Python 这两个Keras模型是等价的吗？如果我在输出层进行压缩，然后提取张量切片

我试图创建一个模型，其最后一层有多个输出。您可以看到

3

输出磁头的模型图

Model: "model_18"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
observations (InputLayer)       [(None, 84, 84, 4)]  0                                            
__________________________________________________________________________________________________
layer1 (Conv2D)                 (None, 20, 20, 32)   8224        observations[0][0]               
__________________________________________________________________________________________________
layer2 (Conv2D)                 (None, 9, 9, 64)     32832       layer1[0][0]                     
__________________________________________________________________________________________________
layer3 (Conv2D)                 (None, 7, 7, 64)     36928       layer2[0][0]                     
__________________________________________________________________________________________________
layer4 (Flatten)                (None, 3136)         0           layer3[0][0]                     
__________________________________________________________________________________________________
agent_indicator (InputLayer)    [(None, 2)]          0                                            
__________________________________________________________________________________________________
concatenate_6 (Concatenate)     (None, 3138)         0           layer4[0][0]                     
                                                                 agent_indicator[0][0]            
__________________________________________________________________________________________________
layer5 (Dense)                  (None, 512)          1607168     concatenate_6[0][0]              
__________________________________________________________________________________________________
gamma_q_layer0 (Dense)          (None, 6)            3078        layer5[0][0]                     
__________________________________________________________________________________________________
gamma_q_layer1 (Dense)          (None, 6)            3078        layer5[0][0]                     
__________________________________________________________________________________________________
gamma_q_layer2 (Dense)          (None, 6)            3078        layer5[0][0]                     
==================================================================================================
Total params: 1,694,386
Trainable params: 1,694,386
Non-trainable params: 0
__________________________________________________________________________________________________
None

由于这是一个我正在与库接口的自定义模型，因此它不起作用，因为库通常需要一个具有输入形状

（batch_size，6）

的单头

如果我将输出磁头连接到一个层中，如图所示，然后将其输入到库中，库中仍然需要一个张量，但我只需将输入形状修改为

（batch_size，6*3）

，然后将张量切成各个部分以提取各个磁头，分别计算每个磁头的损耗，这两种方法是否等效

Model: "model_33"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
observations (InputLayer)       [(None, 84, 84, 4)]  0                                            
__________________________________________________________________________________________________
layer1 (Conv2D)                 (None, 20, 20, 32)   8224        observations[0][0]               
__________________________________________________________________________________________________
layer2 (Conv2D)                 (None, 9, 9, 64)     32832       layer1[0][0]                     
__________________________________________________________________________________________________
layer3 (Conv2D)                 (None, 7, 7, 64)     36928       layer2[0][0]                     
__________________________________________________________________________________________________
layer4 (Flatten)                (None, 3136)         0           layer3[0][0]                     
__________________________________________________________________________________________________
agent_indicator (InputLayer)    [(None, 2)]          0                                            
__________________________________________________________________________________________________
concatenate_18 (Concatenate)    (None, 3138)         0           layer4[0][0]                     
                                                                 agent_indicator[0][0]            
__________________________________________________________________________________________________
layer5 (Dense)                  (None, 512)          1607168     concatenate_18[0][0]             
__________________________________________________________________________________________________
gamma_q_layer0 (Dense)          (None, 6)            3078        layer5[0][0]                     
__________________________________________________________________________________________________
gamma_q_layer1 (Dense)          (None, 6)            3078        layer5[0][0]                     
__________________________________________________________________________________________________
gamma_q_layer2 (Dense)          (None, 6)            3078        layer5[0][0]                     
__________________________________________________________________________________________________
concatenate_19 (Concatenate)    (None, 18)           0           gamma_q_layer0[0][0]             
                                                                 gamma_q_layer1[0][0]             
                                                                 gamma_q_layer2[0][0]             
==================================================================================================
Total params: 1,694,386
Trainable params: 1,694,386
Non-trainable params: 0
__________________________________________________________________________________________________
None

我不确定的是，在这两款车型的backprop期间，重量将如何调整。在第一个模型中，计算每个水头的损失，并将其平均化为最终损失。然后backprop调整权重。如果我像第二个模型一样，通过张量切片提取磁头，计算每个磁头的损耗，并将其平均化为最终损耗，这两个损耗值是否相同

多谢各位