RecipeNet App v2 - Injection Molding Color Recipe Prediction

Enhanced version of the injection molding color recipe prediction system with an interactive recipe simulator.

Features

1️⃣ Recipe Prediction

• Predict pigment recipes from target color names
• Enhanced model with optimized weights
• Top-K pigment selection for practical recipes
• Excel export functionality

2️⃣ Recipe Simulator

• Interactive recipe adjustment with top-N pigment selection
• Real-time color prediction from custom recipes
• ΔE00 calculation to compare with target colors
• Visual color comparison

Installation

Prerequisites

• Python 3.8 or higher
• Conda (recommended) or virtualenv

Setup

1. Clone this repository:

git clone https://github.com/Saymooon/RecipeNet-App-v2.git
cd RecipeNet-App-v2

2. Create and activate a conda environment:

conda create -n ColorMatching python=3.8
conda activate ColorMatching

3. Install required packages:

pip install -r requirements.txt

Usage

Run the Streamlit app:

streamlit run app_v2.py

The app will open in your default web browser at http://localhost:8501

Required Files

Make sure all the following files are in the same directory:

1. app_v2.py - Main application file
2. recipe_model_optimized_weight_0.1.pth - Optimized RecipeNet model
3. name_encoder.pkl - Text encoder for color names
4. xgb_surrogate_2.pkl - XGBoost model for recipe prediction
5. xgb_surrogate_3.pkl - XGBoost model for recipe simulator
6. swatch_recipe_merged_1120.csv - Reference dataset
7. requirements.txt - Python dependencies

How to Use

Recipe Prediction

1. Select a target color from the dropdown (56 colors available)
2. Click "레시피 예측 실행" (Run Recipe Prediction)
3. View the predicted recipe and download as Excel if needed

Recipe Simulator

1. After predicting a recipe, scroll down to the simulator section
2. Click "🔄 예측 레시피 불러오기" to load the predicted recipe
3. Adjust pigment amounts using the sliders
4. Click "🎨 색상 예측 실행" to see the predicted color
5. Compare the result with the target color using ΔE00 metric

Technical Details

• Model Architecture: RecipeNet with 3-head attention mechanism
• Optimization: Model weight = 0.7, Similar recipe weight = 0.3
• Top-K Selection: Default 5 pigments for practical manufacturing
• Color Space: CIE Lab color space
• Color Difference: CIEDE2000 (ΔE00) metric

License

This project is for research and educational purposes.

Contact

For questions or issues, please open an issue on GitHub.