ML Forge

A visual PyTorch pipeline editor. Build, train and run image classification models without writing code.

MLForge v1.0.3

Additions:

• Fixed missing package prefix on import in _build_criterion_and_optimizer
• Preserved plot data in last_run on training completion so Metrics window works post-run
• Implemented update_block_labels/reset_block_labels stubs with live epoch stats
• Fixed GaussianBlur sigma range string parsing crashing at runtime
• Derived _DATASET/_AUG/_LOSS/_OPTIMIZER_BLOCKS from blocks.SECTIONS

What it does

• Visual pipeline - drag nodes onto a canvas, connect them with wires, and ML Forge generates and runs the training code for you
• Three-tab workflow - Data Prep -> Model -> Training
• Live training - watch loss curves update in real time, save checkpoints, run inference on your trained model
• Export - export projects into clean PyTorch

Tutorial

Watch the MLForge tutorial video here: MLForge Tutorial

---

Requirements

IMPORTANT: PyTorch must be preinstalled for training, it is not installed as a dependency.

• Python 3.10 or newer
• PyTorch 2.0 or newer
• torchvision

pip install torch torchvision

GPU training is automatic if CUDA is available. CPU and Apple MPS are also supported.

---

Quick Start

To install MLForge, enter the following in your command prompt

pip install zaina-ml-forge

Then

ml-forge

---

Cloning

git clone https://github.com/zaina-ml/ml-forge
python -m ml_forge

---

Building your first model

1. Data Prep tab

• Add a Dataset node (MNIST, CIFAR10, CIFAR100, FashionMNIST, or ImageFolder)
• Chain transforms: ToTensor is required, add Normalize for best results
• End with a DataLoader (train) node
• For proper validation, add a second chain (same dataset with train=False) ending with DataLoader (val)

2. Model tab

• Start with an Input node - shape is auto-filled from your dataset
• Add layers: Linear, Conv2D, ReLU, BatchNorm2D, Flatten, Dropout, etc.
• End with an Output node - num classes is auto-filled from your dataset
• Connect nodes by dragging from an output pin to an input pin
• in_features and in_channels auto-fill when you connect layers
• After a Flatten node, the next Linear's in_features is calculated automatically

3. Training tab

Add these four nodes from the palette and wire them up:

DataLoaderBlock.images  ->  ModelBlock.images
ModelBlock.predictions  ->  Loss.pred
DataLoaderBlock.labels  ->  Loss.target
Loss.loss               ->  Optimizer.params

Configure epochs, device, checkpointing and early stopping in the right panel, then press RUN.

---

Keyboard shortcuts

Key	Action
Del / Ctrl-Backspace	Delete selected nodes
Ctrl+S	Save project
Ctrl+Z	Undo
Ctrl+Y	Redo
Middle-drag	Pan the canvas

---

Supported datasets

Dataset	Classes	Input shape
MNIST	10	1 × 28 × 28
FashionMNIST	10	1 × 28 × 28
CIFAR-10	10	3 × 32 × 32
CIFAR-100	100	3 × 32 × 32
ImageFolder	custom	3 × 224 × 224

---

Inference

After training, open Run -> Inference, browse to your checkpoint (.pth), and click Run Inference to sample from the test set and see top-k predictions.

---

Metrics

Click the METRICS button to see a summary of your training run: final loss, best validation accuracy, fit diagnosis, and loss/accuracy curves, you may also see the curves on the right training panel.

---

Saving and loading

Projects are saved as .mlf files (JSON). Use File -> Save / Save As or Ctrl+S.

---

Exporting code

File -> Export -> Python -> PyTorch generates a standalone train.py that reproduces your pipeline. No ML Forge required to run it.

---

Future Plans

• Orient to a more beginner-friendly approach
• Add custom block functionality

---

License

MIT