keras 同时训练 2 个模型

Question

I have 2 models: model1 and model2.

I need to take model 1 output and manipulate myData manually and set it (manipulated myData) as input of model2.

model2's output is classification of the responses of myData (to model1 output manipulation), relative to predefined classification (ie supervised).

1. I need to improve model1 output and improve model2 classification simultaneously. However in testing I will work with each model separately.
2. To my opinion I need to use model 2 cost function as model 1 cost function - how can it be done?
3. Any other idea how can it be done?

I emphasize: Concatenate does NOT solve the problem.

Please refer to the attached diagram

Answer 1

The general sketch would be as follows:

# define model 1 architecture
...

# define model 2 architecture
...

# define manipulation logic
out1 = model1.output  # get the output of model1
out1 = SomeLayer()(out1)  # apply any number of layers as you wish
...

out_final = model2(out1) # feed the manipulated output to model2

# define the joint model
final_model = Model(model1.input, out_final)

# compile the model ...
final_model.compile(loss=..., optimizer=...) # loss is computed based on the output of model2

# fit the model
final_model.fit(...)

This way both model1 and model2 will be trained simultaneously and also you can use them independently (eg use model1.predict() or model2.predict() ).