You custom architecture

Part 1 (2019)
1

How can you pass in your custom architecture in create_cnn. I understand we will make a body and head so that we can do differentiable learing rates. I tried directly passing in my arch but it did not work.

The docs say:

You can customize create_cnn for your own model’s default cut and split_on functions by adding them to the dictionary model_meta .

But there is no info on model_meta in the docs. What is it and how can I do this?

2

If you want answers for such advanced questions you have to dive into the source code and not just read the documentation.
model_meta is a variable in learner.py. It’s a dictionnary that stores the integer cut and the function split.

3

Actually, in fastai-0.7, this could be done pretty easily as I remember, you would just pass in your custom model as one of the args.

4

Though agree, I would say that some parts of fastai’s API are a bit obscure =)