Save Learner attributes like recorder

fastai
1

I’d like to save my Learner, but not just its weights.
I want to be able to call learn.recorder.plot_losses() after for instance image, get top_losses, etc.

Is it possible? Because if I do learn.save() and then learn.load(), I get the following error when I try to call learn.recorder.plot_losses() :

  • AttributeError: ‘Learner’ object has no attribute ‘recorder’
2 Likes
2

It is not exactly what you want but you can use callbacks.csv_logger to save your losses after train.

You can find how to use it here: https://docs.fast.ai/callbacks.csv_logger.html

Not sure but I think learn.load() just load the weights.

1 Like
3

Thanks @dhoa,
When i try

  • learn = Learner(data, simple_cnn((3, 16, 16, 2)), metrics=[accuracy, error_rate], callback_fns=[CSVLogger])
    i get the following error:
  • NameError: name ‘CSVLogger’ is not defined

image
image.png1014×196 12.1 KB

Do you know why?

Thanks

4

Try callbacks.csv_logger ? I remember got this problem too.

Hope that helps

5

It works, thanks @dhoa.
But i still want to be able to call any method of learn after, since i’m using a server to train, and i don’t have a visual environement (I can just download outputs from the python code). If anyone knows how to save everything (including recorder), please share with us.

Thanks

6

In any plot, you can pass return_fig=True and then use matplotlib’s .savefig or .save function to save it locally. (I can’t 100% recall which one). It’s not quite what you’re wanting but atleast you can save the graphs.

Hope it helps.

1 Like
7

Hi, for those who want to directly append the CSV logger to the learner’s callback functions, here is the line :

learner.callback_fns.append(partial(callbacks.CSVLogger))

If you want to append futur training sessions logs to your existing file, don’t forget the argument append=True. Hence, the line should look like this :

learner.callback_fns.append(partial(callbacks.CSVLogger, append=True))

I’ve just checked the exact synthax a minute ago. :slight_smile:

8

Hi ,

I’m getting issues while the accuracy/loss plot And I also want t check my final accuracy of the model.

How can I fix that?

image
image648×589 42.9 KB

Thanks,