🔬 CNN Optimization Benchmark Platform

A Scientific Research Workstation & Laboratory Platform for Empirically Benchmarking Metaheuristic Optimization Algorithms on Deep CNN Compression

🌐 Visit Live Deployment: cnn.umeshlabs.in

Live Web App • Features • Architecture • Algorithms • Installation • API • Documentation • Author

---

🎯 Core Research Problem

“Which metaheuristic optimization algorithm achieves the superior multi-objective Pareto trade-off between Top-1 Accuracy, Inference Latency, Continuous Power Draw, Computational Complexity (FLOPs/TOPs), Model Footprint, and Energy Consumption under identical hardware and compression constraints?”

The CNN Optimization Benchmark Platform provides an empirical, reproducible, and standardized research workstation to compare 10 state-of-the-art metaheuristics (plus user-registered custom algorithms) across deep Convolutional Neural Networks (ResNet-18, MobileNetV2, ShuffleNetV2, VGG-16, EfficientNet-B0) on standardized vision benchmarks (CIFAR-10, CIFAR-100, MNIST, Fashion-MNIST, ImageNet-1k Subset) and custom dataset uploads.

---

✨ Key Features

• 🎛️ Algorithm Comparison Workbench: Interactive side-by-side comparison with real-time dynamic objective re-weighting (Accuracy, Latency, Size, Energy sliders) and live score re-computation.
• 📐 Scientific Workstation UI/UX: Designed following laboratory instrumentation principles using IBM Plex Sans and IBM Plex Mono typography, dense analytical data tables, and structured research navigation.
• ⚡ Dual Compute Target (GPU & CPU): Full hardware telemetry for NVIDIA GPUs (torch.cuda with pynvml power sampling) and Multi-Core CPUs with high-resolution synchronized latency timings.
• 🔌 Continuous Power & Compute Throughput Profiling: Real-time measurement and analytical modeling of Average Power Draw ($\text{Watts} = \frac{\text{Joules}}{\text{Latency (s)}}$) and Compute Throughput ($\text{TOPs}$).
• 📉 Multi-Objective Pareto Analysis: Automatic extraction and 6-axis interactive visualization of the empirical non-dominated frontier.
• 📈 Convergence & Stochastic Telemetry: Real-time WebSocket iteration tracking and multi-run statistical boxplots (Mean, Median, Std Dev, Min, Max, 95% Student's $t$ Confidence Intervals).
• 🧩 5-Stage Ablation Decomposition: Isolates the marginal contributions of Quantization (FP16/INT8), Pruning (Structured Channel / Filter), and Metaheuristic Optimization.
• 📊 Confusion Matrix & Per-Class Degradation Suite: Interactive $K \times K$ heatmaps (Raw Counts, Row-Normalized Recall %, Column-Normalized Precision %), Top Confusion Pairs ranking, Per-Class Sensitivity Analysis, Macro/Micro Average metrics, and Differential algorithm comparison grids ($\Delta \text{Cell} = A_{i,j} - B_{i,j}$).
• 🚀 1-Click Local Laptop Automated Installer: Automated zero-friction scripts for Windows PowerShell (install.ps1), Command Prompt (install.bat), and macOS/Linux Bash (install.sh) with auto-dependency bootstrapping and local daemon launching.
• 🏷️ Data Provenance Badging: Explicit scientific provenance labeling (● MEASURED, ◆ CALCULATED, ▲ ESTIMATED, DEMO DATA).
• 📁 Custom Extensibility: 1-Click modal uploads for custom Python optimizers (BaseOptimizer), custom image dataset archives (.zip), and custom PyTorch CNN architectures.
• 📄 Publication-Ready Exports: 1-Click downloads in CSV, JSON, Markdown, Microsoft Word (.doc), and Plain Text (.txt) formats.

---

🔬 Benchmarked Metaheuristic Algorithms

All algorithms adhere to the standardized BaseOptimizer mathematical search contract over continuous decision spaces $\mathbf{x} \in [0.0, 1.0]^D$:

Key	Algorithm Name	Family	Citation	Search Complexity
GWO	Grey Wolf Optimizer	Swarm Intelligence	Mirjalili et al. (2014)	$\mathcal{O}(N \times D)$
WOA	Whale Optimization Algorithm	Swarm Intelligence	Mirjalili & Lewis (2016)	$\mathcal{O}(N \times D)$
ALO	Ant Lion Optimizer	Swarm Intelligence	Mirjalili (2015)	$\mathcal{O}(N \times D)$
MFO	Moth-Flame Optimization	Physics / Biology	Mirjalili (2015)	$\mathcal{O}(N \times D)$
GOA	Grasshopper Optimization Algorithm	Swarm Intelligence	Saremi et al. (2017)	$\mathcal{O}(N^2 \times D)$
MVO	Multi-Verse Optimizer	Physics / Cosmology	Mirjalili et al. (2016)	$\mathcal{O}(N \times D)$
SCA	Sine Cosine Algorithm	Mathematical Trigonometric	Mirjalili (2016)	$\mathcal{O}(N \times D)$
AOA	Arithmetic Optimization Algorithm	Mathematical Algebraic	Abualigah et al. (2021)	$\mathcal{O}(N \times D)$
MGO	Mountain Gazelle Optimizer	Swarm Intelligence	Abdollahzadeh et al. (2022)	$\mathcal{O}(N \times D)$
GMO	Geometric Mean Optimizer	Mathematical Geometric	Mirrashid & Naderpour (2023)	$\mathcal{O}(N \times D)$

---

🏗️ System Architecture & Workflow

graph TD
    subgraph ClientLayer [Client Interface - React 19 & TypeScript]
        UI[Dashboard Command Center]
        WB[Algorithm Comparison Workbench]
        Wiz[New Benchmark Wizard]
        Plots[Pareto & Convergence Visualizers]
    end

    subgraph BackendLayer [Backend Engine - FastAPI & Async Workers]
        API[REST Endpoints & WebSocket Broadcaster]
        Runner[Experiment Task Runner]
        Optimizers[10 Metaheuristic Optimizers]
        Eval[Hardware Evaluation Suite - Latency / Power / Accuracy]
        Analytics[Pareto Analysis & WSM Scoring]
    end

    subgraph StorageLayer [Persistence Layer - SQLite & Cloud DB]
        DB[(benchmark.db - Models / Runs / Metrics)]
    end

    UI --> API
    WB --> API
    Wiz --> API
    Plots --> API
    API --> Runner
    Runner --> Optimizers
    Runner --> Eval
    Eval --> Analytics
    Analytics --> DB
    DB --> API

Loading

---

📐 Mathematical Formulation

1. Weighted Sum Model (WSM) Composite Scoring

$$\text{Composite Score}_i = \left( w_{\text{acc}} \cdot \tilde{A}_i + w_{\text{lat}} \cdot \tilde{L}_i + w_{\text{size}} \cdot \tilde{S}_i + w_{\text{energy}} \cdot \tilde{E}_i \right) \times 100$$

Where $\tilde{A}_i, \tilde{L}_i, \tilde{S}_i, \tilde{E}_i \in [0, 1]$ are min-max normalized metrics across all algorithms with inverse transformation for minimization metrics (Latency, Size, Energy).

2. Multi-Objective Fitness Evaluation

$$f(\mathbf{x}) = w_{\text{acc}} \left(\frac{\Delta \text{Acc}(\mathbf{x})}{\text{Acc}_{\text{baseline}}}\right) + w_{\text{lat}} \left(\frac{\text{Lat}(\mathbf{x})}{\text{Lat}_{\text{baseline}}}\right) + w_{\text{size}} \left(\frac{\text{Size}(\mathbf{x})}{\text{Size}_{\text{baseline}}}\right) + w_{\text{energy}} \left(\frac{\text{Energy}(\mathbf{x})}{\text{Energy}_{\text{baseline}}}\right)$$

3. Compute Throughput Density (TOPs)

$$\text{TOPs} = \frac{\text{FLOPs (M)} \times 10^6}{\text{Latency (ms)} \times 10^{-3} \times 10^{12}} = \frac{\text{FLOPs (M)}}{\text{Latency (ms)} \times 10^6}$$

---

🚀 Quick Start & Installation

Prerequisites

• Python 3.10+ (Tested on Python 3.10, 3.11, 3.12, 3.14)
• Node.js v18+ & npm
• Git

1. Clone Repository & Setup Backend

# Clone the repository
git clone https://github.com/UmeshCode1/cnn-optimization-benchmark.git
cd cnn-optimization-benchmark

# Install Python dependencies
python -m pip install -r requirements.txt

# Launch FastAPI Backend Server
python -m uvicorn backend.main:app --host 0.0.0.0 --port 8000 --reload

2. Setup & Launch Frontend Workstation

In a separate terminal window:

cd frontend
npm install
npm run dev

Visit http://localhost:5173 (or http://localhost:8000 for FastAPI production bundle).

---

🧪 Comprehensive Test Suite

Run full automated test verification covering API endpoints, evaluators, and all 10 optimizers:

python -m pytest tests backend/tests -v

======================== 35 passed, 1 warning in 1.50s ========================

---

📚 REST & WebSocket API Catalog

Method	Endpoint	Description
POST	/api/experiments	Create and trigger an asynchronous benchmark experiment
GET	/api/experiments	List all historical benchmark runs with metadata filters
GET	/api/experiments/{id}	Retrieve comprehensive results, statistics, Pareto points, and ablations
POST	/api/experiments/{id}/cancel	Cancel active benchmark execution
DELETE	/api/experiments/{id}	Delete experiment record and cascading runs/metrics
POST	/api/experiments/{id}/clone	Clone benchmark configuration for reproduction
POST	/api/experiments/{id}/recalculate	Live dynamic recalculation of scores and rankings with custom weights
GET	/api/experiments/{id}/confusion-matrix	Retrieve research-grade $K \times K$ confusion matrix & per-class error analytics
GET	/api/reports/{id}/csv	Export results in CSV format
GET	/api/reports/{id}/json	Export results in JSON format
GET	/api/reports/{id}/markdown	Export formatted Markdown research report
GET	/api/reports/{id}/doc	Export Microsoft Word (.doc) formatted report
GET	/api/reports/{id}/txt	Export formatted Plain Text (.txt) report
WS	/api/experiments/{id}/ws	Real-time WebSocket stream for live iteration telemetry
GET	/api/installer/preflight	Pre-flight diagnostics and automated 1-click install instructions
GET	/install.ps1 / /install.sh / /install.bat	Raw automated 1-click laptop installation scripts
GET	/api/algorithms	List all verified and custom metaheuristic optimizers
POST	/api/algorithms	Register a new custom metaheuristic optimizer plugin
GET	/api/datasets	List available datasets and resolution parameters
POST	/api/datasets/upload	Upload a custom image dataset archive
GET	/api/models	List supported CNN model architectures
GET	/api/hardware	Inspect host CPU/GPU hardware profile and telemetry

---

📖 Complete Documentation & GitHub Wiki

• 🏛️ System Architecture & Data Flows
• 🧮 Mathematical Formulations of 10 Metaheuristics
• 📑 GitHub Wiki Master Blueprint
• ⚙️ Benchmarking & Reproducibility Guide
• 🔌 Complete REST & WebSocket API Reference

---

👨‍💻 Author & Maintainer

Umesh Patel

AI & Deep Learning Systems Researcher / Software Engineer

• Website / Platform: https://cnn.umeshlabs.in/
• GitHub: @UmeshCode1
• Repository: https://github.com/UmeshCode1/cnn-optimization-benchmark
• Specialization: Deep Learning Model Compression, Metaheuristic Optimization, Edge AI Hardware Acceleration, and High-Performance Benchmarking.

---

📄 License & Citation

This project is licensed under the MIT License — see the LICENSE file for details.

If you utilize this benchmark workstation in academic or industrial research, please cite:

@software{umesh_patel_cnn_benchmark_2026,
  author = {Umesh Patel},
  title = {CNN Optimization Benchmark: A Scientific Research Platform for Metaheuristics in Deep Convolutional Neural Network Compression},
  year = {2026},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://cnn.umeshlabs.in/}}
}