Use activations stored by the senter pipe

Before this change, we'd use the senter pipe directly. However, this did
not work with the transformer model without modifications (because it
clears tensors after backprop). By using the functionality proposed in

explosion/spaCy#11002

we can use the activations that are stored by the senter pipe in `Doc`.

projects/biaffine_parser/configs/base-config.cfg

@@ -34,6 +34,7 @@ pooling = {"@layers":"reduce_mean.v1"}
 
 [components.arc_predicter]
 factory = "experimental_arc_predicter"
+senter = "senter"
 
 [components.arc_predicter.model]
 @architectures = "spacy-experimental.PairwiseBilinear.v1"
@@ -78,6 +79,7 @@ pooling = {"@layers":"reduce_mean.v1"}
 
 [components.senter]
 factory = "senter"
+store_activations = true
 
 [components.senter.model]
 @architectures = "spacy.Tagger.v1"

pyproject.toml

@@ -3,6 +3,6 @@ requires = [
     "setuptools",
     "wheel",
     "Cython<3.0",
-    "spacy>=3.3.0,<3.4.0",
+    "spacy>=3.4.0,<3.5.0",
     "numpy",
 ]

requirements.txt

@@ -1,4 +1,4 @@
-spacy>=3.3.0,<3.4.0
+spacy>=3.4.0,<3.5.0
 
 # Development dependencies
 cython>=0.25,<3.0

setup.cfg

@@ -13,7 +13,7 @@ long_description_content_type = text/markdown
 zip_safe = false
 python_requires = >=3.6
 install_requires =
-    spacy>=3.3.0,<3.4.0
+    spacy>=3.4.0,<3.5.0
 
 [options.entry_points]
 spacy_architectures =

spacy_experimental/biaffine_parser/arc_predicter.pyx

@@ -162,7 +162,7 @@ class ArcPredicter(TrainablePipe):
         docs = list(docs)
 
         if self.senter:
-            lengths = split_greedily(docs, ops=self.model.ops, max_length=self.max_length, senter=self.senter, is_train=False)
+            lengths = split_lazily(docs, ops=self.model.ops, max_length=self.max_length, senter=self.senter, is_train=False)
         else:
             lengths = sents2lens(docs, ops=self.model.ops)
         scores = self.model.predict((docs, lengths))
@@ -224,7 +224,7 @@ class ArcPredicter(TrainablePipe):
         docs = [eg.predicted for eg in examples]
 
         if self.senter:
-            lens = split_greedily(docs, ops=self.model.ops, max_length=self.max_length, senter=self.senter, is_train=True)
+            lens = split_lazily(docs, ops=self.model.ops, max_length=self.max_length, senter=self.senter, is_train=True)
         else:
             lens = sents2lens(docs, ops=self.model.ops)
         if lens.sum() == 0:
@@ -316,11 +316,13 @@ def sents2lens(docs: List[Doc], *, ops: Ops) -> Ints1d:
 
     return ops.asarray1i(lens)
 
-def split_greedily(docs: List[Doc], *, ops: Ops, max_length: int, senter: SentenceRecognizer, is_train: bool):
-    split_predictions, _ = senter.model(docs, is_train)
-
+def split_lazily(docs: List[Doc], *, ops: Ops, max_length: int, senter: SentenceRecognizer, is_train: bool) -> Ints1d:
     lens = []
-    for (doc, scores) in zip(docs, split_predictions):
+    for doc in docs:
+        activations = doc.activations.get(senter.name, None)
+        if activations is None:
+            raise ValueError("Greedy splitting requires senter with `store_activations` enabled.")
+        scores = activations['probs']
         split_recursive(scores[:,1], ops, max_length, lens)
 
     assert sum(lens) == sum([len(doc) for doc in docs])