[HN Gopher] Data Science 2020 - Highlights
       ___________________________________________________________________
        
       Data Science 2020 - Highlights
        
       Author : amrrs
       Score  : 11 points
       Date   : 2021-01-03 18:39 UTC (4 hours ago)
        
 (HTM) web link (nulldata.substack.com)
 (TXT) w3m dump (nulldata.substack.com)
        
       | Jugurtha wrote:
       | > _Data Scientists are bad Web Developers, but what if we need to
       | build web apps that can talk or do Machine Learning?_
       | 
       | We do have colleagues like that, but even for those who can go
       | end to end (helping client to state the problem, to data
       | acquisition, model building, writing application, setting up
       | infrastructure, deploying, and monitoring), it's tiring..
       | 
       | We'd get the data, build the models, and then we'd have to write
       | the application. That's a separate repository and software
       | project, often on-premises deployment which we have to maintain.
       | And that's for the "application" proper with user management, and
       | business logic. I'm not even talking about what most people do
       | "Flask app that loads model weights, exposes a form, and returns
       | predictions to show the client, let me set a VM on GCP and send a
       | link to the client", but the client is busy, and you forget about
       | it, so you shut down the VM, the client clicks on the link,
       | nothing there, and you're left to remember which of the VMs and
       | which of the models was there.
       | 
       | This toil is one of the reasons we're building https://iko.ai,
       | because we've done it so many times for clients, that certain
       | patterns and inefficiencies have emerged and we're addressing
       | them in our platform to really focus on high value things.
       | 
       | For example, we have AppBooks[1] where a machine learning
       | practitioner clicks a button and publishes an _automatically_
       | parametrized notebook, without tagging cells or using metadata,
       | or cluttering the notebook with code that does that. It generates
       | a form and the client or domain expert can then change parameters
       | and run the notebook without mutating it.
       | 
       | One other benefit is that this automatically tracks the run: the
       | parameters, metrics, and the model generated is saved, and you
       | can then deploy that model in one click or build a Docker image.
       | These are not much, but solve the problem of "I want to show
       | results and get feedback, how do I do that? VM on GCP and Flask
       | app? Export to PDF?"
       | 
       | > _Finally, Models are easy to build on Jupyter Notebooks. We all
       | know it just takes a few lines of code and your `model.fit()` is
       | ready. But what's next?_
       | 
       | Well, it depends. Long-running notebooks have been problematic
       | when the kernel and the front-end would stop talking with each
       | other. The computation would take place, but the results wouldn't
       | find their way to the front-end. People would be watching
       | training, and then there would be a disconnection or they'd close
       | the browser tab and lose results. Some would add code to save the
       | model/pickle it, but that's clutter. We've added long-running
       | notebook scheduling[2] right from the notebook's interface, so
       | closing a tab or shutting down your computer doesn't impact
       | anything. You can view the output _outside_ of JupyterLab in a
       | simple page, even on your phone.
       | 
       | That's very useful because it sometimes happens where we want to
       | run a _bunch_ of experiments. We just schedule them all in a fire
       | and forget way.
       | 
       | But the gist of it is that so far, it's helping us:
       | 
       | - Load data
       | 
       | - Start notebooks with the most popular libraries pre-installed
       | 
       | - Collaborate in real-time on notebooks and see cursors and
       | changes
       | 
       | - Publish AppBooks: automatically parametrized notebooks
       | 
       | - Schedule long-running training jobs and be able to watch their
       | output as they run
       | 
       | - Automatically track experiments without remembering to do so or
       | polluting the notebook with tracking code
       | 
       | - Deploying models into a "REST endpoint", and monitoring their
       | performance in near-real time dashboards
       | 
       | Basically, a _lot_ of work we 're happy not to be doing :)
       | 
       | - [0]: https://iko.ai
       | 
       | - [1]: https://iko.ai/docs/appbook/
       | 
       | - [2]: https://iko.ai/docs/notebook/#long-running-notebooks
        
       ___________________________________________________________________
       (page generated 2021-01-03 23:02 UTC)