[WIP] Adds dense keras classifier training routine and metrics.

cosmir / dev-set-builder

Boostrapping weak multi-instrument classifiers to build a development dataset with the open-mic taxonomy.

MIT License

4 stars 0 forks source link

[WIP] Adds dense keras classifier training routine and metrics. #12

Open ejhumphrey opened 6 years ago

ejhumphrey commented 6 years ago

Working through adding a flexible dense network classifier and metrics.

Currently designed to operate on a single vector input. This is hands down the easiest thing to do at the moment, some kind of post-filtering might get us something decent to start. Can extend to context windows (len>1) and time-sensitive models later (CNNs, recurrent models, etc).

ToDo's:

[ ] Random parameter generation
[ ] Add class-conditional streaming via pescador
[ ] Introduce exponential (alpha) param to control distribution imbalance.

@bmcfee anything else obvious missing?

ejhumphrey commented 6 years ago

hooray! compiler issues with scipy on travis 😠

potentially related

@bmcfee any experience with this one? I can muck with it later, but I'd rather not troubleshoot things with known answers...

bmcfee commented 6 years ago

@ejhumphrey yeah, I fixed this over in #7 last week. Do you want to merge that and rebase this to get the fix?

ejhumphrey commented 6 years ago

weird... did you fix it on #7? I thought you fixed it on master before rebasing that other branch (which this branch should have)?

bmcfee commented 6 years ago

No, I rebased #7 and then fixed it there. Master has not changed.

ejhumphrey commented 6 years ago

it's not this?

ejhumphrey commented 6 years ago

yep, you're right. fixes will come back from #7