ValueError: [E203] If the transformer embedding layer is not updated during training, make sure to include it in 'annotating components'

[Shreya-Goswami]

Hi,
So when i'm trying to run this code-

for epoch in range(10): 
    losses = {}
    batches = minibatch(train_data, size=compounding(4.0, 128.0, 1.001))
    for batch in batches:
        nlp.rehearse(batch, sgd=optimizer, losses=losses)
    scores = nlp.evaluate(dev_data)
    print(f"Scores at epoch during rehearse {epoch}: {scores}")
    print(f"losses at {epoch}: {losses}")

I'm getting the error - ValueError: [E203] If the transformer embedding layer is not updated during training, make sure to include it in 'annotating components'
I had followed the doc here - https://spacy.io/usage/training#annotating-components and had added annotating_components = ["transformer", "ner] in my config.cfg file under [training] subsection but it's still not working. This is my pipeline - pipeline = ["transformer","ner"]

This is my output -

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[10], [line 14](vscode-notebook-cell:?execution_count=10&line=14)
     [12](vscode-notebook-cell:?execution_count=10&line=12) batches = minibatch(train_data, size=compounding(4.0, 128.0, 1.001))
     [13](vscode-notebook-cell:?execution_count=10&line=13) for batch in batches:
---> [14](vscode-notebook-cell:?execution_count=10&line=14)     nlp.rehearse(batch, sgd=optimizer, losses=losses)
     [15](vscode-notebook-cell:?execution_count=10&line=15) scores = nlp.evaluate(dev_data)
     [16](vscode-notebook-cell:?execution_count=10&line=16) print(f"Scores at epoch during rehearse {epoch}: {scores}")

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/language.py:1235](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/language.py:1235), in Language.rehearse(self, examples, sgd, losses, component_cfg, exclude)
   [1233](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/language.py:1233)         continue
   [1234](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/language.py:1234)     grads = {}
-> [1235](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/language.py:1235)     proc.rehearse(  # type: ignore[attr-defined]
   [1236](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/language.py:1236)         examples, sgd=get_grads, losses=losses, **component_cfg.get(name, {})
   [1237](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/language.py:1237)     )
   [1238](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/language.py:1238) for key, (W, dW) in grads.items():
   [1239](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/language.py:1239)     sgd(key, W, dW)  # type: ignore[call-arg, misc]

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/pipeline/transition_parser.pyx:458](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/pipeline/transition_parser.pyx:458), in spacy.pipeline.transition_parser.Parser.rehearse()

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:328](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:328), in Model.begin_update(self, X)
    [321](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:321) def begin_update(self, X: InT) -> Tuple[OutT, Callable[[OutT], InT]]:
    [322](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:322)     """Run the model over a batch of data, returning the output and a
    [323](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:323)     callback to complete the backward pass. A tuple (Y, finish_update),
    [324](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:324)     where Y is a batch of output data, and finish_update is a callback that
    [325](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:325)     takes the gradient with respect to the output and an optimizer function,
    [326](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:326)     and returns the gradient with respect to the input.
    [327](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:327)     """
--> [328](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:328)     return self._func(self, X, is_train=True)

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:33](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:33), in forward(model, X, is_train)
     [32](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:32) def forward(model, X, is_train):
---> [33](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:33)     step_model = ParserStepModel(
     [34](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:34)         X,
     [35](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:35)         model.layers,
     [36](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:36)         unseen_classes=model.attrs["unseen_classes"],
     [37](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:37)         train=is_train,
     [38](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:38)         has_upper=model.attrs["has_upper"],
     [39](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:39)     )
     [41](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/tb_framework.py:41)     return step_model, step_model.finish_steps

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/parser_model.pyx:213](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy/ml/parser_model.pyx:213), in spacy.ml.parser_model.ParserStepModel.__init__()

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:310](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:310), in Model.__call__(self, X, is_train)
    [307](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:307) def __call__(self, X: InT, is_train: bool) -> Tuple[OutT, Callable]:
    [308](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:308)     """Call the model's `forward` function, returning the output and a
    [309](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:309)     callback to compute the gradients via backpropagation."""
--> [310](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:310)     return self._func(self, X, is_train=is_train)

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:54](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:54), in forward(model, X, is_train)
     [52](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:52) callbacks = []
     [53](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:53) for layer in model.layers:
---> [54](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:54)     Y, inc_layer_grad = layer(X, is_train=is_train)
     [55](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:55)     callbacks.append(inc_layer_grad)
     [56](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:56)     X = Y

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:310](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:310), in Model.__call__(self, X, is_train)
    [307](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:307) def __call__(self, X: InT, is_train: bool) -> Tuple[OutT, Callable]:
    [308](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:308)     """Call the model's `forward` function, returning the output and a
    [309](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:309)     callback to compute the gradients via backpropagation."""
--> [310](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:310)     return self._func(self, X, is_train=is_train)

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:54](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:54), in forward(model, X, is_train)
     [52](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:52) callbacks = []
     [53](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:53) for layer in model.layers:
---> [54](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:54)     Y, inc_layer_grad = layer(X, is_train=is_train)
     [55](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:55)     callbacks.append(inc_layer_grad)
     [56](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/layers/chain.py:56)     X = Y

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:310](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:310), in Model.__call__(self, X, is_train)
    [307](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:307) def __call__(self, X: InT, is_train: bool) -> Tuple[OutT, Callable]:
    [308](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:308)     """Call the model's `forward` function, returning the output and a
    [309](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:309)     callback to compute the gradients via backpropagation."""
--> [310](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/thinc/model.py:310)     return self._func(self, X, is_train=is_train)

File [/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy_transformers/layers/listener.py:68](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy_transformers/layers/listener.py:68), in forward(model, docs, is_train)
     [66](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy_transformers/layers/listener.py:66) for doc in docs:
     [67](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy_transformers/layers/listener.py:67)     if doc._.trf_data is None:
---> [68](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy_transformers/layers/listener.py:68)         raise ValueError(Errors.E203.format(name="transformer"))
     [69](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy_transformers/layers/listener.py:69)     else:
     [70](https://file+.vscode-resource.vscode-cdn.net/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/spacy_transformers/layers/listener.py:70)         outputs.append(doc._.trf_data)

ValueError: [E203] If the transformer embedding layer is not updated during training, make sure to include it in 'annotating components'

[svlandeg]

Hi, thanks for the report!

As I can't run your code, could you try something to confirm my suspicion? I think that the "annotating_components" feature is not actually compatible yet with the experimental rehearse function. If you just run training via nlp.update with the same configuration, does it run then?

[Shreya-Goswami]

Yes, nlp.update with the same configuration works. I actually wanted to avoid catastrophic forgetting in cases where old data is unavailable. For nlp.update, we would need old/revision data + new data for training. Also, can I not use rehearse since it is still in its experimental phase?

[svlandeg]

In cases where old data is unavailable, but the model's original performance on the old data was quite good (let's say >90% F), you could consider creating silver annotations with your model on texts resembling the old data - and then mixing that silver data in with the annotations on your new data. In general, I would advice to train from scratch (old+new data) instead of using the rehearsel functionality (which is indeed in experimental and as we found out, doesn't actually work with annotating components yet :/)

[Shreya-Goswami]

Okay, I think we'll go ahead with your recommendation of training our model from scratch. Thank you so much for your reply!