Component 'relation_extractor' could not be run. Did you forget to call initialize()?

[V20936181]

Hello,

I am trying to implement a custom pipeline component. I have cloned the project template 'rel_component' found here. I have added this component using the 'add_pipe' command as follows:

import spacy
# make the factory work
from rel_component.scripts.rel_pipe import make_relation_extractor

# make the config work
from rel_component.scripts.rel_model import create_relation_model, create_classification_layer, create_instances, create_tensors
from thinc.api import Config

config_str = """
[components]
[components.relation_extractor]
threshold = 0.5

[components.relation_extractor.model]
@architectures = "rel_model.v1"

[components.relation_extractor.model.create_instance_tensor]
@architectures = "rel_instance_tensor.v1"

[components.relation_extractor.model.create_instance_tensor.tok2vec]
@architectures = "spacy.Tok2VecListener.v1"
width = 96

[components.relation_extractor.model.create_instance_tensor.pooling]
@layers = "reduce_mean.v1"

[components.relation_extractor.model.create_instance_tensor.get_instances]
@misc = "rel_instance_generator.v1"
max_length = 100

[components.relation_extractor.model.classification_layer]
@architectures = "rel_classification_layer.v1"
nI = null
nO = null
"""


config = Config().from_str(config_str)
rel_config = config["components"]["relation_extractor"]
nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("relation_extractor", config=rel_config)
print(print(nlp.pipe_names))

This works fine and I get the following output:
['tok2vec', 'tagger', 'parser', 'attribute_ruler', 'lemmatizer', 'ner', 'relation_extractor']

But when I input the text:

text = "Expression of the g-aminobutyric acid (GABA) transporter GAT-3 was selectively decreased within the amygdala of alcohol-choosing rats, whereas a knockdown of this transcript reversed choice preference of rats that originally chose a sweet solution over alcohol."

doc = nlp(text)

I get the following error:

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
File ~/.local/lib/python3.9/site-packages/spacy/language.py:1042, in __call__(self, text, disable, component_cfg)
   1036 def disable_pipes(self, *names) -> "DisabledPipes":
   1037     """Disable one or more pipeline components. If used as a context
   1038     manager, the pipeline will be restored to the initial state at the end
   1039     of the block. Otherwise, a DisabledPipes object is returned, that has
   1040     a `.restore()` method you can use to undo your changes.
   1041 
-> 1042     This method has been deprecated since 3.0
   1043     """
   1044     warnings.warn(Warnings.W096, DeprecationWarning)

File ~/rel_component/scripts/rel_pipe.py:80, in RelationExtractor.__call__(self, doc)
     78     return doc
---> 80 predictions = self.predict([doc])
     81 self.set_annotations([doc], predictions)

File ~/rel_component/scripts/rel_pipe.py:90, in RelationExtractor.predict(self, docs)
     89     msg.info("Could not determine any instances in any docs - can not make any predictions.")
---> 90 scores = self.model.predict(docs)
     91 return self.model.ops.asarray(scores)

File ~/.local/lib/python3.9/site-packages/thinc/model.py:334, in Model.predict(self, X)
    331 """Call the model's `forward` function with `is_train=False`, and return
    332 only the output, instead of the `(output, callback)` tuple.
    333 """
--> 334 return self._func(self, X, is_train=False)[0]

File ~/.local/lib/python3.9/site-packages/thinc/layers/chain.py:54, in forward(model, X, is_train)
     53 for layer in model.layers:
---> 54     Y, inc_layer_grad = layer(X, is_train=is_train)
     55     callbacks.append(inc_layer_grad)

File ~/.local/lib/python3.9/site-packages/thinc/model.py:310, in Model.__call__(self, X, is_train)
    308 """Call the model's `forward` function, returning the output and a
    309 callback to compute the gradients via backpropagation."""
--> 310 return self._func(self, X, is_train=is_train)

File ~/.local/lib/python3.9/site-packages/thinc/layers/chain.py:54, in forward(model, X, is_train)
     53 for layer in model.layers:
---> 54     Y, inc_layer_grad = layer(X, is_train=is_train)
     55     callbacks.append(inc_layer_grad)

File ~/.local/lib/python3.9/site-packages/thinc/model.py:310, in Model.__call__(self, X, is_train)
    308 """Call the model's `forward` function, returning the output and a
    309 callback to compute the gradients via backpropagation."""
--> 310 return self._func(self, X, is_train=is_train)

File ~/.local/lib/python3.9/site-packages/thinc/layers/linear.py:36, in forward(model, X, is_train)
     35 def forward(model: Model[InT, OutT], X: InT, is_train: bool) -> Tuple[OutT, Callable]:
---> 36     W = cast(Floats2d, model.get_param("W"))
     37     b = cast(Floats1d, model.get_param("b"))

File ~/.local/lib/python3.9/site-packages/thinc/model.py:235, in Model.get_param(self, name)
    234 if not self._params.has_param(self.id, name):
--> 235     raise KeyError(
    236         f"Parameter '{name}' for model '{self.name}' has not been allocated yet."
    237     )
    238 return self._params.get_param(self.id, name)

KeyError: "Parameter 'W' for model 'linear' has not been allocated yet."

The above exception was the direct cause of the following exception:

ValueError                                Traceback (most recent call last)
Input In [265], in <cell line: 2>()
      1 text = "Expression of the g-aminobutyric acid (GABA) transporter GAT-3 was selectively decreased within the amygdala of alcohol-choosing rats, whereas a knockdown of this transcript reversed choice preference of rats that originally chose a sweet solution over alcohol."
----> 2 doc = nlp(text)

File ~/.local/lib/python3.9/site-packages/spacy/language.py:1045, in __call__(self, text, disable, component_cfg)
   1037 """Disable one or more pipeline components. If used as a context
   1038 manager, the pipeline will be restored to the initial state at the end
   1039 of the block. Otherwise, a DisabledPipes object is returned, that has
   (...)
   1042 This method has been deprecated since 3.0
   1043 """
   1044 warnings.warn(Warnings.W096, DeprecationWarning)
-> 1045 if len(names) == 1 and isinstance(names[0], (list, tuple)):
   1046     names = names[0]  # type: ignore[assignment]    # support list of names instead of spread
   1047 return self.select_pipes(disable=names)

ValueError: [E109] Component 'relation_extractor' could not be run. Did you forget to call 'initialize()'?

Have I missed any steps in this implementation? Please help.

Package versions:
Python: 3.9.12
SpaCy: 3.6.1
Thinc: 8.1.12

[danieldk]

When you are adding the new pipe to the existing pipeline:

nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("relation_extractor", config=rel_config)

the model of the relation_extractor pipe has not been initialized yet. So, you need to do something along the lines of:

nlp = spacy.load("en_core_web_sm")
relation_extractor = nlp.add_pipe("relation_extractor", config=rel_config)
relation_extractor.initialize(lamda: examples)

where examples is a list of training examples (List[Example]). Also see: https://spacy.io/api/pipe#initialize

An unrelated tip about your configuration:

[components.relation_extractor.model.create_instance_tensor.tok2vec]
@architectures = "spacy.Tok2VecListener.v1"
width = 96

This will use the existing tok2vec pipe of the en_core_web_sm pipeline. If you would train on relation extraction without also continuing to train the other components, the tok2vec pipe will only be updated for relation extraction and performance of the other pipeline components will deteriorate. If you want to keep reusing the existing pipeline components (e.g. part-of-speech tagging or dependency parsing), it's best to give the relation extractor its own tok2vec model and not a listener.

[V20936181]

Thank you @danieldk I was able to add the pipeline by calling initialize method.
You are also right about the configuration changes. I disabled en_core_web_sm model's tok2vec component, and added rel_component's tok2vec instead. Was able to achieve much higher scores. Thank you for pointing that out!