Question

我一直在研究具有2个输入的网络，以评估我的国际象棋引擎的国际象棋位置。为此，我将网络从C ++代码转换为Keras，以便能够在GPU上进行训练。

我的模型如下：

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 20480)        0                                            
__________________________________________________________________________________________________
input_2 (InputLayer)            (None, 20480)        0                                            
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 256)          5243136     input_1[0][0]                    
__________________________________________________________________________________________________
dense_2 (Dense)                 (None, 256)          5243136     input_2[0][0]                    
__________________________________________________________________________________________________
concatenate_1 (Concatenate)     (None, 512)          0           dense_1[0][0]                    
                                                                 dense_2[0][0]                    
__________________________________________________________________________________________________
dense_3 (Dense)                 (None, 32)           16416       concatenate_1[0][0]              
__________________________________________________________________________________________________
dense_4 (Dense)                 (None, 32)           1056        dense_3[0][0]                    
__________________________________________________________________________________________________
dense_5 (Dense)                 (None, 1)            33          dense_4[0][0]                    
==================================================================================================
Total params: 10,503,777
Trainable params: 10,503,777
Non-trainable params: 0

由于大量的输入和大量的训练数据（大约3亿个职位），我在训练期间使用了稀疏矩阵，效果很好。

我想将权重转移回我的手写c ++代码，并且出于调试目的，我想将单个输入馈入Keras模型以将其与我的C ++模型进行比较。

indices =[21768,21769,21770,21771,21773,21774,21775,21788,21825,21830,21890,21893,21952,21959,22019,1288,1289,1290,1291,1292,1293,1294,1295,1345,1350,1410,1413,1472,1479,1539]
eval = -0.24
x_1 = np.zeros(half_input_size)
x_2 = np.zeros(half_input_size)

for i in indices:
    if(i < half_input_size):
        x_1[i] = 1
    else:
        x_2[i-half_input_size] = 1


print(x_1.shape)
print(x_2.shape)

print(model.predict([x_1, x_2]))

两个输入的形状似乎是：

(20480,)
(20480,)

然而Keras给我以下错误：

Traceback (most recent call last):
  File "A:/OneDrive/ProgrammSpeicher/CLionProjects/Koivisto/resources/networkTrainingKeras/Train.py", line 317, in <module>
    print(model.predict([x_1, x_2]))
  File "C:\Users\finne\.conda\envs\DeepLearning\lib\site-packages\keras\engine\training.py", line 1441, in predict
    x, _, _ = self._standardize_user_data(x)
  File "C:\Users\finne\.conda\envs\DeepLearning\lib\site-packages\keras\engine\training.py", line 579, in _standardize_user_data
    exception_prefix='input')
  File "C:\Users\finne\.conda\envs\DeepLearning\lib\site-packages\keras\engine\training_utils.py", line 145, in standardize_input_data
    str(data_shape))
ValueError: Error when checking input: expected input_1 to have a shape (20480,) but got array with shape (1,)

如果有人能简短地告诉我我搞砸了，我会很高兴！

问候芬兰人

Answer 1

进行预测时，您需要添加batch_dim。

如果您的模型接受2D输入，则必须在预测中传递2D样本

您可以简单地扩展维数

model.predict([np.expand_dims(x_1,0), np.expand_dims(x_2,0)])

Answer 2

您应该在输入中使用batch维度。

x_1 = np.expand_dims(x_1, 0)
x_2 = np.expand_dims(x_1, 0)

现在，您的形状为(1, 20480)，这意味着一个示例具有20480个特征

具有2个输入的Keras模型抱怨输入形状

2 个答案: