Benefits of defining experiments in Python

Defining the experiments in pure Python makes it possible reuse experiment configuration code and avoid bugs.

Ever wrote an experiment configuration in a .json file as

{
   "dataset": "mnist",
   "optim": {"name": "sgd", "lr": 0.01},
   "batch-size": 1000,
}

only to realize that batch-size should have been batchsize and you just ran hours of compute on the default batch size of batchsize=128 instead? As the experiment definitions are written in Python, you can check for common mistakes without even running them. Your IDE will tell you about errors and you can even run a static type like mypy.