Kinect WSJ Recipe #131
Kinect WSJ Recipe #131
Conversation
1. Contains recipe to train Deep Clustering on Kinect WSJ TODO: Include the multichannel of DC
There was a problem hiding this comment.
Thanks a lot !
Can you add a README.md directly into egs/kinect-wsj that has a small description of the dataset, a link to your repo and a way to cite the paper please?
Also you can add it to the main README in the supported dataset section.
There are few minor format issues that could be fixed with a linter (black for example)
| os.path.join(exp_dir, 'checkpoints/final.pth')) | ||
|
|
||
|
|
||
| # TODO:Should ideally be inherited from wsj0-mix |
There was a problem hiding this comment.
I'd love to reduce the copy pasting from recipes to recipes but apart from adding this things into the source code, I didn't find a good solution yet. Any good idea welcome !
There was a problem hiding this comment.
Is Deep clustering the only recipe which requires a custom System class?
There was a problem hiding this comment.
No, also in DNS challenge and the two step recipe
There was a problem hiding this comment.
DC models can probably go into blocks.
System callbacks can handle common_steps I think. Will have to analyse a bit more to see if it gets complicated.
|
Tests are failing because Zenodo's sandbox is down, no link to your PR. |
|
Thanks that's perfect for now ! |
TODO:
Multichannel Deep Clustering