I was wondering if timeseriesAI could handle categorical variables using the Categorify transform?
Particularly interested in your dataloader using np.memmap, very interesting stuff for very large datasets. I think this should actually be in fastai itself because it seems so useful for larger datasets.
In theory it could work, you would end up with a discrete ts, maybe passing through and embedding first to get a multivariate continuous ts?
The memap was a neat discovery by @oguiza forced by his usage constraints. I don’t see requirements for this on fastai besides tabular., but probably it would be easier to use a dask df for this task anyway. (I think jeremy had a notebook about this)
It would be awesome to see a tutorial notebook with a multivariate time series to image conversion.
another cool project: https://github.com/jdb78/pytorch-forecasting
It has Temporal Fusion Transformers and N-Beats
wow! thanks for sharing!!!
Hi all,
I’ve just finished a major update to the tsai library and added lots of new functionality, models and tutorial notebooks. It now works well with fastai v2 and Pytorch 1.7.
If you are interested you can find more details in this blog post or in the repo.
Hi Ignacio,
I just learned of your tsai library - amazing! I have used the approach to convert a time-series into an image a while ago (using fastai v1) and it worked really well. My application was to simply plot the altitude of airliners versus time and convert the charts into images for CNN analysis. I am interested in converting this code to use tsai/fastai v2. Here is the link to:
Time-series Classification (Deep Learning to Classify Flight Profiles)
By the way, I published this on a portfolio site I created using fastpages, another of Jeremy’s inventions…
Keep up the good work.
Thanks for sharing @kobus78.
Your notebook is a great example of how time series to image transforms can really add value!!!
This really motivates me to keep improving and adding new functionality to the tsai library.
wow, good work! this could be a paper!
I have a couple of questions:
Did you tried Rocket? or something simpler like a 2channel CNN as baseline?
I don’t know the resolution you have on your time axis, but transforming to images can make you loose a lot of info due to pixel resolution limitations.
When you said inference, do you used some pytorch backend for serving your models? (torchscript, tensor RT, onnx, etc…?) I am curious how those work for TS.
Hi Thomas,
Thanks for the feedback.
I did not try Rocket or something simpler - I specifically wanted to use the opportunity to dig into fastai.
Thanks for the comment about losing resolution - something to keep in mind. In my case, though, airliners change altitude in significant steps - the only changes I were interested in for this problem. Also, a step change takes substantial time relative to a pixel’s resolution.
By inference, I simply meant subjecting the models to unseen test images, i.e. I did not deploy them to production.
Thanks for your reply @oguiza. I actually forgot to add another example: This one also uses the principle of first converting a time-series to an image, but, instead of performing classification, it performs regression. Please see:
Time-series Regression (Deep Learning to Detect Change Points)
Granted, this approach is probably an overkill. Also, it is probably not as accurate as more traditional approaches. But, I was curious about how it would do. This report links to the actual notebook. At the bottom of the notebook I have visualized the regression predictions in green dotted lines: notebook
A faster Rocket model is avialable:
8 minutes to run the full UCR dataset!!!
Not presupposing any particular interest in this, but as I said here, I’m happy to answer any questions anyone might have, help with troubleshooting if you want to try it out, and also just hear any feedback.
https://towardsdatascience.com/synthetic-time-series-data-a-gan-approach-869a984f2239
Meeting with Angus Dempster - Rocket, MiniRocket, and MultiRocket
Hi all,
I’d like to invite you to participate in a web meeting we’ll have with Angus Dempster next week (@angusde ).
For those of you who don’t know him, Angus is a Ph.D. student at Monash University in Australia (a world-class group in time series research) and is one of the authors of several outstanding papers in Time Series Classification.
If interested, please read this.
@oguiza Does your FastAI based TS library include “tft” implementation, like the one in pytorch ? Do you have a separate dataset/dataloader to handle it ?
I was advised by @tcapelle to share my Pytorch based probabilistic time series framework work here. Kindly have a look at https://github.com/zalandoresearch/pytorch-ts for some state of the art uni and multivariate time series methods that are implemented. Thanks!