MLP Visualizer v2.0.0

MLP Visualizer v2.0.0 - Major Update 🚀

🎉 What's New

This release brings significant performance improvements, new features, and architectural changes that make the application more powerful and user-friendly.

⚡ Major Improvements

🔢 Full NumPy Migration

• Complete rewrite of all matrix operations to use NumPy
• 10-100x performance boost in training and visualization
• Vectorized computations throughout the codebase
• Removed all manual loops for matrix multiplications

🎨 Enhanced Autoencoder System

• Two-stage training workflow: Train Autoencoder → Save Encoder → Train MLP
• Encoder Save/Load: Export and import trained encoders for reuse
• Improved reconstruction visualization: 4×5 grid showing all 10 MNIST digits
• Feature expansion strategy: Project 2D data to higher dimensions for better separation
• Deep encoder architecture: 784→128 for MNIST (better feature extraction)

✍️ MNIST Handwriting Tester

• Interactive digit drawing: Draw your own digits and test the model
• Real-time prediction: Instant classification with confidence scores
• MNIST-style preprocessing: Automatic centering, resizing, and normalization
• 280×280 canvas scaled to 28×28 for accurate testing

🆕 New Features

💾 Model Persistence

• Save and load trained models (.pkl format)
• Separate encoder save/load for Autoencoder workflows
• Auto-creates weights/ directory
• Exe-compatible paths: Works in both script and compiled modes

🎯 Improved Training Control

• Min Error stopping criterion as default for all models
• Enhanced stopping criteria visibility and control
• Better status messages during training phases
• Stage completion notifications with actionable next steps

🎨 UI/UX Enhancements

• Context-aware button visibility: Only show relevant controls
• No more black screen: Smooth initialization with proper plot rendering
• Dynamic tab switching: Autoencoder Error, MLP Error, Reconstruction tabs
• Test Split auto-hide: Hidden in MNIST mode (separate test set exists)

🐛 Bug Fixes

Critical Fixes

• ✅ Stop Training button now works immediately (callback-based interruption)
• ✅ Fixed handwriting prediction accuracy (proper MNIST preprocessing)
• ✅ Resolved TabView crash when switching between models
• ✅ Fixed reconstruction visualization (NumPy array boolean checks)

Training Fixes

• ✅ Fixed momentum implementation (normalized update rule)
• ✅ Corrected Perceptron batch size errors
• ✅ Fixed attribute access errors in training pipeline

Visualization Fixes

• ✅ Eliminated initial black screen on MNIST load
• ✅ Fixed decision boundary rendering lag
• ✅ Proper train/test data split visualization

🔧 Configuration Updates

Optimized Defaults

• MNIST Architecture: 784→256→10 (upgraded from 784→128→10)
• Activation Functions: ReLU for hidden layers (faster convergence)
• Stopping Criteria: Min Error (0.01) as global default
• Autoencoder Config:
  • MNIST: Deep encoder 784→128
  • Manual: Feature expansion 2→16

🛠️ Technical Improvements

• Refactored Architecture: Extracted training logic to TrainingManager
• Code Quality: Removed legacy comments, improved modularity
• Error Handling: Comprehensive exception catching and user notifications
• Type Hints: Added throughout critical paths

📝 Breaking Changes

Warning

Autoencoder Workflow Change

• Stage 1 (Autoencoder) and Stage 2 (MLP) now require manual restart
• This allows encoder saving between stages
• Press "START TRAINING" twice: once for encoder, once for MLP