Question

我有一个大小为（5,5）的矩阵A，也有一个大小为（5,5）的矩阵C.矩阵A和矩阵C满足以下关系C = np.dot（A，B）。

我假装矩阵B是未知的。实际上，B也是大小为（5,5）的固定矩阵。

如果我有100个A和C样本，并且我想在有A'个使用角点的情况下预测矩阵C'。以下是我的喀拉拉邦代码：

    x_train, y_train = x_data[:90], y_data[:90]     # x_data and y_data is my data . the shape of x_train and y_train :(90,25)，I have reshape the input A (100,5,5)into (100,25)
    x_test, y_test = x_data[90:], y_data[90:]       #the shape of x_test and y_test :(10,25)

    model = Sequential ()
    model.add(layers.Dense(64,activation='relu', input_dim=25)) 
    model.add(layers.Dense(64,activation='relu'))
    model.add(layers.Dense(25))
    model.compile(optimizer='rmsprop',
                  loss='mse',
                  metrics=['accuracy'])
    for step in range(1001):

    cost=model.fit(x_train,y_train,epochs=20,batch_size=32)

    if step % 100 == 0:
    print('train cost: ', cost)

输出为

Epoch 13/20
90/90 [==============================] - 0s 66us/step - loss: 6.7246 - accuracy: 1.0000
Epoch 14/20
90/90 [==============================] - 0s 55us/step - loss: 8.5773 - accuracy: 1.0000
Epoch 15/20
90/90 [==============================] - 0s 66us/step - loss: 6.8627 - accuracy: 1.0000
Epoch 16/20
90/90 [==============================] - 0s 55us/step - loss: 7.6821 - accuracy: 1.0000
Epoch 17/20
90/90 [==============================] - 0s 66us/step - loss: 13.0541 - accuracy: 1.0000
Epoch 18/20
90/90 [==============================] - 0s 44us/step - loss: 12.1049 - accuracy: 1.0000
Epoch 19/20
90/90 [==============================] - 0s 44us/step - loss: 6.2855 - accuracy: 1.0000
Epoch 20/20
90/90 [==============================] - 0s 55us/step - loss: 6.8535 - accuracy: 1.0000

没有错误，损失是6.8535。

当我输入矩阵A'并通过C''= np.dot（A'，B）将预测的C'与计算的C''进行比较时，我发现它们完全不同。

我不知道为什么以及如何减少损失。我对机器学习和keras非常陌生。我希望有一个人可以帮助我。非常感谢。

如何使用维纳斯拟合二维矩阵

0 个答案: