Colaboratory and Fastai

Part 1 (2018)
21

Thank you. I am also using colab instead of amazon aws and paperspace. I could not locate the notebook for Part 1 version 2. Can anyone link me up?

22

We need to upload them?
Or find them with the fast.ai directory?

23

if we use paperspace the notebooks get downloaded while setting up the environment

24

I searched for the notebooks of part 1 version 2 but couldnā€™t find.

25

The Jupyter notebooks for 2018 version (part 1 v2) of the course is in fastai GitHub repo. Hereā€™s the direct link to get all the notebooks: https://github.com/fastai/fastai/tree/master/courses/dl1

1 Like
26

Yes. Thatā€™s because, during the setup process, one of the command in the Paperspace script git clone the fastai GitHub repo where all the notebooks is.

27

Okā€¦ Thank you @cedric for your help :slight_smile:

28

Thank you so much!
I followed your notebook, when I got to " !mkdir data && wget http://files.fast.ai/data/dogscats.zip && unzip dogscats.zip -d data/," the cell is successfully run. However, the image folder does not appear in my folder.
I tried other commands, they all appear successfully downloaded inside my Colab notebook,however they donā€™t show up in the folder.

When I tried the command above again, it says,ā€œmkdir: cannot create directory ā€˜dataā€™: File exists.ā€ And I couldnā€™t find it anywhere in google drive.

I tried creating new folders and start new colabs,but wget doesnā€™t work.
This doesnā€™t work either:
"!wget https://raw.githubusercontent.com/vincentarelbundock/Rdatasets/master/csv/datasets/Titanic.csv -P drive/app"

29

hi
can you explain how you did it exactly iā€™m getting same error

RuntimeError: cuda runtime error (2) :out of memoryat/pytorch/torch/lib/THC/generic/THCStorage.cu:58

thank you

1 Like
30

Hi,

They are not stored in your google drive. They will be just temp files on that particular instance.

3 Likes
31

Hi Dinesh,

So what I did to fix the issue was to fork the fastai GitHub repository and then in the ā€˜conv_learner.pyā€™ file, on lines 48 and 50 there was a call for ā€˜to_gpu()ā€™ function, I simply deleted the call to gpu and cloned my edited version of the library on colab and it worked. You can clone my altered version of fastai here:

Iā€™d recommend you also check to see if your GPU is actually out of memory, because then that is a completely different issue as well. Also, itā€™s worth noting that since the time I made my original post, I have not had that issue again at all on Google Colab. Iā€™m not sure why that is but Iā€™ve been able to run everything that was throwing the issue previously, without any problems. I hope all of this helps.

Feel free to ask any other questions!

Best,
Jacob

4 Likes
32

Hi Jacob,
Thanks for the reply. I need one more thing from you.
How did you link your edited version onto colab notebook?

33

Hi Dinesh,

Happy to help. First, ensure that you do not have the main version of fastai installed, if you installed using pip, run ā€˜!pip uninstall fastaiā€™ should work, otherwise try ā€˜!rm -rf fastaiā€™. I linked my edited version into the colab notebook using ā€˜!git clone ā€œpaste link hereā€ā€™. Hopefully this should work for you too. If you run into an issue using the edited version of fastai let me know, Iā€™ll do what I can to fix it but of course you may fork your own version on GitHub and work on it yourself if you like!

2 Likes
34

Strange! I ran the notebook as is (at least lesson-1.ipynb) without the to_gpu edits. It seem to work just fine. I wonder what the difference is

1 Like
35

Colab GPU is almost twice as slow as using the CPU on my laptop. Has anyone else noticed this?

36

Itā€™s definitely been weird, Iā€™ve had it work without the to_gpu edits as well but as youā€™ve seen I also had to fix it once before. Since Iā€™ve done that fix, I havenā€™t had an issue. Hereā€™s hoping that issue doesnā€™t come up again!

37

I am getting error
AttributeError: module 'PIL.Image' has no attribute 'register_extensions'
for the cell
img = plt.imread(f'{PATH}valid/cats/{files[0]}') plt.imshow(img);
when I am running the notebook for lesson1 on google colab. I following this guide


Anyone who got the same error, or whizzed past this?

1 Like
38

Yes, Iā€™ve had the same error yesterday for a while, but then as mentioned in the forum I commented out:

#%reload_ext autoreload

#%autoreload 2

%matplotlib inline

and that worked fine.
But today itā€™s not working even after commenting out those 2 lines.
I too am using Google Colab.

2 Likes
39

I had an issue with PIL at one point and was able to fix it by installing version 4.0.0 (I think) of pillow. Not sure if yours is the same issue I had but downgrading pillow may be worth a try since it is a one line bash command, I forget the exact command but a quick google search will show it. Pillow contains PIL and I think some other image library, so that is why it is the thing to get older version of.

1 Like
40

Yes my issue is resolved now by downgrading Pillow to 4.0.0

3 Likes