AttributeError: module 'torch' has no attribute 'compiler' and other various issue about voicecraft

[ajayarora1235]

relax bro. make sure to setup your environment properly according to what's in the readme (torch 0.2.1, that particular version of audiocraft, etc.) this is because of how you installed audiocraft. make sure to install the correct version of it and especially your version of xformers. install xformers 0.22.0 and this fixes it.

[jasonppy]

Checkout quick start with docker, should works for windows https://github.com/jasonppy/VoiceCraft?tab=readme-ov-file#quickstart

[Sewlell]

An update here after I try Docker @jasonppy provided.

Yeah I think I am getting exaggerating that one after hours of troubleshooting and it just failed.

Let talk about the Docker first. It look pretty complete for Docker setup. Except the Issue 1 and 2 still remained during the process. So I switch back to the conventional one, the one without using Docker.

I managed to fix Issue 1 ( wrong folder order ), Issue 2 ( wrong order of environment setup ) and Issue 4 ( basically didn't support in Windows ) in my conventional setup.

But now, new issue, 2 actually.

1. source command didn't supposed to be in Windows
   You can pretty much get away with wget and I believe it is pretty much replicable with other command that supported Windows.

2. Issue 3 in-depth
   I recheck the package list with conda list for @ajayarora1235 's comment. Since I actually kinda reinstalled them with the correct order, so now my torch is 2.0.1 and transformers is 4.38.2.

The issue is, xformers remained 0.0.25.post1, not 0.0.22 according to the environment.yml.

I have already attempt wiping 0.0.25.post1 xformers with conda remove and pip uninstall, but it always directed to the 0.0.22 one and I had no idea. I think Python use 0.0.25.post1 xformers somehow for now the new Cell 7 and it will show up the error of AttributeError: module 'torch' has no attribute 'compiler'.

possibly the conda list make up the list for both base and voicecraft environment? So yeah so far I am still trying to pull out the 0.0.25.post1 xformers to make Cell 7 worked.

FOR THOSE WHO WANT TO KNOW THE SOLUTION FOR ISSUE 1, ISSUE 2 AND ISSUE 4***

Issue 1
Your folder must look like this composition in order to make inference_tts.ipynb detecting data folder when you run Cell 5 and other stuff

whatever thing you got your thing of
     I
     I - VoiceCraft
             I - data
             I - demo
             I - pretrained_models
             I - src ------------------------| - audiocraft
             I - z_scripts                         |  (things like audiocraft.egg-info)
             I - inference_tts.ipynb

first VoiceCraft folder is the folder you git clone https://github.com/jasonppy/VoiceCraft.git of

Issue 2

that thing, along with Cell 5's stuff ( below ), caused by installation issue. So once you fix your installation it should recognize words like data and model in VSCode.

 from data.tokenizer import (
    AudioTokenizer,
    TextTokenizer,
)

My investigation find that pip install -e git+https://github.com/facebookresearch/audiocraft.git@c5157b5bf14bf83449c17ea1eeb66c19fb4bc7f0#egg=audiocraft stuff will download 2.2.2 torch and latest version of pretty much everything else with torch. So what you do is, run that command first, then you run pip install torch==2.0.1, then winget install ffmpeg and then so on.

Issue 4

apt-get is an inexistant command for Windows, so you may need to find alternative for it. Fortunately for me and y'all who watching this, installation of ffmpeg can get get away with winget install ffmpeg, which is an alternative for Windows. And also installation of espeak-ng with downloading their latest release in (https://github.com/espeak-ng/espeak-ng) and adding PATH for espeak-ng manually (bootphon/phonemizer#44 (comment)).

Man this thing is getting me nightmares.

[Sewlell]

The hell... I just casually move on to other fork that got released recently.

When I casually pip install xformers==0.0.20 ( NOT 0.0.22 I mentioned earlier ) in Anaconda Prompt itself, not VSCode's end. It's able to uninstall 0.0.25.post1 xformers and overcome AttributeError: module 'torch' has no attribute 'compiler' in Cell 7.

Another error will be given instead :
AttributeError: module 'os' has no attribute 'uname'

I look into the web and cluster.py and find that, os.uname doesn't support Windows. And thank to some source I import platform the cluster.py and change uname = os.uname to uname = platform.uname.

And this error show up after this modification:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[2], [line 25](vscode-notebook-cell:?execution_count=2&line=25)
     [22](vscode-notebook-cell:?execution_count=2&line=22) phn2num = ckpt['phn2num']
     [24](vscode-notebook-cell:?execution_count=2&line=24) text_tokenizer = TextTokenizer(backend="espeak")
---> [25](vscode-notebook-cell:?execution_count=2&line=25) audio_tokenizer = AudioTokenizer(signature=encodec_fn, device=device)

File [c:\Users\PEY3C\TTS\VoiceCraft\data\tokenizer.py:110](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:110), in AudioTokenizer.__init__(self, device, signature)
    [104](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:104) def __init__(
    [105](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:105)     self,
    [106](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:106)     device: Any = None,
    [107](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:107)     signature = None
    [108](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:108) ) -> None:
    [109](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:109)     from audiocraft.solvers import CompressionSolver
--> [110](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:110)     model = CompressionSolver.model_from_checkpoint(signature)
    [111](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:111)     self.sample_rate = model.sample_rate
    [112](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:112)     self.channels = model.channels

File [c:\users\pey3c\tts\voicecraft\src\audiocraft\audiocraft\solvers\compression.py:287](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/solvers/compression.py:287), in CompressionSolver.model_from_checkpoint(checkpoint_path, device)
    [285](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/solvers/compression.py:285) logger = logging.getLogger(__name__)
    [286](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/solvers/compression.py:286) logger.info(f"Loading compression model from checkpoint: {checkpoint_path}")
--> [287](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/solvers/compression.py:287) _checkpoint_path = checkpoint.resolve_checkpoint_path(checkpoint_path, use_fsdp=False)
    [288](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/solvers/compression.py:288) assert _checkpoint_path is not None, f"Could not resolve compression model checkpoint path: {checkpoint_path}"
    [289](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/solvers/compression.py:289) state = checkpoint.load_checkpoint(_checkpoint_path)

File [c:\users\pey3c\tts\voicecraft\src\audiocraft\audiocraft\utils\checkpoint.py:68](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:68), in resolve_checkpoint_path(sig_or_path, name, use_fsdp)
     [56](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:56) def resolve_checkpoint_path(sig_or_path: tp.Union[Path, str], name: tp.Optional[str] = None,
     [57](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:57)                             use_fsdp: bool = False) -> tp.Optional[Path]:
     [58](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:58)     """Resolve a given checkpoint path for a provided dora sig or path.
     [59](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:59) 
     [60](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:60)     Args:
   (...)
     [66](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:66)         Path, optional: Resolved checkpoint path, if it exists.
     [67](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:67)     """
---> [68](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:68)     from audiocraft import train
     [69](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:69)     xps_root = train.main.dora.dir / 'xps'
     [70](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/checkpoint.py:70)     sig_or_path = str(sig_or_path)

File [c:\users\pey3c\tts\voicecraft\src\audiocraft\audiocraft\train.py:149](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/train.py:149)
    [144](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/train.py:144)         return
    [146](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/train.py:146)     return solver.run()
--> [149](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/train.py:149) main.dora.dir = AudioCraftEnvironment.get_dora_dir()
    [150](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/train.py:150) main._base_cfg.slurm = get_slurm_parameters(main._base_cfg.slurm)
    [152](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/train.py:152) if main.dora.shared is not None and not os.access(main.dora.shared, os.R_OK):

File [c:\users\pey3c\tts\voicecraft\src\audiocraft\audiocraft\environment.py:108](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:108), in AudioCraftEnvironment.get_dora_dir(cls)
    [103](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:103) @classmethod
    [104](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:104) def get_dora_dir(cls) -> Path:
    [105](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:105)     """Gets the path to the dora directory for the current team and cluster.
    [106](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:106)     Value is overridden by the AUDIOCRAFT_DORA_DIR env var.
    [107](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:107)     """
--> [108](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:108)     cluster_config = cls.instance()._get_cluster_config()
    [109](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:109)     dora_dir = os.getenv("AUDIOCRAFT_DORA_DIR", cluster_config["dora_dir"])
    [110](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:110)     logger.warning(f"Dora directory: {dora_dir}")

File [c:\users\pey3c\tts\voicecraft\src\audiocraft\audiocraft\environment.py:81](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:81), in AudioCraftEnvironment.instance(cls)
     [78](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:78) @classmethod
     [79](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:79) def instance(cls):
     [80](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:80)     if cls._instance is None:
---> [81](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:81)         cls._instance = cls()
     [82](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:82)     return cls._instance

File [c:\users\pey3c\tts\voicecraft\src\audiocraft\audiocraft\environment.py:52](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:52), in AudioCraftEnvironment.__init__(self)
     [50](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:50) """Loads configuration."""
     [51](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:51) self.team: str = os.getenv("AUDIOCRAFT_TEAM", self.DEFAULT_TEAM)
---> [52](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:52) cluster_type = _guess_cluster_type()
     [53](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:53) cluster = os.getenv(
     [54](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:54)     "AUDIOCRAFT_CLUSTER", cluster_type.value
     [55](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:55) )
     [56](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/environment.py:56) logger.info("Detecting cluster type %s", cluster_type)

File [c:\users\pey3c\tts\voicecraft\src\audiocraft\audiocraft\utils\cluster.py:31](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/cluster.py:31), in _guess_cluster_type()
     [29](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/cluster.py:29) uname = platform.uname()
     [30](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/cluster.py:30) fqdn = socket.getfqdn()
---> [31](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/cluster.py:31) if uname.sysname == "Linux" and (uname.release.endswith("-aws") or ".ec2" in fqdn):
     [32](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/cluster.py:32)     return ClusterType.AWS
     [34](file:///C:/users/pey3c/tts/voicecraft/src/audiocraft/audiocraft/utils/cluster.py:34) if fqdn.endswith(".fair"):

AttributeError: 'uname_result' object has no attribute 'sysname'

So, OK.... I think the lore thickened. Probably make me feel like this is more certainly designed for Linux than Windows.

[jasonppy]

Thanks for your efforts, I'm unable to test issues regarding windows, but the docker solution seems to work for some people. Thanks for the feedback on audiocraft installation, I have made changes in 991b1fe

[Sewlell]

That's make sense. I got to solve AttributeError: module 'os' has no attribute 'uname' with new .ipynb and my composition.

But I got this error in the same Cell 7.

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
Cell In[4], [line 25](vscode-notebook-cell:?execution_count=4&line=25)
     [22](vscode-notebook-cell:?execution_count=4&line=22) phn2num = ckpt['phn2num']
     [24](vscode-notebook-cell:?execution_count=4&line=24) text_tokenizer = TextTokenizer(backend="espeak")
---> [25](vscode-notebook-cell:?execution_count=4&line=25) audio_tokenizer = AudioTokenizer(signature=encodec_fn, device=device) # will also put the neural codec model on gpu

File [c:\Users\PEY3C\TTS\VoiceCraft\data\tokenizer.py:109](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:109), in AudioTokenizer.__init__(self, device, signature)
    [104](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:104) def __init__(
    [105](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:105)     self,
    [106](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:106)     device: Any = None,
    [107](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:107)     signature = None
    [108](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:108) ) -> None:
--> [109](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:109)     from audiocraft.solvers import CompressionSolver
    [110](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:110)     model = CompressionSolver.model_from_checkpoint(signature)
    [111](file:///C:/Users/PEY3C/TTS/VoiceCraft/data/tokenizer.py:111)     self.sample_rate = model.sample_rate

ModuleNotFoundError: No module named 'audiocraft'

Final Cell when you are going to generate

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
Cell In[11], [line 29](vscode-notebook-cell:?execution_count=11&line=29)
     [27](vscode-notebook-cell:?execution_count=11&line=27) decode_config = {'top_k': top_k, 'top_p': top_p, 'temperature': temperature, 'stop_repetition': stop_repetition, 'kvcache': kvcache, "codec_audio_sr": codec_audio_sr, "codec_sr": codec_sr, "silence_tokens": silence_tokens, "sample_batch_size": sample_batch_size}
     [28](vscode-notebook-cell:?execution_count=11&line=28) from inference_tts_scale import inference_one_sample
---> [29](vscode-notebook-cell:?execution_count=11&line=29) concated_audio, gen_audio = inference_one_sample(model, ckpt["config"], phn2num, text_tokenizer, audio_tokenizer, audio_fn, target_transcript, device, decode_config, prompt_end_frame)
     [31](vscode-notebook-cell:?execution_count=11&line=31) # save segments for comparison
     [32](vscode-notebook-cell:?execution_count=11&line=32) concated_audio, gen_audio = concated_audio[0].cpu(), gen_audio[0].cpu()

NameError: name 'audio_tokenizer' is not defined

The rest is fine so far with sample .wav.

I didn't have time to test more in depth now because school. But yeah I will still assist.

[Sewlell]

ModuleNotFoundError: No module named 'audiocraft' happened because I didn't installed the audiocraft lol. And it worked all the way down to generation now.

I am still looking on generation speed and issue stuff. But somehow I realized the inference_tts.ipynb didn't do the TTS thing, but do the speech editing.

Will elaborate more in another new issue once I done playing and collecting more information/issues with this and other forks.