Starting point for workshop on how to open source an MNIST model with mlhub

mlhub packages ML models so they are easily accessed, run, rebuilt and deployed.

First, to train an MNIST model, run python model.py. The output of this training script is a model checkpoint saved as best.pth.tar and a set of weights metrics_val_best_weights.json.

Given this model checkpoint, can we package the model into a format an end user can get predictions from?

If you don't manage to package the model from the checkpoint, vgg.py contains an example of how to load a pretrained model stored on aws and use the pretrained model to predict the labels of an image.

To make a git project repository mlhub-friendly, simply include a MLHUB.yaml configuration file. To help others understand your model, ideally also include a script to 1. get data eg. dataloader.py and 2. demonstrate what your model does eg. demo.py

A user can then install and run the pre-built model.

1. $ pip install mlhub
2. $ ml install MNIST_mlhub
3. $ ml configure MNIST_mlhub
4. $ ml demo MNIST_mlhub