Real-time-voice-cloning: Cuda Error with low_mem

[PhysWiz314]

Hi,
I have come across the following error when using the toolbox in low memory mode.

On this computer, my GPU only has 2GB so I need to use this mode.

I have tested this on another computer that has a GPU with 4GB RAM. The toolbox works perfectly in normal mode but when I turn on low_mem, I run into the same error.

I'm not sure what other information you would need to look into this so please let me know what else I can provide to help out.

[CorentinJ]

What does demo_cli.py give?

[PhysWiz314]

Without the low_mem flag, it says that all tests pass. With the low_mem flag, I get the same CUDA_ERROR_NOT_INITIALIZED.

[CorentinJ]

Hmm, I don't have a solution right now. I'll have to look into it.

[nitram147]

Same problem.

"demo_cli.py" output without "--low_mem" flag:
Found 1 GPUs available. Using GPU 0 (GeForce MX150) of compute capability 6.1 with 2.1Gb total memory.
...
Use standard file APIs to check for files with this prefix. Testing the vocoder... All test passed! You can now synthesize speech.

"demo_cli.py" output with "--low_mem" flag:

Use standard file APIs to check for files with this prefix. E0707 15:55:23.503854 139813249828608 driver.py:321] Call to cuInit results in CUDA_ERROR_NOT_INITIALIZED
...
multiprocess.pool.RemoteTraceback: /python3.6/site-packages/numba/cuda/cudadrv/driver.py", line 233, in initialize raise CudaSupportError("Error at driver init: \n%s:" % e) numba.cuda.cudadrv.error.CudaSupportError: Error at driver init: [3] Call to cuInit results in CUDA_ERROR_NOT_INITIALIZED: multiprocess/pool.py", line 644, in get raise self._value numba.cuda.cudadrv.error.CudaSupportError: Error at driver init: [3] Call to cuInit results in CUDA_ERROR_NOT_INITIALIZED:

[haoqizhenhao]

I have the same error

[CorentinJ]

Can you try pulling from this new branch and see what gives?

[nitram147]

New branch code without "--low_mem" flag get stucked on "Interactive generation loop" after "all test passed".
Output:

W0710 12:49:35.993566 140170637424384 deprecation.py:323] From /Real-Time-Voice-Cloning/env/lib/python3.6/site-packages/tensorflow/python/training/saver.py:1276: checkpoint_exists (from tensorflow.python.training.checkpoint_management) is deprecated and will be removed in a future version.
Instructions for updating:
Use standard file APIs to check for files with this prefix.
Testing the vocoder...
All test passed! You can now synthesize speech.
This is a GUI-less example of interface to SV2TTS. The purpose of this script is to show how you can interface this project easily with your own. See the source code for an explanation of what is happening.
Interactive generation loop

With "--low_mem" flag same cuda init error.
Output:
https://pastebin.com/PFBEbEnz

[PhysWiz314]

I still run into the same issue as well with the --low_mem flag.

[haoqizhenhao]

When you get stucked on "Interactive generation loop", you should input your own audio file, then the code will work, you can read the source code . But I don't know why the message didn't work.

[ahmed-menshawy]

I have the same error with 2.1G GPU!

[CorentinJ]

I can't reproduce this unfortunately, I'm running a win10 install on two computers (one with 2gb VRAM) and tensorflow-gpu=1.13.1. I'm going to need more information to look into it.

[PhysWiz314]

I am running Linux on both computers. Ubuntu 18.04 with nvidia-430 and Cuda 10.2 (2gb VRAM) and Linux Mint 19.1 with nvidia-410 and Cuda 10.0 (4gb VRAM) on the other.

I have been playing around a bit with the code and it seems that in vocoder/inference.py (almost) any call to numba.cuda results in the error.

[lilydjwg]

I've found out why and how to fix it: multiprocess uses forked workers by default, which inherits some state CUDA isn't expected. Switching to spawned workers fixes it.

patch

diff --git a/synthesizer/inference.py b/synthesizer/inference.py
index 99fb778..b9cc9c0 100644
--- a/synthesizer/inference.py
+++ b/synthesizer/inference.py
@@ -2,12 +2,12 @@ from synthesizer.tacotron2 import Tacotron2
 from synthesizer.hparams import hparams
 from multiprocess.pool import Pool  # You're free to use either one
 #from multiprocessing import Pool   # 
+from multiprocess.context import SpawnContext
 from synthesizer import audio
 from pathlib import Path
 from typing import Union, List
 import tensorflow as tf
 import numpy as np
-import numba.cuda
 import librosa


@@ -80,13 +80,15 @@ class Synthesizer:
             # Low memory inference mode: load the model upon every request. The model has to be 
             # loaded in a separate process to be able to release GPU memory (a simple workaround 
             # to tensorflow's intricacies)
-            specs, alignments = Pool(1).starmap(Synthesizer._one_shot_synthesize_spectrograms, 
-                                                [(self.checkpoint_fpath, embeddings, texts)])[0]
+            specs, alignments = Pool(1, context=SpawnContext()
+                                    ).starmap(Synthesizer._one_shot_synthesize_spectrograms, 
+                                              [(self.checkpoint_fpath, embeddings, texts)])[0]

         return (specs, alignments) if return_alignments else specs

     @staticmethod
     def _one_shot_synthesize_spectrograms(checkpoint_fpath, embeddings, texts):
+        import numba.cuda
         # Load the model and forward the inputs
         tf.reset_default_graph()
         model = Tacotron2(checkpoint_fpath, hparams)

[CorentinJ]

I've implemented it in a rush on the low_mem_fix branch, can you try and see what gives?

[PhysWiz314]

demo_cli.py now passes but I've run into another bug with demo_toolbox.py so I can't verify that one.

[PiotrDabrowskey]

I've got the same problem and can confirm that @lilydjwg's solution solved it.