error log from standalone binary

[deccolquitt]

As per previous comments on solved issue #3, here is my error log when using a custom model in the standalone app

[-] Network - No API response
[+] RAVE - Model successfully loaded: /Users/declancolquitt/Library/ACIDS/RAVE/prior_baisline.ts.ts
 - sr : 1267073031
 - latent size : 32682
 - full latent size : 32682
 - ratio[ ] RAVE - Encode parameters     1
    1
    8
 2048
[ CPULongType{4} ]
[ ] RAVE - Decode parameters     8
 2048
    2
    1
[ CPULongType{4} ]
[+] RAVE - Model successfully loaded: /Users/declancolquitt/Library/ACIDS/RAVE/rave_baisline.ts.ts
 - sr : 1267073031
 - latent size : 32682
 - full latent size : 32682
 - ratio2048
to low; setting rate to : 11
libc++abi: terminating with uncaught exception of type std::runtime_error: The following operation failed in the TorchScript interpreter.
Traceback of TorchScript, serialized code (most recent call last):
  File "code/__torch__.py", line 54, in decode
      _8 = (torch.size(z2))[-1]
      _9 = ops.prim.device(z2)
      pad_latent1 = torch.randn([_7, int(pad_size), _8], dtype=None, layout=None, device=_9)
                    ~~~~~~~~~~~ <--- HERE
      pad_latent = pad_latent1
    z3 = torch.cat([z2, pad_latent], 1)

Traceback of TorchScript, original code (most recent call last):
  File "/home/paperspace/RAVE/export_rave.py", line 176, in decode
            )
        else:
            pad_latent = torch.randn(
                         ~~~~~~~~~~~ <--- HERE
                z.shape[0],
                pad_size,
RuntimeError: Trying to create tensor with negative dimension -32554: [2, -32554, 1]

[domkirke]

@deccolquitt there has been obviously something wrong with your model : where did you get it?

[deccolquitt]

@domkirke I trained it myself from scratch on paperspace using an a100

[deccolquitt]

i have been able to use generation.py with the exported rave .ts file and it produced audio (although I could not with the exported prior .ts file), same applies to using reconstruct.py. The audio wasn't great and had that fuzzy ringing throughout it (which I associated with the early training stages, it was my understanding that things cleared up in the second (prior) stage of training)

[caillonantoine]

Is the sampling rate of 1267073031 intentional ? Same question for the latent space 32682 ?

[deccolquitt]

@caillonantoine nope I didn't specify either of those as hyperparameters

[domkirke]

@deccolquitt there has been a problem with your model. Did you follow the instruments of the RAVE/README.md correcty? Your sampling rate should match your audio files (44100 / 48000 for audio usually), and the number of dimensions should be around 128. I advise you to train the model again using the cli_helper.py command line helper.