VisualTorchIDE is a professional-grade, local-first Integrated Development Environment (IDE) for Deep Learning. It allows engineers to design complex PyTorch architectures as visual graphs, validate tensor shapes in real-time, and execute training loops on their own hardware—all through a cinematic, high-performance holographic interface.

Most visual ML tools are "black boxes" for beginners. VisualTorchIDE is different. It is a Visual Compiler.

• It generates production-ready PyTorch code from your graph.
• It catches Runtime Errors (shape mismatches) before you hit train.
• It runs locally, giving you full access to your own CPU/GPU and datasets.

A high-performance 3D canvas built with Three.js and React-Three-Fiber.

• Holographic Nodes: Real-time distorted spheres and geometries representing layers.
• Data Flow Animation: Pulsing yellow packets travel along wires during training to visualize active gradients.
• Pro-Debugger: Hover over connections to see current tensor shapes (e.g., [64, 128, 28, 28]).

• Real-time Metrics: Low-latency tracking of Loss, Accuracy, Precision, and Recall via WebSockets.
• Hardware Integration: Automatic detection and display of CUDA (GPU) or CPU execution status.
• System Logs: A terminal-style feed for granular engine feedback.

Build more than just linear models. VisualTorchIDE supports:

• Vision: Conv2d, MaxPool2d, BatchNorm2d.
• Sequential: LSTM, GRU, RNN.
• Core: Flatten, Dropout, Linear, Lazy-loading.
• Optimizers: Adam, SGD, RMSProp.

The backend engine translates your visual DAG (Directed Acyclic Graph) into idiomatic, clean Python code. You can export the .py file and run it anywhere.

Building this in 24hr solo was a challange of immense magnitude. I have tried my best to give you the best possible outcome I could, however it has its shortcoming and I want to be transparent and honest about it to you.

• It needs a gemini API key in the chatmodel component

• You can make simple model graphs or load the cnn one. Further updates would include more complex ones as well

• The UI UX is not the best that I have ever made, it will surely be mended and beautified soon.

  ---

• Node.js (v18+)
• Python (3.10+)
• Git

git clone https://github.com/yourusername/VisualTorchIDE.git
cd neural-engine

cd backend
pip install -r requirements.txt
# Alternatively:
# py -m pip install fastapi uvicorn torch torchvision pandas websockets sympy

cd ../frontend
npm install

Return to the root and run uvicorn and react server seperately:

cd ..
cd frontend
react start
# In a different terminal
cd backend
uvicorn main:app --reload

The IDE will automatically open in your browser at http://localhost:3000.

Cloud-based ML platforms are slow to upload datasets and often restrictive. VisualTorchIDE stays local:

• Privacy: Your data and weights never leave your machine.
• Speed: Zero latency for large dataset loading.
• Hardware: If you have an NVIDIA GPU, the engine leverages CUDA automatically.
• Offline: Work on your models anywhere, even without internet.

• Generative Architect: Use Gemini/LLMs to generate graphs from text prompts.
• Live Inference: A drawing canvas to test MNIST/CIFAR models in real-time.
• Data Studio: Integrated CSV cleaning and normalization tools.
• ONNX Export: Deploy models directly to web/mobile.

This project was built for the [Your Hackathon Name] Hackathon. Contributions, issues, and feature requests are welcome!

MIT License - Copyright (c) 2026 Dewashish Lambore

Created with ❤️ for the Machine Learning Community.