Jupyter Notebook is excellent for explorative data science, but with larger data sets, it becomes a hassle to run on your local machine. Alternatively, when running Jupyter in the cloud, it becomes expensive to pay for the cloud unit even when you aren’t using it.
Exploration requires constant changes in the code and data sets. You should have automatic version control to rollback into code that produced superior results a couple of days ago.
Run Jupyter notebooks (or Jupyterhub) on a cheap machine (or even localhost) and do the calculations on demand in the cloud. Valohai launches the machines when needed and automatically shuts them down after the results return to you.
Every experiment is automatically tracked, and you'll have a full audit trail for your experiments from yesterday, the previous week or several years ago with the click of a button.
Data science requires trial-and-error work where running the experiment usually takes considerable time. The Valohai Jupyter Notebook extension lets you do asynchronous experiments. You can launch experiments sequentially or even tens at the same time while trying out different approaches. Read more about asynchronous experiments in deep learning.
Keeping track of what you have done is imperative for having an audit trail into how a model works but also for rolling back into a version of an experiment that behaved better than another one. The Valohai Jupyter Notebooks extension lets you come back to all your previous experiments and compare them between each other at the end of the day. No more guesswork or post it notes about best experiments, but instead real data on which model behaved the best and instant rollback to the code, data, hyperparameters, environment at the click of a button.