Ajouter la documentation pour le projet "Dropout Reduces Underfitting" avec une implémentation TensorFlow/Keras et des objectifs scientifiques

2026-01-14 14:44:21 +01:00 · 2025-12-09 21:18:47 +01:00
parent f5ed4b109b
commit 043872a33a
2 changed files with 374 additions and 0 deletions
--- a/content/projects/artmcp.md
+++ b/content/projects/artmcp.md
@@ -0,0 +1,220 @@
+---
+slug: artmcp
+title: 🤖 ArtMcp
+description: A comprehensive Model Context Protocol (MCP) server exposing professional profile information about Arthur Danjou.
+publishedAt: 2025/10/27
+readingTime: 3
+favorite: true
+tags:
+  - web
+  - nuxt
+  - mcp
+---
+
+🤖 [ArtMcp](https://github.com/arthurdanjou/artmcp) - Arthur Danjou's MCP Server
+
+A comprehensive [Model Context Protocol (MCP)](https://modelcontextprotocol.io) server exposing professional profile information about Arthur Danjou. Built with [Nuxt](https://nuxt.com) and deployed on [NuxtHub](https://hub.nuxt.com) at the Edge.
+
+🔗 **Live Server**: https://mcp.arthurdanjou.fr
+
+## 🎯 Features
+
+### MCP Resources
+
+The server exposes the following resources through the Model Context Protocol:
+
+- **📊 Skills** (`resource://artmcp/skills`) - Complete list of technical skills (programming languages, frameworks, tools)
+- **💼 Experiences** (`resource://artmcp/experiences`) - Professional work experience and projects
+- **🚀 Projects** (`resource://artmcp/projects`) - Portfolio of personal and professional projects
+- **🎓 Education** (`resource://artmcp/education`) - Academic background and degrees
+- **🌐 Languages** (`resource://artmcp/languages`) - Spoken languages with proficiency levels
+- **👤 Profile** (`resource://artmcp/profile`) - Comprehensive profile with bio, location, availability, career goals, and work preferences
+- **🎨 Hobbies** (`resource://artmcp/hobbies`) - Personal interests and activities
+- **📞 Contact** (`resource://artmcp/contact`) - Professional contact information and social links
+- **🛠️ Uses** (`resource://artmcp/uses`) - Tools, hardware, and software setup
+
+### MCP Tools
+
+- **`activity`** - Real-time current activity and status of Arthur Danjou
+- **`resume-link`** - Get download link for resume in English or French
+- **`stats`** - Detailed coding statistics and analytics from WakaTime
+- **`status-page`** - Real-time status and uptime monitoring for homelab infrastructure
+- **`uses-by-category`** - Filter uses by category (homelab, ide, hardware, software)
+- **`weather`** - Get current weather for a city
+
+### MCP Prompts
+
+Pre-configured prompts for common queries about:
+- Resume generation
+- Skills and expertise
+- Projects showcase
+- Real-time activity
+- Contact information
+- And more...
+
+## 🏗️ Architecture
+
+This project uses:
+- **Nuxt 4** with Nitro for server-side rendering
+- **@nuxt/content** for content management
+- **@nuxtjs/mcp-toolkit** for MCP server implementation
+- **NuxtHub** for edge deployment on Cloudflare Workers
+- **nuxt-studio** for content management studio
+- **Zod** for schema validation
+
+## 🚀 Getting Started
+
+### Prerequisites
+
+- Node.js 18+ or Bun
+- pnpm 10.12.1+
+
+### Installation
+
+```bash
+# Install dependencies
+pnpm install
+```
+
+### Environment Variables
+
+Create a `.env` file (optional):
+
+```bash
+# Discord integration (optional)
+NUXT_DISCORD_USER_ID=""
+NUXT_DISCORD_ID=""
+NUXT_DISCORD_TOKEN=""
+
+# Wakatime integration (optional)
+NUXT_WAKATIME_USER_ID=""
+NUXT_WAKATIME_CODING=""
+NUXT_WAKATIME_EDITORS=""
+NUXT_WAKATIME_LANGUAGES=""
+NUXT_WAKATIME_OS=""
+
+# Status page (optional)
+NUXT_STATUS_PAGE=""
+```
+
+### Development
+
+Start the development server on `http://localhost:3000`:
+
+```bash
+pnpm dev
+```
+
+### Production
+
+Build the application for production:
+
+```bash
+pnpm build
+```
+
+### Deployment
+
+Deploy to NuxtHub/Cloudflare:
+
+```bash
+pnpm deploy
+```
+
+## 📚 API Endpoints
+
+All resources are also available as REST API endpoints:
+
+- `GET /api/skills`
+- `GET /api/experiences`
+- `GET /api/projects`
+- `GET /api/education`
+- `GET /api/languages`
+- `GET /api/profile`
+- `GET /api/hobbies`
+- `GET /api/contact`
+- `GET /api/uses`
+- `GET /api/activity`
+- `GET /api/wakatime`
+- `GET /api/status-page`
+- `GET /api/resumes/{en|fr}` - Download resume
+
+## 🧪 Development
+
+### Linting
+
+```bash
+pnpm lint
+```
+
+### Type Checking
+
+```bash
+npx tsc --noEmit --skipLibCheck
+```
+
+## 📂 Content Structure
+
+Content is managed in the `content/` directory:
+
+```
+content/
+├── skills.json              # Technical skills
+├── languages.json           # Spoken languages
+├── profile.md               # Comprehensive profile info
+├── contact.json             # Contact information
+├── hobbies.md               # Personal interests
+├── documentation.md         # MCP documentation
+├── experiences/*.md         # Work experiences
+├── projects/*.md            # Project portfolio
+├── education/*.md           # Academic background
+└── uses/*.md                # Tools and setup
+```
+
+## 🔧 Technologies
+
+- **Frontend/Backend**: Nuxt 4, Vue 3, Nitro
+- **MCP**: @nuxtjs/mcp-toolkit
+- **Content**: Nuxt Content with better-sqlite3
+- **Content Studio**: nuxt-studio
+- **Deployment**: Cloudflare Workers via NuxtHub
+- **Validation**: Zod schemas
+
+## 📖 MCP Integration
+
+To use this server with an MCP client:
+
+1. Configure your MCP client to connect to `https://mcp.arthurdanjou.fr/mcp`
+2. Or use the API directly via REST endpoints
+
+Example MCP client configuration:
+```json
+{
+  "mcpServers": {
+    "artmcp": {
+      "url": "https://mcp.arthurdanjou.fr/mcp"
+    }
+  }
+}
+```
+
+## 🤝 Contributing
+
+This is a personal portfolio project. Feel free to use it as inspiration for your own MCP server!
+
+## 📝 License
+
+Private project - All rights reserved
+
+## 👤 About
+
+**Arthur Danjou**
+- Data Science & Applied AI student at Paris Dauphine-PSL University, passionate about machine learning and mathematical modelling
+- 📍 Paris, France
+- 🔗 [LinkedIn](https://go.arthurdanjou.fr/linkedin)
+- 🐙 [GitHub](https://go.arthurdanjou.fr/github)
+- 📧 [Email](https://go.arthurdanjou.fr/mail-pro)
+
+---
+
+Built with ❤️ using Nuxt and the Model Context Protocol
--- a/content/projects/dropout-reduces-underfitting.md
+++ b/content/projects/dropout-reduces-underfitting.md
@@ -0,0 +1,154 @@
+---
+slug: dropout-reduces-underfitting
+title: 🔬 Dropout reduces underfitting
+description: TensorFlow/Keras implementation of "Dropout Reduces Underfitting" (Liu et al., 2023). A comparative study of Early and Late Dropout strategies to optimize model convergence. 
+publishedAt: 2054/12/10
+readingTime: 4
+favorite: false
+tags:
+  - python
+  - reserch
+  - machine-learning
+  - tensorflow
+---
+
+📉 [Dropout Reduces Underfitting](https://github.com/arthurdanjou/dropoutreducesunderfitting): Reproduction & Analysis
+
+![TensorFlow](https://img.shields.io/badge/TensorFlow-2.x-orange.svg)
+![Python](https://img.shields.io/badge/Python-3.8%2B-blue.svg)
+![License](https://img.shields.io/badge/License-MIT-green.svg)
+
+> **Study and reproduction of the paper:** Liu, Z., et al. (2023). *Dropout Reduces Underfitting*. arXiv:2303.01500.
+
+The paper is available at: [https://arxiv.org/abs/2303.01500](https://arxiv.org/abs/2303.01500)
+
+This repository contains a robust and modular implementation in **TensorFlow/Keras** of **Early Dropout** and **Late Dropout** strategies. The goal is to verify the hypothesis that dropout, traditionally used to reduce overfitting, can also combat underfitting when applied solely during the initial training phase.
+
+## 🎯 Scientific Objectives
+
+The study aims to validate the three operating regimes of Dropout described in the paper:
+
+1.  **Early Dropout** (Targeting Underfitting): Active only during the initial phase to reduce gradient variance and align their direction, allowing for better final optimization.
+2.  **Late Dropout** (Targeting Overfitting): Disabled at the start to allow rapid learning, then activated to regularize final convergence.
+3.  **Standard Dropout**: Constant rate throughout training (Baseline).
+4.  **No Dropout**: Control experiment without dropout.
+
+## 🛠️ Technical Architecture
+
+Unlike naive Keras callback implementations, this project uses a **dynamic approach via the TensorFlow graph** to ensure the dropout rate is properly updated on the GPU without model recompilation.
+
+### Key Components
+
+* **`DynamicDropout`**: A custom layer inheriting from `keras.layers.Layer` that reads its rate from a shared `tf.Variable`.
+* **`DropoutScheduler`**: A Keras `Callback` that drives the rate variable based on the current epoch and the chosen strategy (`early`, `late`, `standard`).
+* **`ExperimentPipeline`**: An orchestrator class that handles data loading (MNIST, CIFAR-10, Fashion MNIST), model creation (Dense or CNN), and execution of comparative benchmarks.
+
+## File Structure
+
+```
+.
+├── README.md                         # This documentation file
+├── Dropout reduces underfitting.pdf  # Original research paper
+├── pipeline.py                       # Main experiment pipeline
+├── pipeline.ipynb                    # Jupyter notebook for experiments
+├── pipeline_mnist.ipynb              # Jupyter notebook for MNIST experiments
+├── pipeline_cifar10.ipynb            # Jupyter notebook for CIFAR-10 experiments
+├── pipeline_cifar100.ipynb           # Jupyter notebook for CIFAR-100 experiments
+├── pipeline_fashion_mnist.ipynb      # Jupyter notebook for Fashion MNIST experiments
+├── requirements.txt                  # Python dependencies
+├── .python-version                   # Python version specification
+└── uv.lock                           # Dependency lock file
+```
+
+## 🚀 Installation
+
+```bash
+# Clone the repository
+git clone https://github.com/arthurdanjou/dropoutreducesunderfitting.git
+cd dropoutreducesunderfitting
+```
+
+##  Install dependencies
+```bash
+pip install tensorflow numpy matplotlib seaborn scikit-learn
+```
+
+## 📊 Usage
+
+The main notebook pipeline.ipynb contains all necessary code. Here is how to run a typical experiment via the pipeline API.
+1. Initialization
+
+Choose your dataset (cifar10, fashion_mnist, mnist) and architecture (cnn, dense).
+```python
+from pipeline import ExperimentPipeline
+
+# Fashion MNIST is recommended to observe underfitting/overfitting nuances
+exp = ExperimentPipeline(dataset_name="fashion_mnist", model_type="cnn")
+```
+
+2. Learning Curves Comparison
+
+Compare training dynamics (Loss & Accuracy) of the three strategies.
+
+```python
+exp.compare_learning_curves(
+    modes=["standard", "early", "late"],
+    switch_epoch=10,  # The epoch where dropout state changes
+    rate=0.4,         # Dropout rate
+    epochs=30
+)
+```
+
+3. Ablation Studies
+
+Study the impact of the "Early" phase duration or Dropout intensity.
+
+```python
+# Impact of the switch epoch on final performance
+exp.compare_switch_epochs(
+    switch_epochs=[5, 10, 15, 20],
+    modes=["early"],
+    rate=0.4,
+    epochs=30
+)
+
+# Impact of the dropout rate
+exp.compare_drop_rates(
+    rates=[0.2, 0.4, 0.6],
+    modes=["standard", "early"],
+    switch_epoch=10,
+    epochs=25
+)
+```
+
+4. Data Regimes (Data Scarcity)
+
+Verify the paper's hypothesis that Early Dropout shines on large datasets (or limited models) while Standard Dropout protects small datasets.
+
+```python
+# Training on 10%, 50% and 100% of the dataset
+exp.run_dataset_size_comparison(
+    fractions=[0.1, 0.5, 1.0],
+    modes=["standard", "early"],
+    rate=0.3,
+    switch_epoch=10
+)
+```
+
+## 📈 Expected Results
+
+According to the paper, you should observe:
+
+- Early Dropout: Higher initial Loss, followed by a sharp drop after the switch_epoch, often reaching a lower minimum than Standard Dropout (reduction of underfitting).
+- Late Dropout: Rapid rise in accuracy at the start (potential overfitting), then stabilized by the activation of dropout.
+
+## 📝 Authors
+
+- [Arthur Danjou](https://github.com/ArthurDanjou)
+- [Alexis Mathieu](https://github.com/Alex6535)
+- [Axelle Meric](https://github.com/AxelleMeric)
+- [Philippine Quellec](https://github.com/Philippine35890)
+- [Moritz Von Siemens](https://github.com/MoritzSiem)
+M.Sc. Statistical and Financial Engineering (ISF) - Data Science Track at Université Paris-Dauphine PSL
+
+Based on the work of Liu, Z., et al. (2023). Dropout Reduces Underfitting.