purzelrakete
commented
2 years ago Experiment
bin/train full-maestro -p batch_size 8 -p batch_norm True -p learning_rate 0.035 -p max_epochs 1Open purzelrakete opened 2 years ago
purzelrakete
commented
2 years ago bin/train full-maestro -p batch_size 8 -p batch_norm True -p learning_rate 0.035 -p max_epochs 1
What
Run a model with around 13MM parameters. This includes 256 channels on everything, which is the model that is able to overfit the Tracks dataset. Reports of well performing Wavenets always seem to point to around 1MM steps.
The maximum time budget for this run is 10 days of uninterrupted training. In order for this to work well, resuming of training runs must work properly.
Hypothesis
We can produce realistic piano samples from the unconditionally trained model. This seems possible because that's what they did in the Maestro paper. One unclear thing though is whether the unconditional samples were obtained from an unconditionally trained Wavenet. It's possible that training was on a conditioned Wavenet, and unconditioned samples were obtained by passing in blank conditioning signals. I would imagine that this sort of scheme might work better, since we gave much more information to the model during the training process.
Results
Write up the results of your experiment once it has completed and has been analysed. Include links to the treatment run, and also to the baseline if appropriate.
Acceptance Criteria