USF daily study group

simplyjk · October 30, 2018, 8:53pm

Where’s the location of the in-person classes today?

Jaghachi · October 30, 2018, 11:06pm

Will this study group become the daily one with at a set time and room?

jeremy · October 30, 2018, 11:57pm

It already is. Details in top post.

Jaghachi · October 30, 2018, 11:58pm

Oh my gosh! how did I miss that! Thank you!

gazay · October 31, 2018, 2:51am

I didn’t look into code of FileDeleter yet, but if it’s deleting images from disk completely, there is an idea in Lesson2 chat to just move it in special e.g. ./trash folder to be able to undo process or use those images later if they needed for submition or anything else

jeremy · October 31, 2018, 4:24am

That sounds like a much better approach @gazay! cc @zachcaceres @lesscomfortable

zachcaceres · October 31, 2018, 5:25am

cool idea @gazay definitely can be done!

gazay · October 31, 2018, 10:26pm

We discussed another idea – why analyze numbers of training by hand when library can analyze it for you and give some advices. For example we trained network for 5 epochs and experienced user will notice that it overfits. But if user just started using library – it’s not so clear what is happening, how to interpret those numbers and what could be done. So we can programm some rules and make it as optional flag – if our rules see that validation loss in every epoch is bigger than train loss – maybe it’s overfitting. And suggest or even try to retrain with different hyperparams.
Possible API:

Total time: 01:07
epoch  train_loss  valid_loss  error_rate
1      1.172217    0.306988    0.088490    (00:17)
2      0.506414    0.436349    0.075183    (00:16)
3      0.314256    0.598335    0.077219    (00:16)
4      0.232976    0.693311    0.081876    (00:16)

!! It's looks like your model is overfitting. Try to apply one of these advices and retrain it: https://link-to-faq-for-overfit/underfit

or

learn.fit_one_cycle(4)
Total time: 01:07
epoch  train_loss  valid_loss  error_rate
1      1.172217    0.306988    0.088490    (00:17)
2      0.506414    0.436349    0.075183    (00:16)
3      0.314256    0.598335    0.077219    (00:16)
4      0.232976    0.693311    0.081876    (00:16)

# in next cell:
learn.recorder.analyze()

=> In epoch 3 and 4 looks like your model is overfitting. Try to apply these technics and retrain your model to get better results: https://link-to-faq-for-overfit/underfit

jeremy · November 1, 2018, 1:53pm

@gazay that sounds like a nice idea for a callback! Will be tricky to get the over-fitting diagnostics really reliable, what certainly worth a try…

eosoba · November 9, 2018, 10:21pm

Is there still space available?

jeremy · November 10, 2018, 6:58pm

Sure!

jyoti3 · November 14, 2018, 4:30am

I would love to join and see if i can contribute anything. Can i still join?

etown · November 14, 2018, 4:34am

I would love to join. I have an interesting medical application I would love to get some feedback on my approach.

Also looking to contribute to the library.

marii · January 8, 2019, 8:06pm

Is anyone still coming to this group? If you are we might need to schedule times as the room is not as busy as before.

Currently in the same room if anyone wants to join.

yubozhao · July 1, 2019, 6:49pm

Just want to check-in. Is the daily study group still active?

marii · July 1, 2019, 9:25pm

No, it is no longer active. There is a study group that meets in the same room on Tuesdays though: https://www.meetup.com/PyLadiesSF/events/wcczcryzkbdb/ for which I am an organizer.

yubozhao · July 2, 2019, 12:35am

Thanks, Molly! I will check it out.