Progressive resizing for time series instead of images

fastai
1

Progressive resizing for a CNN is something Jeremy talked about in almost all his courses and seems to be something that helps with images.

I have been working with time series data recently and you can use Conv1D instead of Conv2D on time series data. I was wondering if anyone tried progressive resizing with CNN, but with time series. I guess gradually increasing the length of the data we feed to the network.

Anyone had any good results with this technique? I could not seem to find any good papers on the subject.

Thanks!

2

Since you can convert time series into images (see: Share your work here ✅ - #367 by oguiza), it may be possible just to use standard progressive resizing.

1 Like
3

Interesting idea! Looking forward to seeing the results of your experiments.

4

Yeah I saw that implemented in tsai! I might try that, but I am not sure this would allow increasing the length of the encoded time series. I will probably try actually playing with tsai with a CNN model and try the progressive resizing.

1 Like