时间戳

首页 > 解决方案 > Tensorflow 输入维度不匹配问题

python - Tensorflow 输入维度不匹配问题

问题描述

我创建了一个 LSTM-NN。我正在传递一个输入，但我得到了错误：
ValueError: Input 0 of layer lstm is incompatible with the layer: expected ndim=3, found ndim=1. Full shape received: (7,)

为了解决这个问题，我参考了一个堆栈溢出帖子，其中提到了参数的使用：input_shape。由于缺乏理解，我仍然无法解决我的问题。请帮忙。这是我的代码

# This is the definition of the model
class LSTMmodel(tf.Module):
    def __init__(self, arg_name=None):
        super().__init__(name=arg_name)
        self.__input = tf.Variable(initial_value=[0 for x in range(7)])
        self.__network = tf.keras.layers.LSTM(units=7, input_shape=(7,))
        self.__output = tf.Variable(initial_value=[0 for x in range(7)])
    @tf.function
    def train(self, arg_data_train, labels, learning_rate):
        with tf.GradientTape() as t:
            self.__input = tf.Variable(initial_value=[0 for x in range(7)])
            self.__output = self.__network(self.__input)
            loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=None, logits=None)
            dw, db = t.gradient(loss, [self.w, self.b])
        self.w.assign_sub(learning_rate * dw)
        self.b.assign_sub(learning_rate * db)

    @tf.function
    def __call__(self, arg_input=[0 for x in range(7)]):
        self.__input = tf.Variable(arg_input)
        self.__output = self.__network(self.__input)
        return self.__output

# This is the input I provide for training where the problem occurs.
# The two vars ````cgm```` and ````labels```` are length 9222 lists.
# Each element of the list is a list with length 7 filled with only integers.
modela = LSTMmodel(arg_name='namea')
modela.train(cgm ,labels, 0.4)
```

标签： pythontensorflow

解决方案

LSTM 将 3D 张量作为输入，并且您正在传递 1D 张量，因此您需要将其重塑为适当的形状。

self.__input = tf.Variable(initial_value=[0 for x in range(7)])
self.__input_reshaped = tf.reshape(self.__input, [1, 7, 1]) # shape of (1, 7, 1)

您还需要更改input_shapeLSTM 的。

self.__network = tf.keras.layers.LSTM(units=7, input_shape=(7,1))

然后将重构后的输入传递给网络以获得输出。

self.__output = self.__network(self.__input_reshaped)

推荐阅读