Fix linking resumed components (#6859)

* link components across enabled, resumed and frozen

* revert renaming

* revert renaming, the sequel

spacy/language.py

@@ -1190,6 +1190,7 @@ class Language:
         get_examples: Optional[Callable[[], Iterable[Example]]] = None,
         *,
         sgd: Optional[Optimizer] = None,
+        link_components: bool = True,
     ) -> Optimizer:
         """Initialize the pipe for training, using data examples if available.
 
@@ -1197,6 +1198,8 @@ class Language:
             returns gold-standard Example objects.
         sgd (Optional[Optimizer]): An optimizer to use for updates. If not
             provided, will be created using the .create_optimizer() method.
+        link_components (bool): Link listener components automatically or not
+            (default True)
         RETURNS (thinc.api.Optimizer): The optimizer.
 
         DOCS: https://spacy.io/api/language#initialize
@@ -1244,7 +1247,8 @@ class Language:
                     proc.initialize, p_settings, section="components", name=name
                 )
                 proc.initialize(get_examples, nlp=self, **p_settings)
-        self._link_components()
+        if link_components:
+            self._link_components()
         self._optimizer = sgd
         if sgd is not None:
             self._optimizer = sgd
@@ -1528,7 +1532,7 @@ class Language:
         """Register 'listeners' within pipeline components, to allow them to
         effectively share weights.
         """
-        # I had though, "Why do we do this inside the Language object? Shouldn't
+        # I had thought, "Why do we do this inside the Language object? Shouldn't
         # it be the tok2vec/transformer/etc's job?
         # The problem is we need to do it during deserialization...And the
         # components don't receive the pipeline then. So this does have to be

spacy/training/initialize.py

@@ -64,8 +64,10 @@ def init_nlp(config: Config, *, use_gpu: int = -1) -> "Language":
         with nlp.select_pipes(enable=resume_components):
             logger.info(f"Resuming training for: {resume_components}")
             nlp.resume_training(sgd=optimizer)
+    # Make sure that listeners are defined before initializing further
+    nlp._link_components()
     with nlp.select_pipes(disable=[*frozen_components, *resume_components]):
-        nlp.initialize(lambda: train_corpus(nlp), sgd=optimizer)
+        nlp.initialize(lambda: train_corpus(nlp), sgd=optimizer, link_components=False)
         logger.info(f"Initialized pipeline components: {nlp.pipe_names}")
     # Detect components with listeners that are not frozen consistently
     for name, proc in nlp.pipeline: