I am trying to do my own pretraining of neural nets on Imagenet, because
- I want to understand how the current state of the art pretraining is being done
- I want to use the same procedure with a more domain-specific data set, and compare the results
- I want to do the same thing with customized and home-grown neural nets
So far, I have found two answers to the pretraining procedure, both in the fastai github repo:
The first one does things in a “clean” fast.ai way, and was last modified in April 2019, while the second one is more pytorch heavy, but although it is from 2018, it seems to be pointed to more often.
So, which one should I use? And where can I find the training script (or something functionally equivalent) that is used to pretrain the “official” nets in fast.ai?
Edit: link trouble