TensorBloom

tensorbloom.org · Docs · Download · Roadmap

Build, train, and export PyTorch neural networks visually. Drag-and-drop layers, connect them, train on real datasets with GPU acceleration, and export to ONNX or TorchScript — all without writing boilerplate code.

Why TensorBloom?

• See your architecture — No more debugging invisible dimension mismatches in code. The graph shows every layer, every connection, and every tensor shape.
• Train instantly — Built-in datasets (MNIST, CIFAR, IMDB, and more), one-click training with live loss curves, and GPU auto-detection.
• See what changed — Every run is kept, so you can compare runs side by side and try a trained model on samples it never saw.
• Learn by doing — 16 starter templates from Simple MLP to nanoGPT. Load one, modify it, train it, see what happens.
• Real PyTorch underneath — Every graph generates real nn.Module code. Export your model and use it anywhere.

Features

Graph Editor

• 70+ PyTorch modules: Linear, Conv2d, Transformer, LSTM, BatchNorm, and more
• Drag-and-drop from the layer palette, Enter on a focused layer, or Ctrl+K spotlight search
• Eight layer color families, each with its own icon, so a layer's kind reads without relying on color
• Skip connections, concatenation, multi-branch architectures; skip connections run dotted beside the block they skip
• Group layers into named frames that collapse to a single card, without changing the model
• Reliable undo/redo, Save vs Save As, an unsaved-changes prompt, and auto-save you can restore or discard

Shape Checking

• Real-time shape inference on every edge and layer
• A shape error marks the edge into the layer that can't take it, shows "Gets [1, 512], expects [1, 500]" on the layer, and lists every problem in the toolbar
• One-click fixes such as "Set in_features to 512", from the problems list or the Properties panel

Training

• One-click training (Ctrl+Enter) with a live status in the toolbar: download, epoch progress, and the result
• Live loss/accuracy charts with a hover readout, and a summary card when the run ends
• Plain-language error panel with hints for shape mismatches, out-of-memory, bad labels, missing packages, and more
• Pre-flight fixes are listed for review (Fix and train, Train without changes, or Cancel) instead of being applied silently
• 10+ built-in datasets: MNIST, CIFAR-10/100, TinyShakespeare, WikiText-2, IMDB, AG News, SpeechCommands
• Custom tensor files (.pt, .npz, .safetensors), ImageFolder, HuggingFace datasets, CSV
• GPU acceleration with mixed precision support and VRAM monitoring
• Optimizers: SGD, Adam, AdamW, RMSprop | Schedulers: StepLR, Cosine, OneCycleLR, ReduceLROnPlateau

Runs and Results

• Every run is kept with its curves, settings, and model, saved in the project file
• The Runs tab overlays runs on shared charts and puts their results and settings side by side, highlighting what differs
• Restore brings back a run's graph and settings
• Try it runs a trained model on validation samples it didn't train on, shows its top guesses, filters to mistakes, or tries your own image

Help Built In

• A welcome card with starter models, a Getting Started guide, and a Help menu
• A "?" next to every layer parameter and training setting explains what it does
• Each layer links to its PyTorch documentation

Code Generation & Export

• Live PyTorch code generation — see the equivalent nn.Module as you build
• Import existing PyTorch code into the graph editor
• Export to ONNX (for deployment) or TorchScript (for PyTorch-native use)

Getting Started

Download

Grab the latest installer from the releases page:

• Windows — .exe installer or .msi
• Linux — .deb, .rpm, or .AppImage

Requirement: Python 3.9+ must be installed. The editor opens right away; if PyTorch isn't set up yet, a card on the canvas offers to install it and shows its progress while you keep building.

Building from Source

For development or contributing:

git clone https://github.com/mdsunbeam/tensorbloom.git
cd tensorbloom/app
npm install
npx tauri dev

Requires Node.js 18+, Rust, and Python 3.9+. Windows: also needs C++ Build Tools. Linux: sudo apt install libwebkit2gtk-4.1-dev build-essential curl wget file libssl-dev libayatana-appindicator3-dev librsvg2-dev

Your First Model in 30 Seconds

1. On the empty canvas, pick Simple MLP from the welcome card (or go to Insert > From Template for all 16)

2. Press Train in the toolbar (or Ctrl+Enter) and watch the loss drop and accuracy climb in real time

3. Press Try it in the run summary to see the model's predictions on digits it didn't train on

From here: click any node to edit parameters, swap the dataset, add layers, or try a different template. New to TensorBloom? Help > Getting Started walks through the steps.

Templates

16 pre-built architectures, ready to train:

Category	Templates
Getting Started	Simple MLP
Vision	LeNet-5, ResNet-18, ViT, MobileNetV2, U-Net
Language	Transformer, nanoGPT, BERT Base, LSTM Classifier
Audio	Whisper Encoder, WaveNet Block
Generative	Autoencoder, Conv Autoencoder, Embedding Classifier
Other	DenseNet Block

Go to Insert > From Template to load any template. Each card shows a preview of the model's layout.

Keyboard Shortcuts

Shortcut	Action
Ctrl+Enter	Start training
Ctrl+K	Add any layer (spotlight search)
Ctrl+S / Ctrl+Shift+S	Save / Save As
Ctrl+O	Open project
Ctrl+N	New project
Ctrl+Z / Ctrl+Shift+Z or Ctrl+Y	Undo / Redo
Ctrl+C / Ctrl+V	Copy / Paste nodes
Ctrl+D	Duplicate selected
Ctrl+A	Select all
Ctrl+G / Ctrl+Shift+G	Group selected layers / Ungroup
Ctrl+L	Auto-layout graph
Ctrl+I	Import PyTorch code
Ctrl+E	Export model
Ctrl+B / Ctrl+.	Toggle side panels
Ctrl+M	Toggle minimap
Ctrl+/	Show all shortcuts
Delete	Delete selected
Escape	Close the open dialog or menu
Scroll / Ctrl+Scroll	Pan / Zoom
Double-click	Rename node

Known Limitations

TensorBloom v0.1 focuses on standard feedforward, convolutional, recurrent, and transformer architectures. The following are not yet supported:

• GANs and adversarial training — no multi-model training loops
• Multi-loss training — only one loss node per graph
• Attention masks — Transformer nodes don't accept mask inputs
• Distributed training — single GPU only (multi-GPU DDP planned for v0.4)
• Dynamic architectures — model structure is static at graph definition time

See the Roadmap for planned features.

Documentation

Full docs at tensorbloom.org/docs, or browse locally:

• Getting Started — Build and train your first model
• Custom Data Guide — Load your own datasets
• Code Blocks — Write custom forward() logic
• Training Guide — Training settings, the run flow, comparing runs, and trying a model
• Exporting Models — ONNX, TorchScript, and Python export
• Template Reference — All 16 built-in architectures
• Architecture — Technical internals for contributors
• Roadmap — Planned features and development direction

Building for Production

cd app
npx tauri build

Produces platform-specific installers in app/src-tauri/target/release/bundle/.

Contributing

See CONTRIBUTING.md for development setup and guidelines.

License

MIT