• 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

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

pip install torch torchvision dearpygui

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

git clone https://github.com/yourname/ml-forge
python main.py

Use the templates (file -> templates) to access premade projects:

File -> Templates -> MNIST Classifier

Then press RUN in the toolbar.

• 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)

• 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

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.

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.

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.

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

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

• Sequential models only - skip connections and branching architectures are not supported in v1.0
• No learning rate schedulers in v1.0
• Tab sizing cosmetic bug - first tab (data) appears far larger than other tabs

APACHE