Fine_tune vs fit_one_cycle, unfreeze, fit_one_cycle

What is the difference between the next two forms of finetuning a model?

  • Method fine_tune
  • Next process:
    1. fit_one_cycle
    2. unfreeze
    3. fit_one_cycle
1 Like

If you look at fine_tunes source code, nothing. But the particular way the learning rate and other hyper parameters are done are tuned for you in fine_tune and they work OOTB pretty good for most cases


I install the latest fastai from github, but the text-learner does not have the “.fine_tune()” method. Where is it?

It’s a new method in fastai v2. So in fastai v1 you will still have to use these steps separately.

1 Like


Thanks, it seems the method language_model_learner() in fastai version 1.0.62dev is not working with arch=AWD_LSTM <error: "attributes not found backwards on line, url = ‘url_bwd’ if data.backwards else ‘url’> or arch=WT103_FWD <error: file not found, KeyError: ‘’>

Is fastai V2 is the only viable option for working with fastai.text module?

No, fastai v1 should definitely be working (I used it a couple of days ago). It looks like your error is linked to backwards, but it’s hard to tell exactly without seeing the full code.

1 Like


You are right. This morning I restart the project with a blank Jupyter Notebook, and it works fine with fastai.text module. If I have time to compare, I will report back to this thread what cause the error messages, and hopefully, others can avoid the mistake.