Remove PyTorch from the code

[JackKelly]

When I started nowcasting_dataset, the intention was to use nowcasting_dataset to generate batches on-the-fly during ML training from separate Zarr stores for the satellite data, NWPs, and PV. But that turned out to be too slow and fragile :) So, we swapped to using nowcasting_dataset to pre-prepare batches ahead-of-time, and save them to disk. During ML training, we just need to load the batches from disk, and we're good-to-go. (Pre-preparing batches has a number of other advantages, too).

But, this development history means that nowcasting_dataset still uses PyTorch (e.g. using the PyTorch DataLoader to run multiple processes). The code may become cleaner and faster and more flexible if we strip out PyTorch, and instead (maybe) use concurrent.futures.ProcessPoolExecutor to use multiple processes.

TODO:

• Remove Datasets and Datamodule. Done in PR Big new design Part 2 :) #307
• Remove pytorch lightning from requirements.txt and environment.yaml. Done in PR Big new design Part 2 :) #307.
• Remove torch from other python files
• Remove torch from requirements.txt and environment.yaml

[peterdudfield]

#213 (comment)
from big issue

Idea is to use optional requirements for pytorch

[JackKelly]

If it's OK, I'll keep this issue open until we've removed the pytorch dataloader and pytorch lightning from the "batch pre-processing" code :)

[peterdudfield]

sure thing, where is that?

[JackKelly]

Tthe specific places where pytorch / pytorch lightning are still used are:

• NowcastingDataModule inherits from pl.LightningDataModule. I think we can remove the dependency from pl.LightningDataModule and have NowcastingDataModule inherit from nothing.
  • NowcastingDataModule.train_dataloader(), val_dataloader(), and test_dataloader() all return torch.utils.data.DataLoader objects.
• NowcastingDataset inherits from torch.utils.data.IterableDataset.

I might strip out these PyTorch things in one of the sub-steps of #202 (but I haven't fully thought this through!)