Hervé BREDIN

I'd also use (and beta test) such a separate config/registry package! In the meantime, I might still consider adding `thinc` as a dependency to my current project just for `thinc.Config`...

I would also try to keep _cannot-link_ constraints for overlapping speakers only (and allow merging non-overlapping speakers). The implementation might be trickier, though.

Despite this huge slowdown, did you estimate the gain in accuracy? How much faster is it on GPU?

The question that ensues is somehow related to #18 (rather a generalization of it). Since paths are now generated automatically, it is not that obvious to know where output files...

FYI, I ended up saving every _automagically_ generated output paths in an attribute of the parent workflow: https://github.com/pyannote/pyannote-workflows/blob/master/pyannote_workflows/utils.py#L56-L68

Indeed, it would be great to have `sciluigi.Workflow` implement the standard `luigi.Task` interface: - `.requires(self)` - `.run(self)` - `.output(self)` But, really, what I am looking for right now is a...

FYI, I ended up adding a dummy `Hyperopt` task at the end of my workflow that write the path of the final task to a temporary file whose path (`temp`)...

I'd like to say yes because I really like the very idea of JOSS. However, I am not sure I can commit to working on this before the end of...

I have been playing with `pescador` and pytorch `DataLoader` lately and ended up doing the following: ```python from torch.utils.data import IterableDataset class _PyTorchDataset(IterableDataset): def __init__(self, stream): super().__init__() self.stream = stream...

@faroit Not yet, no. This is on my TODO list with no ETA :-)