General course chat


(Chris Oosthuizen) #321

Obvious rookie Github question:
When playing and running the lesson file code, in Jupyter Notebook the files are changed and saved.
Then when I follow with a git pull to update the folders, I get an error saying that my changes are uncommitted etc.

  1. I’d like to save some changes under different file names _me. which should not be affected by git pull; and/or
  2. I’d like to chuck the changes away allowing me to keep a clean copy of the fastai repository

What is the best way to maintain these github updates. If I fork the courses v-3 folder will this allow me to play with the lesson files and do a git pull to bring in the updated and new files?


(Brian Smith) #322

My method is to duplicate all of the master copies and always work on the copies, and Jupyter has an option for this. Just select all and click Duplicate at the top of the file list. Sometimes the master copy will change again - depending if I have modified my copy I either delete or end up with Copy and Copy 2 etc.


(Salil Mishra) #323

Hello all, Is anyone taking part in Humpback Whale Challenge on Kaggle, I am facing certain problems regarding how to use the fastai library and looking for someone who can guide/help me or even better participate as team with me?


#324

There’s a thread in this subforum. Jeremy reminded us that sharing outside the official Kaggle forums or on teams is forbidden, so we are not discussing the data itself, but fastai specific questions are still fine I believe.


(Salil Mishra) #325

I am mainly facing issues with fastai library. Can you help ?


#326

Go check that thread. Some common issues already covered.


(Salil Mishra) #327

I was the one who has posted about the problem with metrics not working in create_cnn in that thread.


(EPHRAIM AYEMERE) #328

Hi @jeremy you mentioned in your fireside chat with @isamlambert that you can now have a sort of ‘plug n play’ access to fastai notebooks on GCP.

can you show me how-to in the forum pls? I spend quite tons of hours trying to access the notebook via gcp.


(Bharadwaj Krishnan) #329

I was watching some of the earlier lessons again…based on the image attached(screenshot from lesson 3). Just a random idea…probably already done


In such a scenario, I can imagine that i could have easily added even more epochs in my first attempt. Does it make sense to temporarily save the model after each epoch… so this way, if i end up overfitting, I can just load the model at an earlier epoch cycle instead of training the whole thing again? Wont this save a lot of time? Is this already done? Is it possible?


(Jeremy Howard (Admin)) #331

Please refer to the etiquette section in the FAQ.


(Rick N) #332

I posted this in Part1 a few days ago because it fit a topic there:

but I still have no answer. Since Jeremy’s nb’s always have the input prompt numbers cleared, but the cell outputs still there, clearly there’s a way to do this! Somebody please let me in on the secret… :sunglasses:


#333

It’s all done by this script.


Jupyter Notebook Enhancements [Discussion]
(Shaheen Perveen) #334

@fabsta Were you able to resolve this? I am getting PermissionDenied error. I am loading data froma different location than the one in which I have my fastai installed and it is trying to use that folder location to serach for model


(Edward Ross) #335

It makes sense to me and is possible by adding a callback.

One thing to think about is because the learning rate for fit_one_cycle changes over the course of the training fitting 10 epochs and loading the result after 9 will give you a different result than fitting for 9 epochs. No idea which would work better if underfitting (and in fact maybe it’s just better to up the regularization slightly).


(Akshay Bhardwaj) #336

Hey Edward,
I have been seeing this with my models and was not sure why this happens. Would you please elaborate on this? As in, if I train my model for 4 + 2 epochs, the training and validation error are quite different than just plainly training for 6 epochs (after seeding everything).


(Bharadwaj Krishnan) #337

good point! makes intuitive sense then to run extra epochs first cycle(overfit) and using saved epochs(using callbacks) for quick testing etc. but for getting lowest error rates(production etc), guess it would always make sense to re-run at / around the epoch cycles where error rates started increasing while overfitting. Is my understanding correct?


(Edward Ross) #338

After running a fit_one_cycle have a look at the learning rate:

one_cycle

It starts off small and goes up to lr_max and then drops quite low.

If you did e.g. 4+2 there would be two cycles like, one taking up the first two thirds of the x-axis and the other taking up the second third.

Now it doesn’t seem likely that’s a good idea as you’ll keep bouncing around, but if you drop the max_lr (so you get smaller and smaller waves) and/or increase the number of epochs you may find a better way to train the model.


(Akshay Bhardwaj) #339

Thanks for the explanation. Now it makes some sense to me.


(Dalon Lobo) #340

Hi, Thanks fast.ai and Jeremy for Part 1 of this course. That was great information and inspiring, looking forward to understand it by going through the videos. One question, is there a timeline/tentative on when Part 2 is starting?


(Deena Blumenkrantz) #341

Are the powerpoint slides from each of the class lectures available for download anywhere?
I didn’t see them in the course-v3 repo.