Clément's Projects
Byte pair encoding tokenizer as used in some large language models.
Implementation of some state-of-the-art deep learning architectures for CTR prediction tasks, both in Pytorch and Tensorflow
Fast implementation of Dynamic Time Warping algorithm using numba
Example of a Dash application applied to e-commerce data to represent key metrics in the form of an interactive dashboard
Implementation of "Evidential Deep Learning to Quantify Classification Uncertainty" proposing a method to quantify uncertainty in a neural network.
A set of examples around pytorch in Vision, Text, Reinforcement Learning, etc.
Example of a fully packaged ML model in a Docker image and exposed as a REST API using FastAPI framework
Implements SOTA architecture for CTR predictions tasks, both in Pytorch and Tensorflow
Some strategies of hierarchical time series forecasting in the context of the M5 competition hosted on Kaggle
Tensorflow implementation of three architectures for multi-task learning, a paradigm to learn different prediction tasks jointly using one model
10 or so machine learning algorithms implemented using Numpy only
Implementation of the Wide-ResNet architecture in Pytorch, as described in the original paper and used on a plant's disease image classification problem
Streamlit application to optimally re-balance a portfolio given a target allocation, current positions and market prices
Implementation of SCARF: Self-Supervised Contrastive Learning using Random Feature Corruption in Pytorch, a model learning a representation of tabular data using contrastive learning. It is inspired from SimCLR and uses a similar architecture and loss.
Simple and efficient way of performing deep ensembling to improve robustness as well as estimate uncertainty
Implements a LLM similar to Meta's Llama 2 from the ground up in PyTorch, for educational purposes.
Implements VICReg, NT-Xent and DCL losses for contrastive self-supervised learning to generate semantically meaningful representations of images without labels.
Custom component in Typescript and Python for multipage navigation in Streamlit applications
cookiecutter template for standard data science and machine learning projects in python.
WIP - Web application to optimally rebalance a portfolio (greedily) powered by Svelte
Implements the paper "Wukong: Towards a Scaling Law for Large-Scale Recommendation" from Meta.