NER Training with Frozen Tok2Vec causes "The Tok2Vec listener did not receive any valid input from an upstream component."

[oliviercwa]

How to reproduce the behaviour

• Create a basic configuration for NER training with en_core_web_md
• Set frozen_components = ["tok2vec"] in the configuration
• Start training using the CLI command

Result

  File "spacy\spacy\training\loop.py", line 122, in train
    raise e
  File "spacy\spacy\training\loop.py", line 105, in train
    for batch, info, is_best_checkpoint in training_step_iterator:
  File "spacy\spacy\training\loop.py", line 209, in train_while_improving
    annotates=annotating_components,
  File "spacy\spacy\language.py", line 1123, in update
    proc.update(examples, sgd=None, losses=losses, **component_cfg[name])
  File "spacy\pipeline\transition_parser.pyx", line 395, in spacy.pipeline.transition_parser.Parser.update
  File "D:\src\zendoc\python\envs\spacy311\lib\site-packages\thinc\model.py", line 309, in begin_update
    return self._func(self, X, is_train=True)
  File "spacy\spacy\ml\tb_framework.py", line 38, in forward
    has_upper=model.attrs["has_upper"],
  File "spacy\ml\parser_model.pyx", line 216, in spacy.ml.parser_model.ParserStepModel.__init__
  File "D:\src\zendoc\python\envs\spacy311\lib\site-packages\thinc\model.py", line 291, in __call__
    return self._func(self, X, is_train=is_train)
  File "D:\src\zendoc\python\envs\spacy311\lib\site-packages\thinc\layers\chain.py", line 54, in forward
    Y, inc_layer_grad = layer(X, is_train=is_train)
  File "D:\src\zendoc\python\envs\spacy311\lib\site-packages\thinc\model.py", line 291, in __call__
    return self._func(self, X, is_train=is_train)
  File "spacy\spacy\pipeline\tok2vec.py", line 285, in forward
    model.verify_inputs(inputs)
  File "spacy\spacy\pipeline\tok2vec.py", line 273, in verify_inputs
    raise ValueError(Errors.E954)
ValueError: [E954] The Tok2Vec listener did not receive any valid input from an upstream component.

Expected
Be able to train a NER without affecting the Tok2Vec mappings

Your Environment

• Operating System: Windows 10
• Python Version Used: 3.7.4
• spaCy Version Used: 3.1.2
• Environment Information:

Scenario

I am trying to train a NER model starting from the en_core_web_md model. Once done, I am creating a pipeline that uses the newly trained NER model, plus the Tagger, DependencyParser, attribute_ruler and other components from the original en_core_web_md.
However if the vector mappings are changing, I am worried that these components may work properly.

[adrianeboyd]

If you put the tok2vec in frozen_components, it's not run at all during training. If you want to train a component that depends on its output, then you also need to add tok2vec to annotating_components so that it's run in order to set doc.tensors during training. See: https://spacy.io/usage/training#annotating-components

annotating_components requires spaCy v3.1+.