🚀 NeuralForge AI: YOLO & Cluster MLOps Suite

The official enterprise IDE extension for the train_service2 ecosystem. Designed and architected by William Rodriguez (Wisrovi), this extension bridges the gap between modern code editors (VSCode, Google Antigravity) and enterprise distributed training infrastructure. It delivers end-to-end orchestration for next-generation YOLO26/v8 model training, Optuna Sweeper v2 hyperparameter tuning, Explainable AI (XAI) heatmaps inspection, real-time Kubernetes & Celery cluster telemetry, and selective MLflow artifact retrieval.

🏛️ Architecture & Data Flow

flowchart LR
    subgraph IDE_Workspace [VSCode / Antigravity IDE]
        EXT[NeuralForge Extension]
        XAI[XAI Heatmaps Webview]
        WIZ[YOLO26 Sweeper Wizard]
        DASH[Celery Mission Control]
    end

    subgraph Cluster_Control [Control & Orchestration Layer]
        K8S_API[K8s Cluster Admin API :30001]
        CTRL_SRV[Control Server :23442]
        REDIS[(Redis Broker :23437)]
    end

    subgraph Compute_Grid [Distributed Compute Grid]
        INV1[Worker 192.168.1.40 - RTX 4090]
        INV2[Worker 192.168.1.68 - RTX 3090]
        INV3[Worker 192.168.1.72 - RTX 3090]
    end

    subgraph Storage_Artifacts [Artifacts & Storage Layer]
        MLFLOW[(MLflow Server :5000)]
        MINIO[(MinIO S3 Buckets)]
        CIFS[(Samba CIFS Datasets)]
    end

    EXT -->|Query Telemetry| CTRL_SRV
    EXT -->|Recommend Node & Estimate| K8S_API
    CTRL_SRV -->|Queue Jobs| REDIS
    REDIS -->|Dispatch Tasks| Compute_Grid
    Compute_Grid -->|Mount Datasets| CIFS
    Compute_Grid -->|Log XAI, Weights & Reports| MLFLOW
    Compute_Grid -->|Persist Checkpoints| MINIO
    MLFLOW -.->|Stream Heatmaps & Artifacts| XAI
    CTRL_SRV -.->|Telemetry Stream| DASH

🛠️ Key Technologies and Libraries

The NeuralForge extension is built upon modern IDE, networking, and visualization technologies:

VSCode Extension API (^1.80.0): Native editor integration providing custom Activity Bar views, hierarchical TreeDataProviders, bidirectional Webview message passing, dynamic CodeLens, and configuration schemas.
TypeScript (^6.0.0): Strictly-typed codebase delivering strict null checks, clean interfaces, and robust compile-time guarantees across all telemetry providers.
Webpack 5: Industrial-grade bundling, dead-code elimination, and tree-shaking that compiles all providers and styles into an optimized, self-contained dist/extension.js production artifact.
Node.js Asynchronous HTTP Client: Non-blocking asynchronous network communication interfacing with both the Kubernetes Cluster Admin NodePort (:30001) and Celery Control Server REST endpoints (:23442).
HTML5 & CSS3 Responsive Webviews: Hardware-accelerated, theme-adaptive user interfaces adhering to native VSCode design variables (--vscode-*) for the XAI Heatmaps Explorer, Configuration Wizard, and Mission Control Dashboard.
JSON Schema (Draft-07): Schema specification providing automated autocomplete, inline documentation tooltips, and strict linting for YOLO26 training descriptors and Optuna Sweeper v2 configurations.

🚀 Core Feature Deep Dive

1. 🔬 XAI Heatmaps & Explainability Explorer (`neuralforge.viewHeatmaps`)

Modern computer vision demands interpretability. When training finishes, the worker executes a 20-step scientific analysis. The extension provides an inline interactive Webview to:

Visualize Visual Activations: Inspect Eigen-CAM and Grad-CAM deep-layer feature activation maps overlaying detected objects from extras/model_focus/.
Evaluate Deletion AUC Scores: Verify fidelity metrics (Area Under the Deletion Curve) ensuring the model focuses on salient semantic features rather than background noise.
Inspect Adversarial Robustness: Review accuracy degradation under FGSM and PGD gradient-based adversarial perturbations.
Direct Actions: One-click shortcuts to download final model weights (best.pt) or open the descriptive post-train executive report.

2. ⚡ Live Cluster Telemetry & Celery Mission Control

Say goodbye to black-box distributed computing. The extension connects directly to live cluster APIs:

Cluster TreeView (neuralforgeCluster): Displays real-time status of the central Manager node (NAS_Train_service), individual GPU invoker nodes (e.g. worker-cima0065d (192.168.1.68) [Active GPU Worker]), and active/queued Celery tasks.
Celery Mission Control Dashboard (neuralforge.global.celeryDashboard): High-level view showing Redis broker health, total connected clients, GPU worker queue distribution (gpus_high, gpus_medium, wyolo_gpu_queue), and active study progress.

3. 🎯 Next-Gen YOLO26 & Optuna Sweeper v2 Wizard (`neuralforge.createConfigWizard`)

An intuitive, structured wizard that creates production-ready YAML descriptors:

Architectures: Supports YOLO26 (yolo26n.pt, yolo26s.pt, yolo26m.pt, yolo26l.pt), YOLOv8, and RT-DETR.
Multi-Task CV Tasks: Object Detection, Instance Segmentation, Image Classification, Oriented Bounding Boxes (OBB), and Pose / Keypoints.
Hyperparameter Sweeper (Optuna): Configurable samplers (TPE or CMA-ES), direction, trial counts, and automatic metric mapping (e.g. metrics/mAP50-95(B), metrics/mAP50-95(M), metrics/mAP50-95(P)).
Scheduling & Private Allocation: Target specific worker IPs (target_worker) and schedule off-peak execution (start_at: "22:00").
Post-Train Module Toggles: Granular control to enable or disable XAI, Adversarial testing, and PyLaTeX scientific report generation.

4. 📦 Selective MLflow Artifacts Downloader (`neuralforge.downloadSelectiveArtifact`)

Avoid downloading multi-gigabyte ZIP archives containing checkpoints for dozens of unpromising trials. Surgically download only the critical assets:

Best Checkpoint: weights/best.pt directly into your workspace.
Executive Reports: report/LLM_Report.docx or GLOBAL_RESEARCH_EXPLANATION.md.
Heatmap Gallery: extras/model_focus/ CAM images for presentations and paper figures.
LaTeX Tables: latex/table_results.tex generated via PyLaTeX for scientific publication.

5. 🤖 Node Recommender & Training Duration Estimator

Recommend Node (neuralforge.recommendNode): Queries the Kubernetes scheduler algorithm to identify the optimal worker node based on minimum VRAM requirements and current temperature.
Estimate Duration (neuralforge.estimateDuration): Computes reliable training duration estimates before submitting jobs.

⌨️ Command Palette Reference

All features are registered under the NeuralForge category in the Command Palette (Ctrl+Shift+P / Cmd+Shift+P):

Command ID	Title	Description
`neuralforge.createConfigWizard`	Create Training Config (Wizard)	Opens the interactive Webview to generate YOLO26/v8 Sweeper YAMLs.
`neuralforge.viewHeatmaps`	View XAI Heatmaps & Explainability	Launches the XAI Explorer Webview for Eigen-CAM images and Deletion AUC.
`neuralforge.viewClusterStatus`	Refresh Cluster Telemetry & Node Health	Forces an immediate refresh of the cluster nodes and queues TreeView.
`neuralforge.global.celeryDashboard`	Open Celery Mission Control Dashboard	Displays live cluster overview, broker connections, and active tasks.
`neuralforge.recommendNode`	Query Best Compute Node for VRAM & Load	Recommends optimal GPU worker node based on memory headroom.
`neuralforge.estimateDuration`	Estimate Study Training Duration	Projects training completion time in hours and minutes.
`neuralforge.downloadSelectiveArtifact`	Selectively Download Specific MLflow Artifact	Surgically retrieves `best.pt`, reports, CAMs, or LaTeX tables.
`neuralforge.downloadTrialArtifacts`	Download Complete Study Artifacts (ZIP)	Bundles all trials and checkpoints into a single archive from MLflow.
`neuralforge.launchTraining`	Launch Distributed YOLO Training Study	Submits the active configuration YAML to the cluster via MCP.
`neuralforge.global.viewReport`	View Post-Train MLflow Executive Report	Opens the descriptive post-train LLM analysis report.
`neuralforge.local.viewLogs`	Stream Invoker Worker Execution Logs	Tails live execution logs from worker executor containers.
`neuralforge.local.runTest`	Run End-to-End Invoker Integration Test	Builds local executor and runs validation test.
`neuralforge.global.wpipeScaffold`	Deploy WPipe Modular Pipeline Scaffolding	Scaffolds standard WPipe DAG pipelines.
`neuralforge.local.previewEDA`	Preview Dataset Annotations (Mini-EDA)	Previews dataset bounding boxes and class distribution.
`neuralforge.global.openCodeChat`	Launch OpenCode AI Assistant Chat	Opens interactive AI assistant for training analysis.

⚙️ Configuration & Settings

Configure the extension via Settings > Extensions > NeuralForge AI or in your settings.json:

{
  "neuralforge.controlHost": "192.168.10.252",
  "neuralforge.controlPort": 23442,
  "neuralforge.k8sAdminPort": 30001,
  "neuralforge.mlflowPort": 5000,
  "neuralforge.defaultModel": "yolo26n.pt"
}

neuralforge.controlHost: Host IP or hostname of the Central Manager / Control Server.
neuralforge.controlPort: REST API port for Celery queue status and node registrations (default: 23442).
neuralforge.k8sAdminPort: Kubernetes NodePort for the Cluster Admin API service (default: 30001).
neuralforge.mlflowPort: Port of the MLflow Tracking Server for artifacts and metrics (default: 5000).
neuralforge.defaultModel: Default base model selected when generating new configuration templates.

🧭 End-to-End Workflow Guide

Verify Cluster Health: Open the NeuralForge Activity Bar tab to view active nodes under Cluster Infrastructure & Nodes.
Generate Configuration: Execute NeuralForge: Open YOLO26 / Optuna Sweeper Configuration Wizard, select your dataset path, architecture, and post-train explainability options, then click Generate NeuralForge YAML.
Inspect & Validate: Open the generated .yaml file. The built-in JSON schema automatically lints all fields and highlights parameter errors.
Launch Training: Click the CodeLens action 🚀 Launch Training via MCP at the top of the YAML file.
Monitor Progress: Open the Celery Mission Control Dashboard to watch epochs advance across the assigned GPU invokers.
Analyze Interpretability: When training finishes, run NeuralForge (Global): Open XAI Heatmaps & Explainability Explorer to review Eigen-CAM heatmaps and Deletion AUC fidelity.
Export Checkpoint: Use NeuralForge (Global): Selectively Download Specific MLflow Artifact to pull the optimized best.pt directly into your deployment repository.

📦 Building and Packaging

To compile from source and package into a .vsix file:

# 1. Install dependencies
npm install

# 2. Compile production bundle with Webpack
npm run --prefix . package

# 3. Package into .vsix archive
npx @vscode/vsce package --no-dependencies

# 4. Install in VSCode or Antigravity IDE
code --install-extension neuralforge-0.0.3.vsix --force
antigravity --install-extension neuralforge-0.0.3.vsix --force

📜 License & Commercial Terms

This project is part of the train_service2 ecosystem and operates under a dual-licensing model:

Open Source: AGPLv3 (see LICENSE). Any derivative work or hosted service modifying this software must disclose source code under AGPLv3.
Commercial: Enterprise deployments, closed-source integrations, and commercial licensing require an official commercial license. See COMMERCIAL.md or contact the author.

👤 Author & Official Affiliation

William Steve Rodriguez Villamizar (Wisrovi)
Role: Principal AI Engineer & Applied AI Solutions Architect | Scientific Researcher
📧 Email: wisrovi.rodriguez@gmail.com
🌐 Official Portal: wisrovi.dev
💼 LinkedIn: wisrovi-rodriguez
🆔 ORCID: 0009-0005-0710-1861
📦 PyPI: pypi.org/user/wisrovi/
🐙 GitHub: @wisrovi

NeuralForge AI: YOLO & Cluster MLOps Suite

wirovi