

# 定义需要单层LSTM小算子堆叠的设备类型。

STACK_LSTM_DEVICE = ["CPU"]# 短期内存（h）和长期内存（c）初始化为0# 定义lstm_default_state函数来初始化网络参数及网络状态。def lstm_default_state(batch_size, hidden_size, num_layers, bidirectional):

    """初始化默认输入."""

    num_directions = 2 if bidirectional else 1

    h = Tensor(np.zeros((num_layers * num_directions, batch_size, hidden_size)).astype(np.float32))

    #c = Tensor(np.zeros((num_layers * num_directions, batch_size, hidden_size)).astype(np.float32))

    return h

def stack_lstm_default_state(batch_size, hidden_size, num_layers, bidirectional):

    """init default input."""

    num_directions = 2 if bidirectional else 1



    h_list =  []

    for _ in range(num_layers):

        h_list.append(Tensor(np.zeros((num_directions, batch_size, hidden_size)).astype(np.float32)))

        #c_list.append(Tensor(np.zeros((num_directions, batch_size, hidden_size)).astype(np.float32)))

    h= tuple(h_list)

    return h

# 针对不同的场景，自定义单层LSTM小算子堆叠，来实现多层LSTM大算子功能。class StackRNN(nn.Cell):

    """

    Stack multi-layers LSTM together.

    """



    def __init__(self,

                 input_size,

                 hidden_size,

                 num_layers=3,

                 has_bias=True,

                 batch_first=False,

                 dropout=0.0,

                 bidirectional=True):

        super(StackRNN, self).__init__()

        self.num_layers = num_layers

        self.batch_first = batch_first

        self.transpose = ops.Transpose()



        # direction number

        num_directions = 2 if bidirectional else 1



        # input_size list

        input_size_list = [input_size]

        for i in range(num_layers - 1):

            input_size_list.append(hidden_size * num_directions)



        # layers

        layers = []

        for i in range(num_layers):

            layers.append(nn.RNNCell(input_size=input_size_list[i],

                                      hidden_size=hidden_size,

                                      has_bias=has_bias,

                                      

                                     

                                      ))



        # weights

        weights = []

        for i in range(num_layers):

            # weight size

            weight_size = (input_size_list[i] + hidden_size) * num_directions * hidden_size * 4

            if has_bias:

                bias_size = num_directions * hidden_size * 4

                weight_size = weight_size + bias_size



            # numpy weight

            stdv = 1 / math.sqrt(hidden_size)

            w_np = np.random.uniform(-stdv, stdv, (weight_size, 1, 1)).astype(np.float32)



            # lstm weight

            weights.append(Parameter(initializer(Tensor(w_np), w_np.shape), name="weight" + str(i)))



        #

        self.lstms = layers

        self.weight = ParameterTuple(tuple(weights))

        print(1)

        

    def construct(self, x, hx):

        """construct"""

        print(2)

        if self.batch_first:

            x = self.transpose(x, (1, 0, 2))

        # stack lstm

        h= hx

        hn= None

        for i in range(self.num_layers):

            x, hn,  _, _ = self.lstms[i](x, h[i], self.weight[i])

        if self.batch_first:

            x = self.transpose(x, (1, 0,2))

        

        return x, (hn)

其他部分未作修改

【日志信息】（可选，上传日志内容或者附件）

解答：

从报错信息来看是因为你的函数要求参数是两个但实际提供了三个，那么应该是你有地方删除c变量时没有删除干净，若是寻找比较难你可以开启pynative模式，https://www.mindspore.cn/tutorials/zh-CN/r1.7/advanced/pynative_graph/pynative.html?highlight=pynative， 参考该文档寻找哪一块没有配置好。