search

chathasphere / pno-ai

Music Transformer Sequence Generation in Pytorch
MIT License
102 stars 29 forks source link

Training/Validation Split in Data #5

Closed chathasphere closed 5 years ago

chathasphere commented 5 years ago

How will this trade be affected by the various data augmentation strategies?

chathasphere commented 5 years ago

From what I understand, the idea is to compare training loss vs. validation loss every "x" number of iterations (say 1,000).

Following magenta's example, start with a 90/10 training/validation split. Only augment (transpositions, tempo stretches) the training data.