Open JiyeongCha opened 1 month ago
Hi Jiyeong,
Currently, the model is unable to target monophonic generation (i.e., melody) given a checkpoint trained on polyphonic dataset. However, you can train the model on a melody dataset, so that the generation would likely be melody as well. To achieve this with the built-in POP909 dataset, use the argument --pop909_use_track=0
to select only the melody track for training. For example:
python polyffusion/main.py --model=sdf_chd8bar --output_dir=result/sdf_chd8bar --pop909_use_track=0
Please pull the last commit 1bd4bef0cc438cb9cbdb5a91ec2020d59f6bc8a1 where I fixed a bug regarding this.
Or, you can specify a custom melody dataset using the --data_dir
argument.
As for the ckpt of pnotree, I regret to tell that I don't have it at hand. If this is a must, kindly write to me on this matter.
And can you share the ckpt of pnotree?