Language.from_disk() does not load pipeline

[hertelm]

Loading a saved model via from_disk(...) does not load the pipeline, but the example from the docs suggests that I do not need to take care of this.

How to reproduce the behaviour

Minimum example:

import spacy
from spacy.lang.en import English


model = spacy.load("en_core_web_sm")
print("Original pipeline:", model.pipe_names)
model.to_disk("saved_model")
model = English().from_disk("saved_model")
print("Pipeline after loading from disk:", model.pipe_names)

The result is:

Original pipeline: ['tagger', 'parser', 'ner']
Pipeline after loading from disk: []

Your Environment

• spaCy version: 2.3.0
• Platform: Linux-4.15.0-108-generic-x86_64-with-Ubuntu-18.04-bionic
• Python version: 3.6.9

[adrianeboyd]

When you initialize a model with English() it doesn't have any pipeline components, so it doesn't know to load them when you call from_disk(). If you add the right pipeline components and then run from_disk(), it will deserialize any that are in the current pipeline. See the example here: https://spacy.io/usage/saving-loading#pipeline

There can be cases where you want to handle things differently, but in general you probably want to use spacy.load() to load models, since it takes care of these details for you. The docs are here: https://spacy.io/api/top-level#spacy.load and you can see what it's doing underneath here:

spaCy/spacy/util.py

Lines 194 to 222 in 2d71545

	def load_model_from_path(model_path, meta=False, **overrides):
	"""Load a model from a data directory path. Creates Language class with
	pipeline from meta.json and then calls from_disk() with path."""
	if not meta:
	meta = get_model_meta(model_path)
	# Support language factories registered via entry points (e.g. custom
	# language subclass) while keeping top-level language identifier "lang"
	lang = meta.get("lang_factory", meta["lang"])
	cls = get_lang_class(lang)
	nlp = cls(meta=meta, **overrides)
	pipeline = meta.get("pipeline", [])
	factories = meta.get("factories", {})
	disable = overrides.get("disable", [])
	if pipeline is True:
	pipeline = nlp.Defaults.pipe_names
	elif pipeline in (False, None):
	pipeline = []
	# skip "vocab" from overrides in component initialization since vocab is
	# already configured from overrides when nlp is initialized above
	if "vocab" in overrides:
	del overrides["vocab"]
	for name in pipeline:
	if name not in disable:
	config = meta.get("pipeline_args", {}).get(name, {})
	config.update(overrides)
	factory = factories.get(name, name)
	component = nlp.create_pipe(factory, config=config)
	nlp.add_pipe(component, name=name)
	return nlp.from_disk(model_path, exclude=disable)