Error importing torch

[tobeyond]

When I try:

C:\Users\...\jukebox>python jukebox/sample.py --model=5b_lyrics --name=sample_5b --levels=3 --sample_length_in_seconds=20 \ --total_sample_length_in_seconds=180 --sr=44100 --n_samples=1 --hop_fraction=0.5,0.5,0.125

I get this:

Traceback (most recent call last):
  File "jukebox/sample.py", line 3, in <module>
    import torch as t
  File "C:\Users\...\.conda\envs\jukebox\lib\site-packages\torch\__init__.py", line 81, in <module>
    ctypes.CDLL(dll)
  File "C:\Users\...\.conda\envs\jukebox\lib\ctypes\__init__.py", line 364, in __init__
    self._handle = _dlopen(self._name, mode)
OSError: [WinError 126] The specified module could not be found

Any ideas?

[johndpope]

do you have nvidia card? cuda ? Looks like it didn't install properly. UPDATE https://stackoverflow.com/questions/61488902/cannot-import-pytorch-winerror-126-the-specified-module-could-not-be-found

[tobeyond]

Not sure. How do I know that?

I proceeded with conda install pytorch torchvision cpuonly -c pytorch

Now I get this:

Using cuda False
Traceback (most recent call last):
  File "jukebox/sample.py", line 275, in <module>
    fire.Fire(run)
  File "C:\Users\...\.conda\envs\jukebox\lib\site-packages\fire\core.py", line 127, in Fire
    component_trace = _Fire(component, args, context, name)
  File "C:\Users\...\.conda\envs\jukebox\lib\site-packages\fire\core.py", line 366, in _Fire
    component, remaining_args)
  File "C:\Users\...\.conda\envs\jukebox\lib\site-packages\fire\core.py", line 542, in _CallCallable
    result = fn(*varargs, **kwargs)
  File "jukebox/sample.py", line 267, in run
    rank, local_rank, device = setup_dist_from_mpi(port=port)
  File "c:\users\...\jukebox\jukebox\utils\dist_utils.py", line 55, in setup_dist_from_mpi
    torch.cuda.set_device(local_rank)
  File "C:\Users\...\.conda\envs\jukebox\lib\site-packages\torch\cuda\__init__.py", line 245, in set_device
    torch._C._cuda_setDevice(device)
AttributeError: module 'torch._C' has no attribute '_cuda_setDevice'

[tobeyond]

I also tried adding --device cpu and --gpu_ids -1 to my command line but it didn't work. Nor did changing torch._C._cuda_setDevice(device) to torch._C._cuda_setDevice(-1).

[johndpope]

frankly - if you have an AMD card / radeon or are using a mac - then this project isn't for you unless you get an nvidia card. People are struggling to run this on cards with 8gbs of gpu ram. To run on a cpu would take > days to create seconds of audio. Use the google collab / TPU route if you want to get results today.

[tobeyond]

I'm on Windows 10 with 16GB of RAM and it's Intel Core i7. Do you think it works?

[johndpope]

Not talking about desktop RAM - it's GPU RAM. You need a high end graphics card - like the hightest. Volta 100 / RTX 5000 - it works if you run on google colab / they use Google's TPUs.

[tobeyond]

I don't have a dedicated graphics card. Is it mandatory?

[combs]

yes--they say it requires a graphics card (GPU) with 16gb of ram (usd $8000). Or you can play with it in Google CoLab for $0 and a lot of patience

[xandramax]

they say it requires a graphics card (GPU) with 16gb of ram

The stated 16gb requirement is for the 5_b model. This model generates with better quality but was also trained on fewer artists and genres (the v2 artist/genre IDs are for 5_b and the v3 IDs are for 1_b).

The 1_b model might not satisfy if you're after something like 1-click generation of Kanye performing Eminem, but I find it's very exciting when it comes to co-composition with a little patience. I run batches of 3 four-second samples at a time when co-composing, and almost always find something interesting within 2 or 3 batches.

Speaking from this experience, it's very much possible to explore model 1_b with an RTX 2080 Super. The card has only 8gb of VRAM. Having watched the GPU memory usage and having run into out-of-memory errors when attempting larger batch sizes, I believe that 8gb is likely the minimum possible.

That means it is probably possible to run the 1_b model on a card as old as a GTX 1070 (also 8gb VRAM).

With a 1080 Ti or 2080 Ti (11gb VRAM), or a Titan X or Titan V (12gb VRAM), it might even be possible to run the 5_b model with a small batch size but that's a total guess. It would have to be tested.

The Titan RTX has 24gb VRAM and costs $2.5k; much less than the $8k for a V100.

Worth considering is Nvidia's stated "compute capability" for their various cards (viewable here). The V100 is listed with a compute capability of 7.0, whereas the RTX line is 7.5. For comparison, the K80 that Colab sometimes hands out only has compute capability of 3.7 (still enough for exploring 1_b), and a GTX 1070 has compute capability 6.1.

All that to say, yes, a dedicated graphics card is mandatory but you may be able to spend as little as ~$300 to get going at home with the 1_b model on a 1070.

[combs]

thank you, that's good to know!

[tobeyond]

Thank you very much for the clarification. I've never used google colab and I'm a little confused here. Does anyone have a step-by-step explanation on how to run it there?

[johndpope]

https://www.youtube.com/results?search_query=google+colab&page=&utm_source=opensearch

[cicinwad]

I'm on Windows 10 with 16GB of RAM and it's Intel Core i7. Do you think it works?

You need to be on windows 8, it will work better.