Module Not Found Error: No module named 's3prl'

[gooran]

After the new update, my code is broken and gives an error. When I build the code, I get this error:

|     import wespeaker
|   File "/usr/local/lib/python3.11/site-packages/wespeaker/__init__.py", line 1, in <module>
|     from wespeaker.cli.speaker import load_model  # noqa
|     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|   File "/usr/local/lib/python3.11/site-packages/wespeaker/cli/speaker.py", line 36, in <module>
|     from wespeaker.frontend import frontend_class_dict
|   File "/usr/local/lib/python3.11/site-packages/wespeaker/frontend/__init__.py", line 16, in <module>
|     from .s3prl import S3prlFrontend
|   File "/usr/local/lib/python3.11/site-packages/wespeaker/frontend/s3prl.py", line 19, in <module>
|     import s3prl
| ModuleNotFoundError: No module named 's3prl'
service exited with code 1 (restarting)

[gooran]

After adding the following line to the requirements.txt,

s3prl==0.4.18
git+https://github.com/wenet-e2e/wespeaker.git

I got new errors:

  |     import wespeaker
  |   File "/usr/local/lib/python3.11/site-packages/wespeaker/__init__.py", line 1, in <module>
  |     from wespeaker.cli.speaker import load_model  # noqa
  |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  |   File "/usr/local/lib/python3.11/site-packages/wespeaker/cli/speaker.py", line 36, in <module>
  |     from wespeaker.frontend import frontend_class_dict
  |   File "/usr/local/lib/python3.11/site-packages/wespeaker/frontend/__init__.py", line 16, in <module>
  |     from .s3prl import S3prlFrontend
  |   File "/usr/local/lib/python3.11/site-packages/wespeaker/frontend/s3prl.py", line 20, in <module>
  |     from s3prl.nn import Featurizer, S3PRLUpstream
  |   File "/usr/local/lib/python3.11/site-packages/s3prl/nn/__init__.py", line 12, in <module>
  |     from .upstream import Featurizer, S3PRLUpstream
  |   File "/usr/local/lib/python3.11/site-packages/s3prl/nn/upstream.py", line 14, in <module>
  |     from s3prl import hub
  |   File "/usr/local/lib/python3.11/site-packages/s3prl/hub.py", line 16, in <module>
  |     from s3prl.upstream.hf_hubert.hubconf import *
  |   File "/usr/local/lib/python3.11/site-packages/s3prl/upstream/hf_hubert/hubconf.py", line 1, in <module>
  |     from .expert import UpstreamExpert as _UpstreamExpert
  |   File "/usr/local/lib/python3.11/site-packages/s3prl/upstream/hf_hubert/expert.py", line 4, in <module>
  |     from transformers import HubertModel, Wav2Vec2FeatureExtractor
  |   File "/usr/local/lib/python3.11/site-packages/transformers/__init__.py", line 27, in <module>
  |     from . import dependency_versions_check
  |   File "/usr/local/lib/python3.11/site-packages/transformers/dependency_versions_check.py", line 16, in <module>
  |     from .utils.versions import require_version, require_version_core
  |   File "/usr/local/lib/python3.11/site-packages/transformers/utils/__init__.py", line 24, in <module>
  |     from .auto_docstring import (
  |   File "/usr/local/lib/python3.11/site-packages/transformers/utils/auto_docstring.py", line 30, in <module>
  |     from .generic import ModelOutput
  |   File "/usr/local/lib/python3.11/site-packages/transformers/utils/generic.py", line 465, in <module>
  |     _torch_pytree.register_pytree_node(
  |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  | AttributeError: module 'torch.utils._pytree' has no attribute 'register_pytree_node'. Did you mean: '_register_pytree_node'?
 exited with code 1 (restarting)

[gooran]

It would be better for the wespeaker library to provide an installable package with proper versioned releases (e.g., on PyPI) instead of requiring users to install directly from the Git repository via pip install git+https://github.com/wenet-e2e/wespeaker.git. Installing from a Git URL bypasses standard release management, makes dependency resolution and reproducibility harder, and complicates downstream usage in production, CI pipelines, and pinned environments. A cleaner approach is to define proper packaging metadata (pyproject.toml or setup.py), tag semantic versions (e.g., v0.0.1), and publish release artifacts so users can simply run pip install wespeaker==0.0.1, benefiting from version pinning, caching, and compatibility guarantees.

[gooran]

s3prl pulls in a newer version of transformers that relies on a PyTorch API (register_pytree_node) which is not compatible with PyTorch 2.1.2. For environments using PyTorch 2.1.2, transformers versions below 4.31.0 are known to be compatible.

To avoid this incompatibility, I added an explicit version constraint for transformers in requirements.txt:

transformers<4.31.0

After applying this constraint, a new issue appeared:

ModuleNotFoundError: No module named 'whisper'

To resolve this, I added the missing dependency explicitly:

openai-whisper

But it seems the errors never end:

ImportError: cannot import name 'Wav2Vec2BertModel' from 'transformers'

[cdliang11]

Thank you very much for your suggestion. We will fix this issue as soon as possible.