Satellite and nwp channel strings --> torch about nowcasting_dataloader HOT 13 CLOSED

Awesome, SGTM! Thank you!

from nowcasting_dataloader.

Could I get your advice on this @jacobbieker , happy for the channels strings to be changed to an index of numbers?

from nowcasting_dataloader.

What array of strings is it? The issue I see with changing strings ti numbers using the channel name is that if we change what channels we use, it could mess that up easily, incase we don't use all 12 channels

from nowcasting_dataloader.

yea so i was thinking of using a mapping from 'channel_name' to an index?

from nowcasting_dataloader.

Will we ever need to give the channel names to Torch? Is it sufficient to identify the channels by the ordering of the channel dimension?

from nowcasting_dataloader.

We probably want to make sure that

• ['name_1', 'name_2']
• ['name_1', 'name_3']
  are represented differently,

but other that that i think its ok

from nowcasting_dataloader.

Yeah, I guess it wouldn't be important for giving Torch the names, I was more going for us to know what is happening, but I guess thats saved in the config anyway!

from nowcasting_dataloader.

Is the main use case to tell the "modality encoding" code which channel is which? In which case, is that encoding done before pytorch?

from nowcasting_dataloader.

That isn't encoded at all right now actually, its only encoded in time and space, not channels

from nowcasting_dataloader.

PR: #48
This is what I was thinking

1. Dataset should save the channgel names in the xr.dataset as it useful to see. Although it is also in the config, so we could remove it
2. converting Dataset to torch, need to change it to indexes (away from strings)

from nowcasting_dataloader.

1 sounds great, thank you!

2 before bothering to convert names to Torch, please can we figure out what the use case is? :)

from nowcasting_dataloader.

That isn't encoded at all right now actually, its only encoded in time and space, not channels

Although, now that we are using native resolution images, I guess there will be two modalities, one for HRV and one for the rest of the channels

from nowcasting_dataloader.

1 sounds great, thank you!

2 before bothering to convert names to Torch, please can we figure out what the use case is? :)

2. so perhaps for the moment, we just dont pass the channels names to Torch, and if we come up with a use case we can add it?

from nowcasting_dataloader.