I’d like to participate. Thank you.
I would like to participate 
Hi, I am new here. And I really like the contents, it is very informative. However, I have a doubt I was going through inceptionTime paper, and other papers also, I just wanted to ask if I am using convolution neural networks over time-series then convnet will treat the time series as a bag-of-seq(if I am just applying over time-series) as convolution neural networks are translational invariant(in the case of computer vision) thus in the case of time series data convnet will become time-invariant. And I also remembered someone has mentioned uber’s paper on cord-convnet, where they are encoding the position coordinate to make convnet translation-invariant(correct me if I am wrong), similarly in Transformers NN there they use positional encoding to avoid this issue(in the case of NLP it is bag-of-words). So using positional encoding, will it help the model? @hfawaz.
sorry If I have mistaken something
Interested
Hi oguiza,
For those of us unable to join, can we get a recording of this? Maybe upload it to YouTube?
Me too, please.
Hi @shanya,
Welcome to fastai and the time series/ sequential data community!
I’ll try to answer your question since Fawaz has not been participating lately.
You raise an interesting paradox. Why do we use a Conv1d (temporal invariant) to learn time series where time is so important?
My view is that there are different types of time series. In an electrocardiogram (unique, long time series split in chunks), for example, the different heart beat waves may occur at any time, although the duration and relative wave lengths are important. In other cases (“end-to-end” time series), like in a food spectrogram, the position of the peaks is critical, and cannot happen at any time, so they are critical.
Based on this idea, I added CoordConv to InceptionTime in the tsai library. There’s a model called InceptionTimePlus with a kwarg coord (in tsai models ending with Plus are models I use for research, so I add different options that may not be available in the original model). When coord=True, an additional positional channel is added to each convolution layer. As everything in DL you’ll need to test it to see if it can add any value. I’ve seen everything from slight decrease to a good performance boost. But coord is something I always test, especially with the “end-to-end” time series.
1 Like
I’m in! 
1 Like
Hi @boris, based on your suggestion I’m planning to record and upload the meeting to YouTube unless the presenter or any of the attendees withdraw the permission to do so.
2 Likes
Sure! No problem.
All of you who have requested the link to the meeting should have received it by email. Otherwise please let me know.
Will be very helpful if the video could be recorded and uploaded to Youtube. Meeting is slightly late hours for me in India.
1 Like
Hey, @oguiza thanks for the reply. Yeah, it seems paradoxical to me too. Anyways in my work I used this paper PDE, I used it to solve one dimensional PDE and it has saved a lot of our time 
. The authors instead of using convolution they are mapping the data to the frequency domain and then they are applying a linear transformation. Which kind of make sense as the solution of any PDE(which satisfy boundary value condition) can be written as the superposition of different solutions thus fourier series(or analytical series).
But I believe any time series can be represent as a PDE and you don’t need know what is that PDE and how to solve it. If you have data and corresponding labels then mapping the data to frequency domain you may find a solution. Hope that would help someone
1 Like
I’m looking to do something similar to groupkfold from sklearn in tsai but can’t quite wrap my head around what I could do.
Let’s say each group is a season with no date column, only a column for the season i.e. the row of data could fall on any date within a season.
I understand I want to have
group 1 = winter 1
group 2 = spring 1
…
group 5 = winter 2
group 6 = spring 2
How can I mimic this prior to modelling? I understand their is slidingwindow but the lack of date is proving hard for me to get working.
Any thoughts?
Hi, This is just to inform you that I will not be able to record today’s session with Angus Dempster due to a technical problem. Apologies to those who won’t be able to attend live.
Sent.