Can I use the model on a machine without CUDA? about wavernn

[fatchord]

@vitaly-zdanevich Yes you can. I would recommend opening up the repo in a decent IDE and using the search and replace functionality across all files:

".cuda()" replaced with an empty string ""

You should be able to run everything on CPU after that small change.

[vitaly-zdanevich]

Removed ".cuda()" with find . -type f -exec sed -i 's/.cuda()//g' {} + (you can check before and after with grep -r ".cuda()" .), but looks like this is not enough:

$ python3 quick_start.py 

Initialising WaveRNN Model...

Trainable Parameters: 4.481M

Loading Weights: "quick_start/voc_weights/latest_weights.pyt"

Traceback (most recent call last):
  File "quick_start.py", line 62, in <module>
    voc_model.restore('quick_start/voc_weights/latest_weights.pyt')
  File "/home/vitaly_zdanevich/WaveRNN/models/fatchord_wavernn.py", line 375, in restore
    self.load(path)
  File "/home/vitaly_zdanevich/WaveRNN/models/fatchord_wavernn.py", line 378, in load
    self.load_state_dict(torch.load(path), strict=False)
  File "/home/vitaly_zdanevich/.local/lib/python3.6/site-packages/torch/serialization.py", line 368, in load
    return _load(f, map_location, pickle_module)
  File "/home/vitaly_zdanevich/.local/lib/python3.6/site-packages/torch/serialization.py", line 542, in _load
    result = unpickler.load()
  File "/home/vitaly_zdanevich/.local/lib/python3.6/site-packages/torch/serialization.py", line 505, in persistent_load
    data_type(size), location)
  File "/home/vitaly_zdanevich/.local/lib/python3.6/site-packages/torch/serialization.py", line 114, in default_restore_location
    result = fn(storage, location)
  File "/home/vitaly_zdanevich/.local/lib/python3.6/site-packages/torch/serialization.py", line 95, in _cuda_deserialize
    device = validate_cuda_device(location)
  File "/home/vitaly_zdanevich/.local/lib/python3.6/site-packages/torch/serialization.py", line 79, in validate_cuda_device
    raise RuntimeError('Attempting to deserialize object on a CUDA '
RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location='cpu' to map your storages to the CPU.

[vitaly-zdanevich]

Ok I edited models/fatchord_wavernn.py at line 378 and models/tacotron.py at 411 (was: torch.load(path), now: torch.load(path, map_location='cpu')).

[vitaly-zdanevich]

Without CUDA speech synthesis is slow, I will try to use a machine with CUDA. Just asking - no way to increase rendering speed? Maybe it is possible to decrease a quality somehow?

[vitaly-zdanevich]

I want some TTS on AWS Lambda, but this technology does not support CUDA.

[vitaly-zdanevich]

Tried on CUDA, yes it much faster, but I see that from 8 CPU cores only one is busy, just saying.

[vitaly-zdanevich]

Tried instance on Google Cloud with 4 GPUs with CUDA - looks like the same performance. @fatchord is it correct that only one GPU is used anyway?

time python3 quick_start.py -u --input_text "What will happen if I run this command?"
/home/vitaly_zdanevich/.local/lib/python3.6/site-packages/numba/errors.py:105: UserWarning: Insufficiently recent colorama version found. Numba requires colorama >= 0.3.9
  warnings.warn(msg)

Initialising WaveRNN Model...

Trainable Parameters: 4.481M

Loading Weights: "quick_start/voc_weights/latest_weights.pyt"


Initialising Tacotron Model...

Trainable Parameters: 11.078M

Loading Weights: "quick_start/tts_weights/latest_weights.pyt"

+---------+---------------+-----------------+----------------+-----------------+
| WaveRNN | Tacotron(r=1) | Generation Mode | Target Samples | Overlap Samples |
+---------+---------------+-----------------+----------------+-----------------+
|  804k   |     196k      |    Unbatched    |      N/A       |       N/A       |
+---------+---------------+-----------------+----------------+-----------------+
 

| Generating 1/1
| ████████████████ 54400/54450 | Batch Size: 1 | Gen Rate: 1kHz | 

Done.


real    0m53.960s
user    0m43.638s
sys     0m10.492s

[vitaly-zdanevich]

@fatchord maybe it is possible to use GPUs and multiple CPU cores?

[TheButlah]

Switching between single GPU and cpu training has been added as of #107