experiments code, or example train and generation?

[rlabs-oss]

An example showing instantiation from a checkpoint to generate audio would be appreciated.

[flavioschneider]

You are passing a dataset to the input instead of a batch of elements, please have a look at the examples in the readme to understand the correct tensor shape.

[flavioschneider]

Have you installed version 0.0.15?

[rlabs-oss]

0.0.19 was installed, reverted back to 0.0.15 and restarted the kernel - the checkpoint via url returns "All keys matched successfully"

Thank you

[matigekunstintelligentie]

When I change the model to a checkpoint, which was resumed from the pretrained model, I trained with the audio-diffusion-pytorch-trainer repo. I get missing keys too. I only changed
model.load_state_dict(torch.load("last.ckpt"))
The yaml has in_channels=2, diffusion_sigma_data=0.2
I've also tried both 0.0.19 and 0.0.15 to no avail.
What am I doing wrong?

[flavioschneider]

The created checkpoint is a pytorch-lightning .ckpt checkpoint, not directly compatible with pytorch .pt checkpoints. Please check the FAQ in the README.

[rlabs-oss]

In practice that's something resembling:-
The current directory is the trainer folder (audio_diffusion_pytorch_trainer)

from main.module_base import Model

model = Model.load_from_checkpoint(
    checkpoint_path='/logs/ckpts/YYYY-MM-DD-HH-MM-SS/last.ckpt',
    learning_rate=1e-4,
    beta1=0.9,
    beta2=0.99,
    in_channels=2,
    patch_size=16
)
torch.save(model.model, 'torchckpt.pt')

With instantiation using one of the existing code examples;

model=torch.load('torchckpt.pt', map_location="cuda")

with torch.no_grad():
.... rest of code...

Play from notebook, or save sample:

mysample = Audio(sample_long.squeeze().cpu(), rate = sampling_rate)

with open('generated_test.wav', 'wb') as f:
    f.write(mysample.data)

[rlabs-oss]

above example no longer works with the latest version.

Im attempting (with no examples)

from main.module_base import Model
#main.module_base.Model

model = Model.load_from_checkpoint(
checkpoint_path='logs/ckpts/2022-10-10-23-42-25/last.ckpt',
lr=1e-4,
lr_beta1=0.95,
lr_beta2=0.999,
lr_eps=1e-6,
lr_weight_decay=1e-3,
ema_beta=0.9999,
ema_power=0.7,
model=AudioDiffusionModel << not sure thats correct?
)

which also doesnt work.

1878 Module: self
1879 """
-> 1880 for p in self.parameters():
1881 p.requires_grad_(requires_grad)
1882 return self

AttributeError: 'bool' object has no attribute 'parameters'

parameters obfuscated in more parameters.