我正在嘗試使用 keras 為我的簡化神經網絡實現 CTC 損失：  def ctc_lambda_func(args):    y_pred, y_train, input_length, label_length = args     return K.ctc_batch_cost(y_train, y_pred, input_length, label_length)x_train = x_train.reshape(x_train.shape[0],20, 10).astype('float32')input_data = layers.Input(shape=(20,10,))x=layers.Convolution1D(filters=256, kernel_size=3,  padding="same", strides=1, use_bias=False ,activation= 'relu')(input_data)x=layers.BatchNormalization()(x)x=layers.Dropout(0.2)(x)x=layers.Bidirectional (LSTM(units=200 , return_sequences=True)) (x)x=layers.BatchNormalization()(x)x=layers.Dropout(0.2)(x)y_pred=outputs = layers.Dense(5, activation='softmax')(x)fun = Model(input_data, y_pred)# fun.summary()label_length=np.zeros((3800,1))input_length=np.zeros((3800,1))for i in range (3799):    label_length[i,0]=4    input_length[i,0]=5   y_train = np.array(y_train)x_train = np.array(x_train)input_length = np.array(input_length)label_length = np.array(label_length)   loss_out = Lambda(ctc_lambda_func, output_shape=(1,), name='ctc')([y_pred, y_train, input_length, label_length])model =keras.models.Model(inputs=[input_data, y_train, input_length, label_length], outputs=loss_out)model.compile(loss={'ctc': lambda y_train, y_pred: y_pred}, optimizer = 'adam')model.fit(x=[x_train, y_train, input_length, label_length],  epochs=10, batch_size=100)我們有 y_true （或 y_train） ，尺寸為（3800,4），因此我將 label_length=4 和 input_length=5 （+1 表示空白）我面臨這個錯誤：ValueError: Input tensors to a Model must come from `tf.keras.Input`. Received: [[0. 1. 0. 0.] [0. 1. 0. 0.] [0. 1. 0. 0.] ... [1. 0. 0. 0.] [1. 0. 0. 0.] [1. 0. 0. 0.]] (missing previous layer metadata).y_true 是這樣的： [[0. 1. 0. 0.] [0. 1. 0. 0.] ... [1. 0. 0. 0.] [1. 0. 0. 0.] [1. 0. 0. 0.]]我的問題是什么？