在 multi_gpu_model 之后访问 Keras 中的中间层

Question

As stated in this issue , Keras' multi_gpu_model alters the apparent layer structure of a model.

For instance, when the layers from an original model looks are:

>>> [l.name for l in my_model.layers]
[input, conv2d_1, conv2d_2, maxpool_1, conv2d_3, maxpool_2, dense_1, dense_2, dense_3]

After using multi_gpu_model(my_model) , then the layers will become:

>>> new_model = multi_gpu_model(my_model)
>>> [l.name for l in new_model.layers]
[input, lambda_1, lambda_2, lambda_3, lambda_4, model_1, dense_3]

I am trying to access the output from the layer maxpool_2 from the original model, in order to use it into a different, subsequent model. How can I access the original maxpool_2 layer's output after using multi_gpu_model ?

Note 1: as old_model has not been trained, I have no interest in getting the output from old_model.get_layer('maxpool_2') . Only the outputs from the trained new_model are interesting here.

Note 2: trying to call new_model.get_layer('maxpool_2') will trigger ValueError: No such layer: maxpool_2.

Answer 1

From Keras source code it seems that the original model shares its weights with the multi-gpu model:

Therefore suffice to do the following the access the intermediate layer:

old_model.get_layer('maxpool_2').output

Which will use the output of the trained model.