Question

我有一个Keras模型，该模型由其他3个Keras模型（嵌套模型）组成。我的问题是关于Keras训练日志中显示的损耗值的含义。

以下是我的全局模型的摘要：

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_16 (InputLayer)           (None, 256, 256, 3)  0                                            
__________________________________________________________________________________________________
model_1 (Model)                 (None, 16, 16, 128)  690368      input_16[0][0]                   
__________________________________________________________________________________________________
model_4 (Model)                 [(None, 17, 4), (None, 17, 4), (None, 16, 16, 128)] 5103826     input_16[0][0]                   
__________________________________________________________________________________________________
concatenate_8 (Concatenate)     (None, 16, 16, 256)  0           model_1[1][0]                    
                                                                 model_4[1][2]                    
__________________________________________________________________________________________________
decoder (Model)                 (None, 256, 256, 3)  582843      concatenate_8[0][0]              
==================================================================================================

这些嵌套模型是2个编码器（model_1和model_4）和1个解码器（decoder）。

我也有3种损失：2种损失直接应用于model_4输出中的2种，一种损失应用于解码器的输出。

训练完整模型时，我只看到model_4有一个损失，称为model_4_loss：

Epoch 34/60
13548/19512 [===================>..........] - ETA: 34:57 - loss: 0.6764 - decoder_loss: 0.0944 - model_4_loss: 0.2797

但是当我尝试单独训练model_4时，我在训练日志中清楚地看到了2个损失（此处concatenate_xxx的损失对应于model_4的前两个输出）：

Epoch 35/60
 5430/19512 [=======>......................] - ETA: 1:20:14 - loss: 0.8475 - concatenate_5_loss: 0.2998 - concatenate_7_loss: 0.2767

对此我有几个问题：

训练完整模型时，我不应该看到3个损失而不是2个损失（model_4 2个损失，decoder 1个损失吗？
model_4_loss代表什么？ model_4造成的2次损失的平均值？总和？两者中只有一个？
如何使训练日志清楚地显示model_4的两个损失而不是一些合计值？

为提供更多背景信息，以下是我如何构建整个模型的摘要：

encoder1 = build_encoder1()   # returns an object of type `Model` with a single (None, 16, 16, 128) output
encoder2 = build_encoder2()   # returns an object of type `Model` with a list of 3 tensors as output
decoder = build_decoder()     # returns a `Model` with a single (None, 256, 256, 3) output

inp = Input(shape=input_shape)      # input_shape is (None, 256, 256, 3)
z_1 = encoder1(inp)                 # (None, 16, 16, 128)
out1, out2, z_2 = encoder2(inp)     # [(None, 17, 4), (None, 17, 4), (None, 16, 16, 128)]

concat = concatenate[z_1, z_2]      # (None, 16, 16, 256)
out3 = decoder(concat)              # (None, 256, 256, 3)

outputs = [out3, out1, out2]
losses = [loss1(), loss2(), loss2()]     # loss1 is a custom loss function managing the (None, 256, 256, 3) output and loss2 is another managing the (None, 17, 4) outputs
model = Model(inputs=inp, outputs=outputs)
model.compile(loss=losses, optimizer=RMSprop(lr=start_lr))

非常感谢您！

在训练Keras嵌套模型期间会显示哪些损失？

0 个答案: