Lesson 1 In-Class Discussion

(Ching June Hao) #210

If no mistaken, he said that backpropagation is not a real thing that happens in the brain and we should start all again

(Nikhil B ) #211

This works, thanks.

(sergii makarevych) #212

Code complexity in course part1-v2 is significantly different from code level of part1-v1. Nested classes (ModelData>ImageData>ImageClassifierData + FilesIndexArrayDataset + FilesIndexArrayRegressionDataset ), decorators, tons of parameter passing

def __init__(self, path, datasets, bs, num_workers, classes):
    trn_ds,val_ds,fix_ds,aug_ds,test_ds,test_aug_ds = datasets
    self.path,self.bs,self.num_workers,self.classes = path,bs,num_workers,classes
    self.trn_dl,self.val_dl,self.fix_dl,self.aug_dl,self.test_dl,self.test_aug_dl = [
        self.get_dl(ds,shuf) for ds,shuf in [

It looks like spaceship much more complicated than it was a year ago and I wonder why?

(Jeremy Howard) #213

(Jeremy Howard) #214

(Jeremy Howard) #215

We’ve built a whole new framework, not just a few little wrapper classes. It’s so we can go far deeper than last year, get much better results, but do it with less/simpler code in the notebooks.

More advanced students can help contribute to the new library.

(Jeremy Howard) #216

I’ve added the recordings to the top post of this thread, BTW.

(Santhanam Elumalai) #217

@jeremy is the framework contribution is restricted to the advanced level candidates? Because, If you have a thread with the SOP’s, Test cases and design documents people can pull over and start forking the framework’s repo, you can evaluate the merge request with the test cases, This will give you more quality and productivity in-terms of frame-work.

(Santhanam Elumalai) #218

You can ask some of us to create the SOP and design documents and you can validate it for the pre-design phase.

(Jeremy Howard) #219

Anyone is welcome to contribute!

(Santhanam Elumalai) #220

Thank you, Jeremy. Any Forum thread for the contributors?


Is there a way to set the code to run in CPU only mode? ( like v1 class).

When i ran the model, got the below Assertion error:

~/anaconda3/lib/python3.6/site-packages/torch/cuda/init.py in _check_driver()
56 Found no NVIDIA driver on your system. Please check that you
57 have an NVIDIA GPU and installed a driver from
—> 58 http://www.nvidia.com/Download/index.aspx""")
59 else:
60 # TODO: directly link to the alternative bin that needs install

Found no NVIDIA driver on your system. Please check that you
have an NVIDIA GPU and installed a driver from

(Jeremy Howard) #222

Not currently. It wouldn’t take much to modify fastai to make this possible, but I haven’t had the time yet. If anyone wants to contribute this, feel free! :slight_smile:

(Jeremy Howard) #223

Not as yet, but if you have some code to discuss, feel free to create one.

(Santhanam Elumalai) #224

Sure, I will.


Ok. Thanks for your response.

(yinterian) #226

No. There are two plots. In one plot you have x=mini-batch y = learning rate. This is showing how learning rate is increasing as a function of mini-batch. This is for the algorithm that find best learning rate.
The other plot is learning rate as a function of loss.

(Nandamuri Hari Naga Sumanth) #227

when will be remaining part of lesson available ??


What do you mean @HariSumanth9 ? I think it’s all already there.

(James Requa) #229

Check the top of this thread, Jeremy has posted the recorded video links for both section 1 and 2.