zaina-ml/ml_forge ? reverse-engineered prompt

Build me a simple desktop app for training image classification models without writing code.

I want a visual workspace where I can drag blocks onto a canvas, connect them with lines, and have the app turn that into a working PyTorch training pipeline. It should be beginner friendly, with a clear flow for data setup, model building, and training. Let me pick common datasets like MNIST, CIFAR 10, CIFAR 100, FashionMNIST, or a custom image folder, then chain together basic transforms and data loaders.

For the model side, I want to add common layers like convolution, linear, ReLU, batch norm, flatten, and dropout, with useful values filled in automatically when blocks are connected. Then I want a training screen where I can wire the model, loss, and optimizer, choose epochs and device, and hit run to see live loss and accuracy updates.

Please also include saving projects, loading them back, running inference on a checkpoint, and exporting the pipeline to a standalone PyTorch script. Look up current docs online if you need to.