Fastai v2 academic paper now available

Here’s the peer reviewed paper that @sgugger and I wrote, with lots of help from @AlexisGallagher, about fastai v2:


Congrats! Great reading before the next course!

Great article and an excellent summary of fastai v2. I have a minor remark. I think the article was written before refactoring DataBlock.databunch() to DataBlock.dataloaders(); the code excerpts included in the paper might be updated in order to be in sync with current version :slightly_smiling_face:

1 Like

Slight error in case it wasn’t already pointed out: It says “For instance, here is how fastai adds the read() method to the pathlib.Path class” but the code example patches the write method.

Also, I used fastai v1 for my paper and currently have a reference to the github. Not sure if I should keep that as that’s what I used or link this one instead.


Maybe cite the paper, but mention you used v1?

1 Like

Good news!

Congratulations @jeremy and @sgugger! Thank you for the important and immensely useful work that you do.

For Figure 11, the image on the right actually looks like it has greater loss of texture, especially if you look at the eyes and the fur. Did the references to left and right images in the caption get reversed?

I am looking forward to reading the upcoming book.

1 Like

Indeed, will fix that, thanks for pointing it out!

I am coming to this post to point out this and found that someone have already found it. :slight_smile:

I was going to comment that but you beat me to it :slight_smile:

I just completed reading my first academic paper end to end without sleeping!

Merci @sgugger for keeping the writing conversational and easy to read.

Honestly speaking, having just finished part 1 of fastai v1, I was getting very anxious about the buzz around v2 in the groups. Somehow, reading this paper made me eager to look in those groups and follow along. So anyone else out there feeling anxious about v2, I’d highly suggest reading the paper. It reads like fastai’s life story!

With no intention to nit-pick, thought I’d call out two things which I noticed on my first read.

  1. In the introduction there is a line in the last paragraph
    While that code made change slightly as the library
    I imagine you meant may*

  2. In nbdev, there is a line:
    Automatically create searchable, hyperlinked documentation from your code (as seen in figure [fig:nbdev];
    The hyperlink doesn’t link to any image. (blog only problem)

Also, is this post available in .md format? I just began using fast_template and I was curious to see the formatting, for eg. all images have nice boxes around them.

Thanks guys! See you on the v2 forums!


Hi isn’t this a typo? I just can’t comprehend this sentence :slight_smile:

Numericalization and vocabulary creation often requires many lines of code, and careful management here fails and caching.

Let’s now see who’s the first to cite it :slight_smile:

As I came to revisit the post, maybe you forgot to fix it? :stuck_out_tongue:

I definitely fixed it for the version on MDPI. Maybe we forgot to update the version on arxiv.


  1. Missing/Extra bracket in code example in Section 2.2 Text. metrics=Perplexity()]

    learn = language_model_learner(dls_lm, AWD_LSTM, metrics=Perplexity()])
  2. In code example for tabular(2.3) and collab(2.4) section fastai2 is used instead of fastai.

    from fastai2.tabular.all import *
    from fastai2.collab import *