About the fastai dev category

Just wanted to point to Pytoune in case it turns out to be helpful.

Also note that we’re using pytorch from master plus a recent PR that’s not merged yet.


I might be late here, is there still help needed? What would be the best way to figure what to pick up? Post here: Fastai dev study group at USF ?

Yup. We’re posting what we’re working on as we go, so just follow this category to see what’s happening! :slight_smile:

1 Like

I just added links to the various design notes to the top post.


Do we need to install from source or can we just conda install pytorch-nightly -c pytorch to use the latest v.1 builds?

I haven’t tested it, but I think the latest pytorch-nightly should work.

Can confirm it works. Thanks.

FYI: I’m running this against a conda environment that uses python 3.7 on a machine with previously installed conda environments that use python 3.6 … and I noticed that without logging back into my server, jupyter would always set my python incorrectly to 3.6 (even though I had installed jupyter into the 3.7 environment and activated it before running).

Don’t know if there is a better way of handling this than restarting the server or just logging back in, but I post this here in case folks notice the same behavior.

BTW you can use py36 with fastai v1, as long as you pip install dataclasses.

Are any folks running this on ubuntu 18.04?

Thinking about upgrading my 16.04 machine (which is working just fine) and wondering what headaches I would create for myself if I go through with it.

btw, the first notebook ran flawlessly using that latest preview of pytorch v1 and python 3.7.


from the SourceURL:https://github.com/fastai/fastai_v1

Requires the installation of pytorch from the source.
to use the master branch of pytorch master

For Windows it requires Visual Studio to compile it.

Is there any restriction/limitation to use the pytorch version that is currently available with the pip or conda commands?

Sadly yes: you will get an error with data augmentation in computer vision with the default options (the pytorch function we used doesn’t have padding_mode=reflect in 0.4.1). If anyone manages to get wheels for windows that work to compile master, that would also interest me deeply.

Hopefully, pytorch v1 will be available the day of the conference so we only have a week to wait for it. In the meantime, you can get around this problem by adding the option padding_mode=‘zeros’ or padding_mode=‘border’ when possible. I’ll try to finish the documentation for the vision transforms a bit later today (notebook docs_src/vision.transform.ipynb) so you can see how it’s done.

Turns out it’ll just be a preview. But you can install the preview from conda now (see the ‘nightly’ links on the pytorch home page).

Not on windows :frowning:

Many of us working on “Shadow IT” are not allowed to run installers without admin rights on the PCs.

Hello Everyone,

First congratulations on the lauch.

Question: those amazing developer notebooks dev_nb are gone until the new course starts?

Thank you


They are in the fastai_docs repository:


Thank you Marco.

Congratulations again to the team.

How classes like *Dataset like FilesDataset will be substituted? I cannot find them in the new fastai/fastai folder.

Also there if a problem after moving old fastai code to /old folder since now I cannot pull latest changes without breaking my code. Wouldn’t it be better to move it to another branch and label it with fastai version instead?