Fastai2 Blog Posts, Projects, and Tutorials

marcossantana · August 29, 2020, 6:42am

Thanks ahahha
I will update with fine-tuning as soon as I find out how to save the vocab in fastaiv2.
Edit:
For those wondering how to save the vocab:

with open(LM/'vocab.pkl', 'wb') as f:
  pickle.dump(learn.dls.vocab, f)

muellerzr · August 29, 2020, 12:38pm

You can also do torch.save(dls.vocab) then load it back in with torch.load()

morgan · August 31, 2020, 9:52am

@muellerzr do you think its worth me submitting a dls.save type PR? v1 had similar right? Especially hand for nlp where you’d like to keep the vocab intact.

jwuphysics · August 31, 2020, 1:28pm

Hi y’all! I only started writing a research blog on Deep Learning x Astrophysics a few months ago, and wanted to share some highlights:

Using the Fastai (v2) library, we can train a CNN regression model to predict elemental abundances from the morphologies of galaxies (post):
By using Pytorch/Fastai Hooks, we can use dimensionality reduction tools to visualize the feature space (activation maps) of galaxy morphology (post):

image773×535 49 KB
We can adapt our previous model to perform classification, and then interpret the CNN using Grad-CAM for image attribution (post):

image1260×434 121 KB

Also, I joined Twitter not too long ago – feel free to follow if you’re interested in astronomical applications for deep learning!

lschmiddey · September 18, 2020, 8:46am

Hey guys,
I only started my own data science blog following along the beautiful book “Deep Learning for Coders with fastai&PyTorch”. So yes I use fastpages for my blog

In my latest blogpost I used fastai’s DataBlock to build a classifier on Lego figurines. Absolutely glorious.

https://lschmiddey.github.io/fastpages_/2020/09/16/Lego-Classification.html

I hope you enjoy it!

Lasse

joedockrill · September 18, 2020, 9:33am

meh. not really as impressive as teddies vs grizzly bears though is it?

mrfabulous1 · September 19, 2020, 4:15am

Hi lschmiddey Hope your having a lovely day!
Great concise and enjoyable post!
Cheers mrfabuous1

zerotosingularity · September 19, 2020, 5:19am

Upvote!

muellerzr · September 20, 2020, 6:29pm

Just posted @Juvian and I’s results on CamVid. Short answer: there is no consistency in how models are being trained for benchmarks so we can’t truly compare anything. A better option is to train on the CityScapes dataset, as it’s setup more like Kaggle so there is no possible way to actually get the true test set and reach the same conclusion (though it is for research purposes only). Along with this, if we were to “compare” how fastai’s dynamic unet does, it’s pretty decent with only a resnet34 backbone, and we saw a significant boost when applying Mish to the head of the network (couldn’t quite get it to replace the actual backbone w/ the activation functions well). If you’d like to read more see here: https://muellerzr.github.io/fastblog/papers/2020/09/18/CAMVID.html

marcossantana · September 26, 2020, 3:02am

Hi guys,

I’m experimenting with transfering Pytorch/Keras/TF models to Fastai. My last post was about molecule generation using LSTM. Today I bring you a Message Passing Neural Net for bioactivity prediction!

The model architeture comes from this paper: https://jcheminf.biomedcentral.com/articles/10.1186/s13321-019-0407-y

And you can find the original implementation here: https://github.com/edvardlindelof/graph-neural-networks-for-drug-discovery

For those of you looking for new, cool models to use for drug discovery studies using fastai, here’s my notebook showing the working model:

github.com

Marcosuff/Marcosuff.github.io/blob/master/nbs/Graph_NN.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "Graph-NN.ipynb",
      "provenance": [],
      "collapsed_sections": [
        "H9Rwatasfevc",
        "FopBJgee1B_J"
      ],
      "authorship_tag": "ABX9TyO/xBLHCoa6i/GvltdhK/mO",
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "accelerator": "GPU"
  },

This file has been truncated. show original

imrandude · September 26, 2020, 6:05am

File not found, for the link shared.

marcossantana · September 26, 2020, 6:49am

just edited

lschmiddey · September 28, 2020, 6:00pm

Hey guys!
I started a little project where I created a web app with binder to predict my ratings for a book given the text of an uploaded image.

I didn’t get the example of Jeremy working (voila_bears) so if anyone of you tries to make Binder work, check out my repo: https://github.com/lschmiddey/book_recommender_voila

My blogposts to this project can be found here: https://lschmiddey.github.io/fastpages_/2020/09/28/Build-binder-app-Part4.html

I hope you enjoy it!

Lasse

butchland · October 7, 2020, 1:14am

Hi all,

I just published a two-part introductory article on building your own image classifier for deep learning beginners - if you’ve already completed Lessons 1-3 this is nothing new, but my target audience is myself before I started Deep Learning 2 years ago.

Its a hands-on article so it walks you through building a dataset and a model on Colab and running it on binder. If you have problems in the walk-through, I’d appreciate a comment so I can fix it…

Here’s the link : https://medium.com/@butchland/build-and-run-your-own-image-classifier-using-colab-binder-github-and-google-drive-part-1-bd1aebc626e

Best regards,

butchland · October 7, 2020, 3:44am

btw, forgot to mention thanks to @joedockrill for the jmd_imagescraper package and @vikbehal for the binder instructions – hope you get to see your contributions in the article!

Best regards,
Butch

dinodelao · October 8, 2020, 11:23pm

Hi all, my name is Dino and I am new to fastai. I absolutely love the material and the teaching style. I’ve read the first 7 chapters in the book and listened to the lectures online. Before moving further into the book, I decided to jump into a project and I was able to create a multilabel classification model on the “chinese-mnist” dataset. I would appreciate any feedback and I am really excited about moving forward with this course.

https://www.kaggle.com/dinodelao/fastai-v2-gpu-for-chinese-mnist-prediction

Rakeshsuku · October 21, 2020, 6:32am

Hi All,

I have written a couple of blog posts explaining the workings of FastAI Optimizers and FastAI training loop. I have read the source code for these and have demonstrated my understanding with an Image classification example. Find the link to my posts below. Hopes this will be helpful.

Study the resnet34 model architecture and build it using plain Python & PyTorch.
Deep dive into FastAI optimizers & implement a NAdam optimizer.
Study FastAI Learner and Callbacks & implement a learning rate finder (lr_find method) with callbacks.

Regards
Rakesh Sukumar

sukhdeepgill · November 16, 2020, 5:19am

I’ve written my blog around the stacking of tensors using PyTorch.

For e.g. I’ve used MNIST dataset. This is my first blog in deep learning. So, please give it a read and let me know if any changes need to be made.

Thank you

karynaur · November 29, 2020, 2:12pm

Just added my first blog on image classification using fast.ai’s vision library and jmd_imagescraper !! Please feel free to critique it.
You can find it at: A Beginner’s guide to Deep Learning

johannesstutz · December 16, 2020, 12:21pm

New blog post:

Check it out and please let me know if you find the visualizations helpful, or what can be improved