Question

经过几次不成功的尝试，我想请求你帮忙解决这个错误。我正在尝试通过使用本地csv文件来训练深度自动编码器网络，然后将其转换（通过csv和numpy库）到numpy数组中。但是这些数据永远不会影响我占位符的张量。

这是深度自动编码器的摘要：

class Deep_Autoencoder:
    def __init__(self, input_dim, n_nodes_hl = (32, 16, 1), epochs = 400, batch_size = 128, learning_rate = 0.02, n_examples = 10):

    # Hyperparameters
    self.input_dim = input_dim
    self.epochs = epochs
    self.batch_size = batch_size
    self.learning_rate = learning_rate
    self.n_examples = n_examples

    # Input and target placeholders
    X = tf.placeholder('float', [None, self.input_dim])
    Y = tf.placeholder('float', [None, self.input_dim])
    ...

    self.X = X
    print("self.X : ", self.X)
    self.Y = Y
    print("self.Y : ", self.Y)
    ...

def train_neural_network(self, data, targets):

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        for epoch in range(self.epochs):
            epoch_loss = 0
            i = 0

            # Let's train it in batch-mode
            while i < len(data):
                start = i
                end = i + self.batch_size

                batch_x = np.array(data[start:end])
                print("type batch_x :", type(batch_x))
                print("len batch_x :", len(batch_x))
                batch_y = np.array(targets[start:end])
                print("type batch_y :", type(batch_y))
                print("len batch_y :", len(batch_y))

                hidden, _, c = sess.run([self.encoded, self.optimizer, self.cost], feed_dict={self.X: batch_x, self.Y: batch_y})
                epoch_loss +=c
                i += self.batch_size

        self.saver.save(sess, 'selfautoencoder.ckpt')
        print('Accuracy', self.accuracy.eval({self.X: data, self.Y: targets}))

在这里我创建了输入数据，在下面你可以看到我将打印出他们的主要功能以供参考（注意我实际上只对第3列感兴趣）：

features_DeepAE = create_feature_sets(filename)

Train_x = np.array(features_DeepAE[0])
Train_y = np.array(features_DeepAE[1])

print("type Train_x : ", type(Train_x))
print("type Train_x.T[3] : ", type(Train_x.T[3]))
print("len Train_x : ", len(Train_x))
print("len Train_x.T[3] : ", len(Train_x.T[3]))
print("shape Train_x : ", Train_x.shape)
print("type Train_y : ", type(Train_y))
print("type Train_y.T[3] : ", type(Train_y.T[3]))
print("len Train_y : ", len(Train_y))
print("len Train_y.T[3] : ", len(Train_y.T[3]))
print("shape Train_y : ", Train_y.shape)

在这里我运行代码：

DAE = Deep_Autoencoder(input_dim = len(Train_x))
DAE.train_neural_network(Train_x.T[3], Train_y.T[3])

这些是打印输出，fyi：

type Train_x :  <class 'numpy.ndarray'>
type Train_x.T[3] :  <class 'numpy.ndarray'>
len Train_x :  3433
len Train_x.T[3] :  3433
shape Train_x :  (3433, 5)
type Train_y :  <class 'numpy.ndarray'>
type Train_y.T[3] :  <class 'numpy.ndarray'>
len Train_y :  3433
len Train_y.T[3] :  3433
shape Train_y :  (3433, 5)
self.X :  Tensor("Placeholder_142:0", shape=(?, 3433), dtype=float32)
self.Y :  Tensor("Placeholder_143:0", shape=(?, 3433), dtype=float32)
type batch_x : <class 'numpy.ndarray'>
len batch_x : 128
type batch_y : <class 'numpy.ndarray'>
len batch_y : 128

最后是错误：

ValueError：无法为Tensor'Plamber_142：0'提供形状值（128，），其形状为'（？，3433）'

并且是的...我在占位符＃143 ...测量了很多失败（重塑批次和/或张量，转换一个和/或另一个，在互联网上寻找变通方法......）！如果需要，请随时提出更多信息。

Answer 1

解决了，感谢Avishkar Bhoopchand和amirbar：

将input_dim设置为1并为batch_x和batch_y添加“虚拟”维度，使其适合形状[？，1]的占位符，如下所示：batch_x = np.array（data [start：end]） [：，None]和batch_y = np.array（targets [start：end]）[：，None]。 None为Numpy数组添加空维度。

Tensorflow - ValueError：无法为Tensor'Plamber_142：0'提供形状值（128，），其形状为'（？，3433）

1 个答案: