Train experimental models on chunked dataset

ubclaunchpad / minutes

:telescope: Speaker diarization via transfer learning

https://medium.com/ubc-launch-pad-software-engineering-blog/speaker-diarisation-using-transfer-learning-47ca1a1226f4

27 stars 5 forks source link

Train experimental models on chunked dataset #86

Closed iKevinY closed 6 years ago

iKevinY commented 6 years ago

Even though we haven't implemented HDF5 yet, we can train on the smaller (5-class) data chunks.

grig-guz commented 6 years ago

K I'll try to pull something out this week. I suggest we use this one: https://keras.io/applications/#densenet Dense nets are quite hot and fresh right now, they're like Reznets but ALL layers are connected to each other instead of skip connections. 121-layer one is pretty lightweight too (33 MB).

Lol, it looks like it's completely plug and play. Input dimention is customizable, and the number of classes too.

iKevinY commented 6 years ago

Seems like we've tackled this issue sufficiently! 🎉