TensorFlow

Machine-learning models in the browser with TensorFlow.js and Bun.

Ce projet constitue mon projet de fin d'études et porte sur la conception et le déploiement d'une solution de vision par ordinateur destinée à la détection automatique des risques et non-conformités liés au port des équipements de protection individuelle (EPI) sur les chantiers.

L'objectif est de développer une solution capable d'analyser une image d'une scène de chantier, d'identifier les personnes présentes et d'exploiter les informations relatives au port des équipements de protection afin de déterminer les situations conformes et non conformes.

Le projet s'inscrit dans une démarche de Computer Vision appliquée à la sécurité industrielle, avec une attention particulière portée à la possibilité de déployer le modèle sur un environnement mobile.

Travaux réalisés :

Constitution et préparation d'un dataset de scènes de chantier. Annotation et structuration des données d'entraînement. Nettoyage et préparation des données. Expérimentation de modèles de Deep Learning pour la détection d'objets. Utilisation de YOLO pour la détection des personnes et des équipements de protection. Mise en place d'une logique de décision permettant d'interpréter les détections et d'identifier les situations de non-conformité. Évaluation des performances du modèle. Optimisation du pipeline d'inférence. Conversion du modèle vers TensorFlow Lite. Expérimentation du déploiement du modèle dans une application mobile.

Ce projet porte sur le développement d'un modèle de scoring bancaire destiné à prédire le risque associé à une demande de crédit à partir des données historiques de prêts du dataset Lending Club.

L'objectif est de construire une chaîne complète de Data Science, allant de l'analyse et de la préparation des données jusqu'à la modélisation prédictive et l'évaluation des performances.

Travaux réalisés :

Analyse exploratoire du dataset et étude des différentes variables disponibles. Nettoyage et préparation des données. Traitement des valeurs manquantes et des variables catégorielles. Sélection et transformation des variables pertinentes. Réalisation de Feature Engineering afin de construire des variables adaptées à la prédiction. Préparation des jeux de données pour l'entraînement et l'évaluation. Expérimentation de différents modèles de Machine Learning. Expérimentation d'un modèle de Deep Learning avec TensorFlow Évaluation des modèles à l'aide de métriques de classification. Analyse comparative des performances afin d'identifier l'approche la plus pertinente

Practical tasks on Deep Learning (DL) and Neural Networks (NN). 🤖

This project leverages the Brain Tumor Image database to develop a semantic segmentation model for detecting brain tumors. The model is based on an encoder-decoder U-Net architecture, which classifies each pixel in a medical image as either tumor or healthy tissue. The resulting segmentation is visualized as a binary mask, clearly delineating the tumor region from the surrounding healthy areas.

The model successfully localizes the tumor’s global position with high accuracy but the precise shape and boundaries remain approximate (moderate Dice coefficient). Extending the training phase with additional epochs could refine the segmentation quality, leading to a higher Dice coefficient and more accurate tumor delineation.

A custom convolutional neural network (CNN) was developed from scratch and trained to recognize flower types using pcituires from the Flowers Recognition Dataset. The model’s architecture and hyperparameters were optimized to maximize recognition accuracy.

The custom CNN is compared against MobileNetV2, a pre-trained model fine-tuned using transfer learning. The comparison highlights the trade-offs between custom-built models and transfer learning approaches in terms of accuracy, training efficiency, and computational cost.

Photo-realistic single-image super-resolution (SRGAN, x4) in TensorFlow/TensorLayer — refactored OO training/evaluation pipeline with research experiment branches.

Fundamental theory and practice in Data Science (DS). 🧮

This project focuses on binary sentiment classification of movie reviews from the IMDb Reviews dataset, using three distinct models: a multilayer perceptron and a recurrent neural network (RNN). Each model is designed to classify reviews as positive or negative, leveraging advanced techniques in natural language processing. The word embeddings generated by each model are visualized using dimensionality reduction techniques (PCA, t-SNE), providing an intuitive representation of the semantic space.

Models are assessed using confusion matrices to analyze classification accuracy and ROC curves to evaluate the trade-off between true positive and false positive rates. The project compares the performance of MLP, Word2Vec, and RNN in capturing sentiment from text data.

This project leverages Physics-Informed Neural Networks (PINNs) to solve partial differential equations (PDEs) governing stellar structure. The goal is to train a neural network using four fundamental equations that describe the internal physics of stars. Once trained, the model predicts key structural properties such as core temperature, pressure field, mass, and chemical composition, for well-known stars, including Vega, Betelgeuse, Sirius, and the Sun.

The PINN model is coded using TensorFlow, a framework that allows for building custom architectures, using specialized loss functions and performing custom training with Adam followed by BFGS (from TensorFlow Probability) as the optimizer.

For this project, the idea is to implement a graph neural network (GNN) model with an encoder-processor-decoder architecture and train it using results from mesh-based simulations of a wing in an airflow. The simulations are generated using OpenVSP, a medium-fidelity solver that applies the vortex lattice method to compute the scalar pressure coefficient field across the wing mesh. This field is used to derive aerodynamic curves (lift, drag and moment coefficients) and draw conclusions about the wing profile’s performance. Simulations are automated to cover a wide range of parameters such as Mach number, angle of attack, sweep angle and for different airfoil profile (NACA0012, NACA2412, and NACA23012)

The goal of this project is to compare the results generated by the conventional OpenVSP software with the model’s predictions for new cases not used in training. The limitations of each model as well as the computational resource costs are also studied.

Introduction to classification using machine learning and deep learning (PyTorch, TensorFlow, Keras)

A Python library for Secure and Explainable Machine Learning

Documentation available @ https://secml.gitlab.io

Follow us on Twitter @ https://twitter.com/secml_py

A real-time yoga pose classification web application built with Angular and TensorFlow.js. The app uses your webcam to detect whether you are performing a Downward Dog or Goddess yoga pose, powered by a CNN model trained in Keras and converted for in-browser inference.

Trading Bot – Algorithmic Crypto Trading with AI Integration

This project is a powerful algorithmic trading bot for cryptocurrency markets. It combines traditional technical analysis with modern machine learning to generate accurate and intelligent trading decisions.

Key Features:

Candlestick Pattern Detection: Identifies classic reversal patterns such as Hammer, Doji, Engulfing, Shooting Star, and complex formations like triangle patterns.

Technical Indicators: Includes standard indicators (RSI, MACD, Moving Averages, Bollinger Bands) and advanced tools like Ichimoku Clouds, SuperTrend, Fibonacci Retracements, and more.

Machine Learning Integration: Uses LSTM-based models for time-series forecasting and momentum strategies, combined with indicator signals through weighted evaluation.

Dynamic Signal Weighting: Customizable signal weighting for patterns, indicators, and ML predictions with automatic adjustments to market volatility.

Trade Execution Engine: Supports long/short positions with stop-loss, take-profit, and trailing stop features. Automatically includes fees and tax deductions in profit calculations.

Backtesting & Debugging: Simulates strategies on historical data with detailed equity/value curve visualization and comprehensive debug logs.

Robust Error Handling: Detects and logs data inconsistencies, index errors, and processing issues to ensure stability.

Modular architecture with key components such as TraderBot, SignalHandler, PatternManager, IndicatorManager, MLModelHandler, SequenceManager, DataAPI, and CryptoCurrency. Additional support provided by PatternCalculator, IndicatorCalculator, and DataProcessing.

Version: V1.3.0.0 | GUI: V1.0.0 Author: Marian Seeger – info@seegersoftwaredevelopment.de

web-interface

Old functions to play with neural networks. Developed in December 2024

ML model training

A TensorFlow/Keras neural network for regression on noisy sine wave data, predicting continuous values with real-time visualization of predictions and loss using Matplotlib.