magenta / ddsp

DDSP: Differentiable Digital Signal Processing
https://magenta.tensorflow.org/ddsp
Apache License 2.0
2.87k stars 332 forks source link

DDSP VST Model training & autoencoder training both instantly start to slow down significantly #481

Open olaviinha opened 1 year ago

olaviinha commented 1 year ago

Describe the bug Training starts to slow down instantly after you start it. After a few thousand steps the pace is already considerably over 1 second per step. Around 8k steps it is already taking several seconds for each step. "Training typically takes ~2-3 hours with free Colab" mentioned in the VST Training notebook is not true at all with this bug.

System Info

To Reproduce

  1. Run Train VST notebook
  2. Observe the training pace after a few thousand steps (e.g. around 5000 steps) -> is extremely slow.
  3. Interrupt training.
  4. Restart/continue training.
  5. Observe the training pace after continuing training from the last saved checkpoint -> is fast again. You can observe the difference clearly with naked eye, no need to add measurements.

Additional context Stopping and restarting !ddsp_run programmatically e.g. every 1000 steps seems to work around the issue. Otherwise it will take literally days or weeks to train a model.