spaCy loads "spacy-transformers" library even when using CPU model

[xxyzz]

Usually this is not a problem, but when running spaCy with Python optimize level set to 2 this would cause AttributeError: 'NoneType' object has no attribute 'format' error in the "transformers" package.

Error traceback:

Traceback (most recent call last):
  File "calibre\gui2\threaded_jobs.py", line 82, in start_work
  File "calibre_plugins.worddumb.parse_job", line 168, in do_job
  File "calibre_plugins.worddumb.parse_job", line 224, in create_files
  File "calibre_plugins.worddumb.parse_job", line 591, in load_spacy
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy\__init__.py", line 54, in load
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy\util.py", line 432, in load_model
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy\util.py", line 468, in load_model_from_package
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\en_core_web_lg\__init__.py", line 10, in load
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy\util.py", line 649, in load_model_from_init_py
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy\util.py", line 506, in load_model_from_path
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy\util.py", line 554, in load_model_from_config
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy\language.py", line 1788, in from_config
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy\language.py", line 163, in __init__
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\catalogue\__init__.py", line 119, in get_all
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\catalogue\__init__.py", line 134, in get_entry_points
  File "importlib\metadata\__init__.py", line 162, in load
  File "importlib\__init__.py", line 126, in import_module
  File "<frozen importlib._bootstrap>", line 1050, in _gcd_import
  File "<frozen importlib._bootstrap>", line 1027, in _find_and_load
  File "<frozen importlib._bootstrap>", line 992, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 241, in _call_with_frames_removed
  File "<frozen importlib._bootstrap>", line 1050, in _gcd_import
  File "<frozen importlib._bootstrap>", line 1027, in _find_and_load
  File "<frozen importlib._bootstrap>", line 1006, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 688, in _load_unlocked
  File "<frozen importlib._bootstrap_external>", line 883, in exec_module
  File "<frozen importlib._bootstrap>", line 241, in _call_with_frames_removed
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy_transformers\__init__.py", line 1, in <module>
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy_transformers\architectures.py", line 6, in <module>
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy_transformers\layers\__init__.py", line 1, in <module>
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy_transformers\layers\listener.py", line 4, in <module>
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\spacy_transformers\data_classes.py", line 5, in <module>
  File "C:\Users\xAppData\Roaming\calibre\plugins\worddumb-libs-py3.10\transformers\tokenization_utils.py", line 26, in <module>
  File "C:\Users\x\AppData\Roaming\calibre\plugins\worddumb-libs-py3.10\transformers\tokenization_utils_base.py", line 3646, in <module>
AttributeError: 'NoneType' object has no attribute 'format'

Could this behavior be changed to only load the spacy-transformer package when a transformer model is actually used?

Relevant code:

spaCy/spacy/language.py

Line 162 in df0cb4b

util.registry._entry_point_factories.get_all()

[adrianeboyd]

There's no way for spacy to use entry points like those for the additional architectures in spacy-transformers without loading the packages to some degree.

I would suggest filing an issue in the transformers repo because this could be updated to check for whether __doc__ is defined first, which they do in other places in the code. (I checked and couldn't find any related issues or PRs, but it's kind of hard to search for __doc__, so I might have missed something.)

(Also, due to issues with numpy and torch we haven't published spacy-transformers packages that support the very newest version of transformers yet, but I've tested newer releases of transformers and they were fine even if spacy-transformers complains about the version conflict.)

[xxyzz]

Thanks for your reply. I was kind surprised "transformers" code was called when loading a CPU model and thought maybe something is off. I'll create a issue there and hope it could be fixed soon.

[xxyzz]

Hi, transformers v4.26.0 is released. It fixes the doc string error, could it be supported in the next spacy-transformers and transformers models release?

[adrianeboyd]

Sure, we can do that. If we have a new patch release of spacy-transformers, the most recent models will automatically support it because they should have spacy-transformers>=1.2.0,<1.3.0.