问题
When using Inception v3 model in Keras, nor the graph of the network nor the model.summary() indicate the presence of the auxiliary classifiers (as in Szegedy et al..
Why is it so ? Is Keras still using the right architecture but hiding this specificity from the user ? If so, how can we then customize upper layers from the network ? Indeed, we may want the 3 classifiers to have distinct architectures.
Thank you !
来源：https://stackoverflow.com/questions/44906653/why-is-there-no-trace-of-auxiliary-classifiers-of-the-inception-v3-model-in-kera