I am doing a lot of work with transfer learning at the moment (using keras and tensorflow if that is relevant). I am having a lot of issues in sufficiently summarizing the very large models.

This post: How do you visualize neural network architectures? shows a lot of useful methods for visualizing architectures, and they are great for networks such VGG16, but none of them are reasonable to include in a report if the models are very large (such as InceptionResNetV2 based networks).

My current approach is to simply include the depth, number of parameters and data like size and accuracy on the imagenet validation set. I would like to include more fine grained information however.

What I have tried: Several of the methods included in the post above. Exporting keras summarize to a table (issue here is that the tables for the largest networks will have over 700 rows, which just clutters the report so much in my opinion even if just included in the appendix).

So what I would like to know: do you have any recommendation for how to summarize very large network architectures in a way that is actually informative as to the inner workings without taking up 7 pages. It does not need to be in any particular format, but I would prefer a table or figure solution if it exists, and it would be perfect if it did not take up more than one or two pages for the largest models.

I feel like I have searched quite thoroughly for a solution, but I can not seem to find any.


1 Answer 1


A way to summarize a complex architecture is to use "abstract boxes" for certain known and well defined parts and computations, instead of their complete detailed architecture.

Thus a complex large model can be represented as a set of simpler abstract boxes representing different computations and layers. In this sense conciseness is retained in représentation without sacrificing clarity.

For example see the following summarization of VGG16

enter image description here

In an even larger net which includes VGG16 as subnetwork one can include only one "box" for the whole VGG16 network, since it can be considered as known and well-defined.

One can use any schematics software to draw the network summary (eg DIA)

  • 1
    $\begingroup$ Thank you for your answer. I have tried the exact approach you described but it still became far too large for the larger networks (VGG16 is one of the most shallow transfer learning networks available even though it is large). It is interesting, your point about representing the entire pre-trained network as a single block however, and just pointing to online resources. Maybe I should just do that. $\endgroup$
    – Oskar
    Commented May 29, 2022 at 15:04
  • $\begingroup$ I think you should do just that $\endgroup$
    – Nikos M.
    Commented May 29, 2022 at 17:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.