zaina-ml/ml_forge

What it solves

ML Forge is a visual editor for PyTorch pipelines, allowing users to build, train, and run image classification models without writing code. It removes the technical barrier of manual coding for those who want to experiment with neural network architectures and training loops.

How it works

Users interact with a three-tab workflow:

1. Data Prep: Users drag and drop nodes to define datasets (like MNIST or CIFAR) and apply transformations.
2. Model: Users connect layers (such as Conv2D, Linear, and ReLU) via a visual canvas where input/output shapes are automatically calculated.
3. Training: Users wire together data loaders, models, loss functions, and optimizers to execute the training process.

Once configured, the tool generates and runs the PyTorch code internally. Users can monitor loss curves in real-time, save checkpoints, and run inference on trained models, and eventually export the entire pipeline as a standalone train.py PyTorch script.

Who it’s for

It is designed for beginners or researchers who prefer a visual approach to building image classification models rather than writing boilerplate PyTorch code.

Highlights

• Visual Pipeline Editor: Drag-and-drop interface for data preparation, model architecture, and training logic.
• Automated Shape Calculation: Automatically fills in_features and in_channels when connecting layers.
• Live Monitoring: Real-time loss curves and a metrics summary including validation accuracy and fit diagnosis.
• Code Export: Ability to export projects into clean, standalone PyTorch code.
• Built-in Dataset Support: Native support for MNIST, CIFAR-10/100, FashionMNIST, and custom folders via ImageFolder.