Share your work here ✅

I’ve been keeping a list of docker containers since v1.0.60, and continue to build new versions as they get released.

Today I added support for 2.7.2 with PyTorch 1.10.0 or 1.11.0:


I’ve just trained my first decent ensemble image model on kaggle and am moving up the leaderboard thanks to Jeremy’s Walkthrus and kaggle experiments!


I’m also using fastai to take part in an ongoing UW-Madison GI Tract Image Segmentation competition. (Shared some examples before.) My results aren’t that impressive, and the competition is still running. But I haven’t been at such rank (even temporarily) for a long time.

No matter how it ends, it feels like now I am iterating pretty quickly and spend much less time writing boilerplate while focusing on modeling instead. Also, competing became less stressful and more joyful because of the shifted focus.


Top 10% is pretty good!


In this one I wanted to show how you can BOTH train and deploy a :red_car: license plate recognition system :brain: in a fully Dockerized way.

The serving container just adds a couple of libraries on top of the training one, guaranteeing a completely reproducible end2end pipeline.

If you are into Docker, Amazon Textract, IceVision and FastAPI feel free to dive in!

Twitter, LinkedIn posts


I participated in the fastai hackathon this week and with the help of my team @AllenK @Rkap and @Sanjib we created an incident priorisation app: Risk Predictor - a Hugging Face Space by mrosinski


Based on Jeremy’s notebook on Kaggle - Getting Started with NLP for absolute beginners, I tried replicating the same approach on the dataset Women’s Clothing Reviews to determine if a customer will recommend the product based on clothing reviews.

Kaggle Notebook: getting-started-with-nlp-womens-clothing-reviews | Kaggle

Using the text fields alone such as title and review, trained a deberta-v3-small transformer model to predict the output.

Next Steps

  • This is a rich dataset consisting of both text, categorical and numerical fields and will be experimenting with some of the ideas shared in this topic Tabular Model with NLP Text features

I have created two gradio apps based on two kaggle competitions which I have been participating in. Both are on computer vision models with one to identify the type of disease in the paddy crop and another to identify the name of flowers(which I am terribly bad at remembering names).


Thanks for posting. I was curious to see your use of: learn = from_pretrained_fastai(repo_id)
so found the doc here: Mixins & serialization methods
and you model repo here: kurianbenoy/paddy_convnext_model at main

What steps did you take to upload your model?
Was it using push_to_hub_fastai()? …like described here: Google Colab

1 Like

Yeah it’s exactly like the steps in the google colab notebook. If you are not developing from google colab, you need to install git-lfs and generate huggingface token which was slightly covered in the course also it seems :slight_smile:

1 Like

I manage to build and train a graph convolutional neural network (GCN) for lipophilicity (logP) prediction of small molecules :).

GCN are not available in fastai, but we can build our own architecture thanks to the flexibility in the Learner and DataLoaders classes!

The hardest part was preprocessing the data and putting it in a format suitable for training. I also had problems understanding the update function, but I followed an important piece of advice from the course: pay attention to tensor shapes!

I made the dataset and model available at: GitHub - marcossantanaioc/graph_neural_networks: Fastai implementation of Kipf & Welling (2017) Graph convolutional neural network (GCN) for Cheminformatics tasks


Looks great, I think the link is broken.


Second time around doing the course. First time I didn’t give myself to learn and enjoy being a beginner so I never used to make small projects like these for learning purposes.

Anyway, second chance now and I plan on doing a lot of small learning projects. In this one I trained a model that can differentiate between a caterpillar, a centipede and a millipede. The inspiration came from when I had a caterpillar sting but I was freaking out and thought it the insect was a centipede until I googled the difference.

Enjoy :slight_smile:


I made my first successful submission to HuBMAP + HPA - Hacking the Human Body competition. I use fastai as much as I can. My objective was to make a successful submission, which was not easy for me. It is a code competition, so you need to keep your kernel work offline. My score is not great :poop: :poop: :poop:, but I trained my model for only 2 epochs with a very tiny dataset.
It says 3rd submission but at the first two, I’ve got scoring error. I’ll try to make it better every day.


I’m going to share my fire detection model. Thanks, Jeremy for the inspiration from the first lecture. You can find all the steps in my medium post.


These couple of days I worked on polishing my model that detects fire in the forests for my thesis. I uploaded it to hugging face. I encourage anyone to try and break it. Thanks, @jeremy for all the tips and tricks learned in the lessons. Later I will share the notebook. I just need time to make it prettier so that anyone can see what I did.


Hey all, sharing my first DL project ever that I created using fastAI by following Jeremy’s lesson 1 notebook. I used 3 Different datasets ( Tiger Vs Cat, Happy Vs Sad Vs Angry images, Backpack Vs Purse) , and used those to create the models and predict. I’m not much aware about the technicalities of how everything is working or what are the exact underlying metrics, but just trusting the top-down approach and hoping to learn along the way :slight_smile:
Sharing the project here, would love any feedback.



After the first lesson and seeing the 101-food classification app by suvash, I modified the code by modifying the searches to be a list of fruits to search for all the images. I then use the same code to do a classification model, using resnet50. It’s amazing that it’s not too hard to make an app like this. The accuracy is about 75% at the moment. Cannot wait to improve the accuracy using the strategy in lecture 2: doing several rounds of cleaning the data.


Greetings, I’m sharing my first project that uses fast AI. I downloaded the Rice Image dataset from Kaggle and followed the video tutorial.

My model is good at predicting rice, but I uploaded a picture of a bear instead of rice, saying it is a rice type. Does that mean the model is too good?

Thanks, and I’m hoping to hear from you.