I have developed a face skin analyser using fast.ai which classifies different face conditions.
Link to medium article: Face skin analyzer with Fast.ai and Gradio | by Pratik Smitin Karnik | Sep, 2022 | Medium
Link to code: pratikskarnik/face_problems_analyzer at main
Link to app:
I created a simple image classifier that can be used for detecting whether the area shown in an image is flooded or not. This can help optimize relief logistics, and in much less time too.
The creation of this model was spurred by the recent flooding in Pakistan; the notebook also shows how easy it is to create powerful models in minimal time and with little effort.
Even though the model was trained on top-down images as shown below…
…it was able to perform really well on images that weren’t top down.
You can view the making of the model on my blog here: 📌 Detecting Floods for Disaster Relief | ForBlog
Or, you can view the model on Kaggle here: Detecting Floods for Disaster Relief | Kaggle
Hi folks;
I have just started the second lesson of the Fastai course. In this case, the question was to make a image classifier, whether an image is a velutine wasp or a goat wasp. And then put the model into production using Hugging Spaces.
This is a crab and velutina wasps:
As the result of the training was not satisfactory I did not have high expectations that the application in Hugging Spaces would be efficient. As expected, already with the first image I uploaded, a very clear image of a velutine wasp, the model predicted that it was a crabo wasp with the following result:
Final conclusion
it is obvious that my model has not worked satisfactorily. The reason I think is in the data, on the one hand it is not easy for the machine to differentiate one species of wasp from another, they are quite similar although to us with our vision it seems false to distinguish. On the other hand, the downloaded data contained many dubious images, nests of both species, images containing both types of wasp and others showing species other than these two.
I assume that correct data handling is one of the essential conditions for the predictive model to work. In pre-trained models this kind of problem does not arise, but with in-house data it is a challenge to handle the correct images for the model to work.
Completed my first lesson over here and ended up creating a Model for detecting if the road travelled has potholes or is it perfect road to travel on!!
Every year the amount of accidents occuring in india has been increasing , while this might certainly not be the best solution i can offer , this can be seen as a step in that direction
The link to the notebook:
Flamingos are of different species. I wanted to see if a model can identify between two species of flamingos. So I built a model to identify if the image of the flamingo is of a Lesser or Greater Flamingo.
Greater Flamingos:
Lesser Flamingos:
I get a 404 “can’t find that page” error. Have you shared the notebook to public?
I’ve made the notebook public now in the edited comment. My bad. 
In the future when robots are trying to kill us all, we’ll need a way to tell if someone is a robot disguised as a human. So I built it! Upload a pic of someone and see how likely it is that they are a host/robot from Westworld.
I built this using the Fast AI Deep Learning course jupyter notebook, Chapter 2, Production. It’s my first public model!
At DEFCON this year, the AI Village created a Kaggle competition Capture the Flag @ DEFCON Which I would recommend everyone to take a quick look. The challenges walk through ML concepts that emphasize why we need to be careful with securing models and preventing abuse as they go to production.
It was a good time, and I will definately do it again next year!
Then here is my notebook:
Just finished Lesson 4 on NLP. To help me learn the concepts, I applied them to a current Kaggle Competition (English Language Learners). Check out my blog post describing what I did.
Looks great!
Looks nice 
! I’m also still working on the challenge and figured out in the meantime
Thanks to Fastai community and Jeremy, I could deploy end-to-end model on hugging face Dogorforest - a Hugging Face Space by rmadupu, this has always been a dream to complete something end-to-end.
I will be extending this to sittingorsleepbaby, as that will help us when we are away from baby for fun.
Thank you, looking forward to learn Lesson-3
Thanks for sharing your approaches, Simon! I’d love to see your notebook or writeup for regression if you get the chance to post it. What kind of changes did you make to get deberta-v3-base to work in a Kaggle environment? I think I tried deberta-v3-base at some point, but the training failed once I hit some computing limits.
Find my notebooks here.
I hope its ok for now i just pushed everything to github as i am not in mood to write a blog post 
Please note that the code is still a raw version and i want to improhe it further but already now you can use it to generate your own models track them with wandb etc. And in its version now it runs perfectly fine
If u have any questions just let me know :))
To the training part: i trained in paperspace pushed the models to a kaggle dataset and inferenced on kaggle (as of today my score improvef to 0.45)
Huge thanks Jeremy, that’s very generous. Unfortunately I do not have twitter or social media (only Linkedin for professional reasons!), after reading Deep Work I decide to get rid of the others and focus on the actual work and the pleaure of reading and creating stuff. I will make a post in Linkedin and tag you if that is ok.
The vast majority of machine learning discussion happens on Twitter FYI. There’s very little of value in this space on LinkedIn. So if you’re just gonna use one social media thing, I would recommend twitter over LinkedIn. Just my $0.02 of course.
Hi!
I’ve deployed a mushroom classifier with 81 classes here.
I first scraped a list of species found in my home country to figure out which classes I wanted my model to recognize.
I tried using different variants of ResNet and I still need to play around with data augmentation to see if I can improve model performance.
In addition, I used Isolation Forest for anomaly detection among images in order to locate and remove irrelevant images.
Thanks for an amazing course!
Awesome, I will check them out! Congrats on your score of 0.45 
that’s very good!
