make it possible for predict to load very large spectrograms

[NickleDave]

will need to turn them into multiple batches, probably within predict function?

or maybe the data loader can take care of this using a clever collate_fn

[NickleDave]

since we're already using ViewAsWindowBatch, quick-and-dirty solution might be to set a max_batch_size attribute and then split above that. Then let the collate_fn deal with those splits (?)

[atomfried]

You should be able to reuse the data generator you use for training, no? (Without shuffling ofc).