Dealing with 2D data about neuralcde HOT 6 CLOSED

I think conv2d should work. As I say, flatten before interacting with things like cdeint, and unflatten as the first operation inside the vector field (before calling conv2d or whatever).

from neuralcde.

Not sure what you mean by this. How do you plan to process 2D data sequentially?

from neuralcde.

For example if you have data in the format: (batch, time dimension, feature, subfeatures)
Aka an Image like but the feature channel is a time dimension.

from neuralcde.

Relative to the standard (batch, time, channel) format, then you're asking about having the channels in two different axes?
There's no (mathematical) reason you can't do that, it's just a shape thing.

If you're needing compatibility with my code (or anyone else's), then you can always flatten/unflatten (batch, time, feature, subfeature) <-> (batch, time, feature * subfeature) at the appropriate points.

from neuralcde.

Thanks, I totally understand this as this can work with your code.
But in my case the (batch, time dimension, feature, subfeatures) can change from a batch to batch, thus the usage of Linear layers will be incompatible, same with your code. This is why I was looking for an example that's totally built on Conv2d layers.

from neuralcde.

I think using a Conv1D can solve this issue and it will be compatible with your code, yet the spatial configurations will be lost.
Thanks

from neuralcde.