Add option to replace listeners for sourced components

spacy/language.py

@@ -8,7 +8,7 @@ from contextlib import contextmanager
 from copy import deepcopy
 from pathlib import Path
 import warnings
-from thinc.api import Model, get_current_ops, Config, Optimizer
+from thinc.api import get_current_ops, Config, Optimizer
 import srsly
 import multiprocessing as mp
 from itertools import chain, cycle
@@ -670,6 +670,47 @@ class Language:
         self._pipe_configs[name] = filled
         return resolved[factory_name]
 
+    def replace_listeners(
+        self,
+        tok2vec_name: str,
+        pipe_name: str,
+        listeners: Iterable[str] = SimpleFrozenList(),
+    ):
+        if tok2vec_name not in self.pipe_names:
+            raise ValueError  # TODO:
+        if pipe_name not in self.pipe_names:
+            raise ValueError  # TODO:
+        tok2vec = self.get_pipe(tok2vec_name)
+        tok2vec_cfg = self.get_pipe_config(tok2vec_name)
+        if (
+            not hasattr(tok2vec, "model")
+            or not hasattr(tok2vec, "listener_map")
+            or "model" not in tok2vec_cfg
+        ):
+            raise ValueError  # TODO: likely bug in spaCy if this happens
+        pipe_listeners = tok2vec.listener_map.get(pipe_name, [])
+        pipe_cfg = self._pipe_configs[pipe_name]
+        if listeners:
+            util.logger.debug(f"Replacing listeners of component '{pipe_name}'")
+            if len(listeners) != len(pipe_listeners):
+                # The number of listeners defined in the component model doesn't
+                # match the listeners to replace, so we won't be able to update
+                # the nodes and generate a matching config
+                raise ValueError(f"{listeners}, {pipe_listeners}")  # TODO:
+            pipe = self.get_pipe(pipe_name)
+            # Go over the listener layers and replace them
+            for listener in pipe_listeners:
+                util.replace_model_node(pipe.model, listener, tok2vec.model.copy())
+            # Update the config accordingly by coping the tok2vec model to all
+            # sections defined in the listener paths
+            for listener_path in listeners:
+                # Check if the path actually exists in the config
+                try:
+                    util.dot_to_object(pipe_cfg, listener_path)
+                except KeyError:
+                    raise ValueError  # TODO:
+                util.set_dot_to_object(pipe_cfg, listener_path, tok2vec_cfg["model"])
+
     def create_pipe_from_source(
         self, source_name: str, source: "Language", *, name: str
     ) -> Tuple[Callable[[Doc], Doc], str]:

spacy/tests/pipeline/test_tok2vec.py

@@ -1,5 +1,4 @@
 import pytest
-
 from spacy.ml.models.tok2vec import build_Tok2Vec_model
 from spacy.ml.models.tok2vec import MultiHashEmbed, CharacterEmbed
 from spacy.ml.models.tok2vec import MishWindowEncoder, MaxoutWindowEncoder
@@ -9,12 +8,11 @@ from spacy.tokens import Doc
 from spacy.training import Example
 from spacy import util
 from spacy.lang.en import English
-from ..util import get_batch
-
 from thinc.api import Config
-
 from numpy.testing import assert_equal
 
+from ..util import get_batch
+
 
 def test_empty_doc():
     width = 128
@@ -187,3 +185,29 @@ def test_tok2vec_listener_callback():
     Y, get_dX = tagger.model.begin_update(docs)
     # assure that the backprop call works (and doesn't hit a 'None' callback)
     assert get_dX(Y) is not None
+
+
+def test_replace_listeners():
+    orig_config = Config().from_str(cfg_string)
+    nlp = util.load_model_from_config(orig_config, auto_fill=True, validate=True)
+    examples = [Example.from_dict(nlp.make_doc("x y"), {"tags": ["V", "Z"]})]
+    nlp.initialize(lambda: examples)
+    tok2vec = nlp.get_pipe("tok2vec")
+    tagger = nlp.get_pipe("tagger")
+    assert isinstance(tagger.model.layers[0], Tok2VecListener)
+    assert tok2vec.listener_map["tagger"][0] == tagger.model.layers[0]
+    assert nlp.config["components"]["tok2vec"]["model"]["@architectures"] == "spacy.Tok2Vec.v2"
+    assert nlp.config["components"]["tagger"]["model"]["tok2vec"]["@architectures"] == "spacy.Tok2VecListener.v1"
+    nlp.replace_listeners("tok2vec", "tagger", ["model.tok2vec"])
+    assert not isinstance(tagger.model.layers[0], Tok2VecListener)
+    t2v_cfg = nlp.config["components"]["tok2vec"]["model"]
+    assert t2v_cfg["@architectures"] == "spacy.Tok2Vec.v2"
+    assert nlp.config["components"]["tagger"]["model"]["tok2vec"] == t2v_cfg
+    with pytest.raises(ValueError):
+        nlp.replace_listeners("invalid", "tagger", ["model.tok2vec"])
+    with pytest.raises(ValueError):
+        nlp.replace_listeners("tok2vec", "parser", ["model.tok2vec"])
+    with pytest.raises(ValueError):
+        nlp.replace_listeners("tok2vec", "tagger", ["model.yolo"])
+    with pytest.raises(ValueError):
+        nlp.replace_listeners("tok2vec", "tagger", ["model.tok2vec", "model.yolo"])

spacy/tests/test_misc.py

@@ -205,6 +205,25 @@ def test_dot_to_dict(dot_notation, expected):
     assert util.dict_to_dot(result) == dot_notation
 
 
+def test_set_dot_to_object():
+    config = {"foo": {"bar": 1, "baz": {"x": "y"}}, "test": {"a": {"b": "c"}}}
+    with pytest.raises(KeyError):
+        util.set_dot_to_object(config, "foo.bar.baz", 100)
+    with pytest.raises(KeyError):
+        util.set_dot_to_object(config, "hello.world", 100)
+    with pytest.raises(KeyError):
+        util.set_dot_to_object(config, "test.a.b.c", 100)
+    util.set_dot_to_object(config, "foo.bar", 100)
+    assert config["foo"]["bar"] == 100
+    util.set_dot_to_object(config, "foo.baz.x", {"hello": "world"})
+    assert config["foo"]["baz"]["x"]["hello"] == "world"
+    assert config["test"]["a"]["b"] == "c"
+    util.set_dot_to_object(config, "foo", 123)
+    assert config["foo"] == 123
+    util.set_dot_to_object(config, "test", "hello")
+    assert dict(config) == {"foo": 123, "test": "hello"}
+
+
 @pytest.mark.parametrize(
     "doc_sizes, expected_batches",
     [

spacy/training/initialize.py

@@ -1,4 +1,4 @@
-from typing import Union, Dict, Optional, Any, List, IO, TYPE_CHECKING
+from typing import Union, Dict, Optional, Any, IO, TYPE_CHECKING
 from thinc.api import Config, fix_random_seed, set_gpu_allocator
 from thinc.api import ConfigValidationError
 from pathlib import Path
@@ -33,7 +33,7 @@ def init_nlp(config: Config, *, use_gpu: int = -1) -> "Language":
     if use_gpu >= 0 and allocator:
         set_gpu_allocator(allocator)
     # Use original config here before it's resolved to functions
-    sourced_components = get_sourced_components(config)
+    sourced = get_sourced_components(config)
     nlp = load_model_from_config(raw_config, auto_fill=True)
     logger.info("Set up nlp object from config")
     config = nlp.config.interpolate()
@@ -57,7 +57,7 @@ def init_nlp(config: Config, *, use_gpu: int = -1) -> "Language":
     # Components that shouldn't be updated during training
     frozen_components = T["frozen_components"]
     # Sourced components that require resume_training
-    resume_components = [p for p in sourced_components if p not in frozen_components]
+    resume_components = [p for p in sourced if p not in frozen_components]
     logger.info(f"Pipeline: {nlp.pipe_names}")
     if resume_components:
         with nlp.select_pipes(enable=resume_components):
@@ -68,10 +68,18 @@ def init_nlp(config: Config, *, use_gpu: int = -1) -> "Language":
         logger.info(f"Initialized pipeline components: {nlp.pipe_names}")
     # Detect components with listeners that are not frozen consistently
     for name, proc in nlp.pipeline:
-        if getattr(proc, "listening_components", None):
+        if getattr(proc, "listening_components", None):  # e.g. tok2vec/transformer
             for listener in proc.listening_components:
-                if listener in frozen_components and name not in frozen_components:
+                # If it's a component sourced from another pipeline, we check if
+                # the tok2vec listeners should be replaced with standalone tok2vec
+                # models (e.g. so component can be frozen without its performance
+                # degrading when other components/tok2vec are updated)
+                paths = sourced.get(listener, {}).get("replace_listeners", [])
+                if paths:
+                    nlp.replace_listeners(name, listener, paths)
+                elif listener in frozen_components and name not in frozen_components:
                     logger.warning(Warnings.W087.format(name=name, listener=listener))
+                # We always check this regardless, in case user freezes tok2vec
                 if listener not in frozen_components and name in frozen_components:
                     logger.warning(Warnings.W086.format(name=name, listener=listener))
     return nlp
@@ -173,16 +181,18 @@ def init_tok2vec(
     return False
 
 
-def get_sourced_components(config: Union[Dict[str, Any], Config]) -> List[str]:
+def get_sourced_components(
+    config: Union[Dict[str, Any], Config]
+) -> Dict[str, Dict[str, Any]]:
     """RETURNS (List[str]): All sourced components in the original config,
     e.g. {"source": "en_core_web_sm"}. If the config contains a key
     "factory", we assume it refers to a component factory.
     """
-    return [
-        name
+    return {
+        name: cfg
         for name, cfg in config.get("components", {}).items()
         if "factory" not in cfg and "source" in cfg
-    ]
+    }
 
 
 def convert_vectors(

spacy/util.py

@@ -8,7 +8,7 @@ import re
 from pathlib import Path
 import thinc
 from thinc.api import NumpyOps, get_current_ops, Adam, Config, Optimizer
-from thinc.api import ConfigValidationError
+from thinc.api import ConfigValidationError, Model
 import functools
 import itertools
 import numpy.random
@@ -738,6 +738,24 @@ def get_package_path(name: str) -> Path:
     return Path(pkg.__file__).parent
 
 
+def replace_model_node(model: Model, target: Model, replacement: Model) -> None:
+    """Replace a node within a model with a new one, updating refs.
+
+    model (Model): The parent model.
+    target (Model): The target node.
+    replacement (Model): The node to replace the target with.
+    """
+    # Place the node into the sublayers
+    for node in model.walk():
+        if target in node.layers:
+            node.layers[node.layers.index(target)] = replacement
+    # Now fix any node references
+    for node in model.walk():
+        for ref_name in node.ref_names:
+            if node.maybe_get_ref(ref_name) is target:
+                node.set_ref(ref_name, replacement)
+
+
 def split_command(command: str) -> List[str]:
     """Split a string command using shlex. Handles platform compatibility.
 
@@ -1279,6 +1297,25 @@ def dot_to_object(config: Config, section: str):
     return component
 
 
+def set_dot_to_object(config: Config, section: str, value: Any) -> None:
+    """Update a config at a given position from a dot notation.
+
+    config (Config): The config.
+    section (str): The dot notation of the section in the config.
+    value (Any): The value to set in the config.
+    """
+    component = config
+    parts = section.split(".")
+    for i, item in enumerate(parts):
+        try:
+            if i == len(parts) - 1:
+                component[item] = value
+            else:
+                component = component[item]
+        except (KeyError, TypeError):
+            raise KeyError(Errors.E952.format(name=section)) from None
+
+
 def walk_dict(
     node: Dict[str, Any], parent: List[str] = []
 ) -> Iterator[Tuple[List[str], Any]]: