Synth1b1 dataset

[arlorostirolla]

Is it possible to download the synth1b1 dataset? I can't seem to find it anywhere

[jorshi]

@arlorostirolla -- synth1B1 isn't saved anywhere, it is deterministically generated on the fly. Check out the quickstart for info on how to generate the dataset https://torchsynth.readthedocs.io/en/latest/getting-started/quickstart.html. In this way, you can just iterate over the batches of the dataset during training. You could of course save the audio to disk and use the audio/parameters in a pytorch Dataset or similar.

A bit more info on the synth1B1 https://torchsynth.readthedocs.io/en/latest/reproducibility/synth1B1.html

[arlorostirolla]

Thanks for your help! just one more quick question, is there a way to load the tensor of parameters returned when generating synth1b1 data into the Voice module? In the docs the set parameters function takes a dictionary with (module, parameter) : tensor entries. Whereas output from Voice is a tensor of size 78. I'm not quite sure how to interpret the output or convert it into the dictionary format.
For clarity I'm trying to create a model that maps from audio data to the parameters, and want to eventually generate parameters for audio not found in synth1b1.

[jorshi]

No worries! If you're wanting to update all the parameters at once the easiest would be to directly set them in the voice. This would expect that all the parameters are in a normalized range between 0 and 1.

import torch
from torchsynth.synth import Voice
from torchsynth.config import SynthConfig

# We set reproducible False so we can call Voice without passing in a batch_num
config = SynthConfig(reproducible=False, batch_size=64)
voice = Voice(config)

audio, params, _ = voice()

# (batch_size, num_params)
assert params.shape == (64, 78)

# Generate new random params in range [0,1], which is required for internal
# torchsynth parameters
new_params = torch.rand_like(params)

# Now directly set Voice params
for p, new_p in zip(voice.parameters(), new_params.T):
    p.data = new_p.detach()

audio_2, params_2, _ = voice()

# output of voice is different now and used updated params
assert not torch.allclose(audio_2, audio)
assert torch.allclose(params_2, new_params)