Just finished lesson 2 - inspired by the risks of misinformation from Deepfake videos I’ve made a proof of concept around using Convolutional Neural Networks for identification of deepfakes. Here, I have put together a - Real or Deepfake Tom Cruise classifier to showcase this functionality.
It seems pretty accurate with anything I feed it - give it a go here! I’ve also written about the idea and the code here as part of my learning blog (the blog is still very much in development so please bare with the amount of broken links and formatting issues etc.!). Any helpful tips would be greatly appreciated!
Finally I was able to complete the tutorial 2, deploying the model to production on hugging face.
As the course was updated 2 years ago, there are some changes in versions and syntax.
I have collated all the issues which I faced while doing this, hope this might help if anyone faces the similar issues.
Did a pretty good job to classify europe and bangladesh images though. Don’t know why this worked so well. It even correctly classified turkey even though we have similar architecture.
The Tree of Thoughts framework demonstrates the power of combining the reasoning capabilities of LLMs with tree search algorithms (BFS & DFS). I’ve written a blog post that makes the framework easy to understand by taking you through the details in a step-by-step manner: Understanding and Implementing the Tree of Thoughts Paradigm
If you are interested in the topics of LLM reasoning, planning and tree/graph search, then you may find the blog post useful.
I finished the second lecture and got my first deployed model up! I tried to classify images of faces and determine which emotion they were most likely to be experiencing. I go through the ups and downs of this process here Emotions_Classifier
Finished Lesson 1.
Trying to modify the ‘bird’ identification homework.
I just used a simple comparison between a School bus & Transport buses. It was interesting to see the results.
For my L2 Deployment Project I decided to create a Tournament Anime Classifier. The model was trained on images of characters from the anime Baki/Baki Hanma, Kengan Ashura, and Record of Ragnarok. Due to the crossover movie Baki vs Kengan Ashura I had to ensure that I cleaned my dataset to remove images containing characters from both anime.
My anime lovers, please take a look :)! Check out @DEVUDILIP post if you are running into issues with deploying to huggingface spaces! His recent post should push you in the right direction.
For lesson three I basically just worked through the MNIST multi-class classification project. I broke it up into 2 posts on my blog here. You can also just draw on this huggingface spaces app to get a sense of what I was able to accomplish!
For lesson 1, I’ve trained resnet152 on mel spectrograms of different genres of music. It seems to work really well out of the box but still confuses reggae to be blues or hip hop and few other similar mistakes.
You can find the code here
I’ve recently been working on a deep learning project to classify plant leaf images based on their health status and crop type. Using the PlantVillage dataset and FastAI, I built a robust multiclass image classifier with ConvNeXt-Small22k and fine-tuned it to improve performance.
Model: Trained with fastai, exported to ONNX
Dataset: Skin lesion images from Roboway