Most of the projects used some variant of the following architecture, deploying and running their systems on free serverless services (such as Modal, Hopsworks, and Hugging Face):
KTH ID2223 Projects 2022/3
This page showcases some of the “Serverless ML Services” developed by the students in the “Scalable ML and DL” masters level course. The main requirements for the project were to build a complete ML system that includes:
Most of the projects used some variant of the following architecture, deploying and running their systems on free serverless services (such as Modal, Hopsworks, and Hugging Face):
Project | Technologies | Github Repo | User Interface |
Football Betting | Modal Hopsworks |
Project | Technologies | Github Repo | User Interface |
Bitcoin Price Prediction | Hopsworks |
Project | Technologies | Github Repo | User Interface |
Ski Snow Depth Prediction | Modal | https://github.com/scalable-ml-deep-learning/predicting-snow-conditions |
Project | Technologies | Github Repo | User Interface |
Air Quality in Poland | Hopsworks |
Project | Technologies | Github Repo | User Interface |
Stockholm house prices | Hopsworks | https://github.com/Nathanotal/ScalableMachineLearning/tree/main/Projekt |
Project | Technologies | Github Repo | User Interface |
Song Recommender for Spotify Playlists | HF Spaces Datasets/Models |
Project | Technologies | Github Repo | User Interface |
Vocal Remover from music/sound files | Hopsworks Modal |
Project | Technologies | Github Repo | User Interface |
Stock Market Price Prediction using Sentiment Analysis | Hopsworks |
Project | Technologies | Github Repo | User Interface |
Hopsworks LSTMs | https://github.com/Alexanderm0/Serverless_Stock_Market_Prediction |
Project | Technologies | Github Repo | User Interface |
Water Flow Prediction for Stockholm Region | Hopsworks Notebooks |
Project | Technologies | Github Repo | User Interface |
Tesla stock price prediction | Hopsworks Modal |
Project | Technologies | Github Repo | User Interface |
Real-time traffic prediction near Kista | Hopsworks |
Project | Technologies | Github Repo | User Interface |
Reddit - predict if a post will be liked or not | Hopsworks |
Project | Technologies | Github Repo | User Interface |
Show news articles for a specified sentiment | Hopsworks |
Project | Technologies | Github Repo | User Interface |
Double the resolution of your image | Colab CNN, Google Drive | https://github.com/GianlucaRub/Scalable-Machine-Learning-and-Deep-Learning/tree/main/Project |
Project | Technologies | Github Repo | User Interface |
Electricity Price Prediction for Sweden (SE3) | Hopsworks, Modal, |
Project | Technologies | Github Repo | User Interface |
Name that Face - face recognition | Colab Google Drive |
Project | Technologies | Github Repo | User Interface |
Aurora Prediction for Kiruna | Hopsworks Modal | https://github.com/NeoForNew/ID2223_scalable_machine_learning_and_deep_learning |
Project | Technologies | Github Repo | User Interface |
Bitcoin Price Prediction | Hopsworks Modal | https://github.com/Akseluhr/ID2223-Scalable-ML/tree/master/project |
Project | Technologies | Github Repo | User Interface |
Predicting Electricity Prices in NYC | Hopsworks Modal | https://github.com/aykhazanchi/id2223-scalable-ml/tree/master/proj |
Project | Technologies | Github Repo | User Interface |
Predicting Crime in San Francisco | Hopsworks Modal |
Project | Technologies | Github Repo | User Interface |
Stance Prediction for News Topics | Hopsworks Modal |
Project | Technologies | Github Repo | User Interface |
Anime Recommender System - Two Tower model | Tensorflow Recommenders | https://github.com/backgroundhumeur/serverless-ml-anime-recommender |
This lab involved transforming this example notebook from Hugging Face that fine-tunes Whisper for Hindi. A problem with this notebook is that the first 40 minutes are feature engineering, where you don’t need a GPU (CPUs only needed). So, if you train the notebook with a GPU, the first 40 mins will be wasted. And if you want to perform hyperparameter tuning, you have to wait 40 mins for it to start. So, students refactored the notebook into a feature pipeline that wrote out 21 GBs of data (sound data in arrow file format). Then a training pipeline read in the features to train a model that was written to a model registry. Then, the Hugging Face Spaces Gradio UI (see examples below) used Whisper to transcribe sound to text.
Some students were creative and chained pre-trained models together (e.g., use Chat-GPT to summarize the transcribed text), translate the transcribed text into other languages.
Gamified Language Learning with Whisper: speak the word for the picture shown in Swedish, then it tells you if you were correct in your pronunciation or not, Github Link.
Swedish Transcription with translation to English, Github Link.
Lithuanian transcription and translation, Github Link.
Galician transcription with multiple UIs, Github Link.
Dutch transcription and summarization, Github Link.
