Great writeup, always interesting to see people digging into issues like that in such detail. I work at Google and shared the article with the Gboard team. Thanks!
Google employee here, xmanger is one of the main ML experiment tracking/orchestration tool we use internally, I'm pretty excited that it is now available for other to use!
In a nutshell, xmanager allows you to:
- define an experiment, which is a collection of one or more work units (think combination of hyperparamters)
- manage the different jobs/executable required to run this experiment (TPU workers, tensorboard job, etc.)
- collect and display measurements from work units (loss, other metrics)
- keep a reproducible artifact which allows you to re-run the same experiment at any point in the future
If you are interested in time series predictions, I would suggest you had a look at darts (https://github.com/unit8co/darts). It's a well designed library which provides a unified API to deal with time series and try/compare different algorithms/frameworks like Prophet, recurrent NNs, etc.
Yeah, that's a pretty common problem. I'm sure they are actively working on making sending feedback easier :). In the meantime, your technique worked! The issue is now fixed!
That's pretty weird, thanks for reporting this! I work at Google in the Maps org, I reported your bug internally and will update this once I have a resolution.
I would highly recomment David Duvenaud's PhD thesis [1] for anybody interested in Gaussian Processes. I find his chapter [2] on expressing structure with kernels super interesting and quite intuitive. It's also one of the PhD theses with the best-looking graphics I've read!
Instead of a tactical exploration of a thorny technical problem or new tool, this article digs into patterns and practices that can be critical to a successful machine learning project. The article delves into four key aspects of machine learning management: Connecting with users to enable closed-loop feedback, applying machine learning to the right objectives, building systems to support machine learning-based solutions, and understanding what goes into running a machine learning-based system.
Indeed, TikZ is great to create beautiful-looking plots! A friend of mine is quite good at it, e.g. look at Figure 2 in [1] and Figure 2.1 (p. 18) in [2]. For complex figures, though, I find that TikZ can be a bit hard to master and sometimes results in longer compilation times.
Indeed, I remember having to switch to Agg for this reason but can't remember why I switched back. Maybe some rendering issues I had with Agg, not sure...
That's a great point! Templates are a great tool to generate big tables from results, I usually do that for most of the results in my papers, makes it easier to have the odd copy/paste error. I might add this to the tips and tricks, thanks!