Attempt to change dimension 'nV' for model 'hashembed' from 2000 to 5000

[vrunm]

I am trying to load my custom model from spacy amd I am running into this issue:

`ValueError                                Traceback (most recent call last)
/tmp/ipykernel_27/2150472599.py in <module>
----> 1 nlp = spacy.load("my_custom_model")
      2 doc = nlp("Tim Cook is the CEO of Apple")
      3 print([(ent.text, ent.label_) for ent in doc.ents])

/opt/conda/lib/python3.7/site-packages/spacy/__init__.py in load(name, vocab, disable, enable, exclude, config)
     58         enable=enable,
     59         exclude=exclude,
---> 60         config=config,
     61     )
     62 

/opt/conda/lib/python3.7/site-packages/spacy/util.py in load_model(name, vocab, disable, enable, exclude, config)
    432             return load_model_from_package(name, **kwargs)  # type: ignore[arg-type]
    433         if Path(name).exists():  # path to model data directory
--> 434             return load_model_from_path(Path(name), **kwargs)  # type: ignore[arg-type]
    435     elif hasattr(name, "exists"):  # Path or Path-like to model data
    436         return load_model_from_path(name, **kwargs)  # type: ignore[arg-type]

/opt/conda/lib/python3.7/site-packages/spacy/util.py in load_model_from_path(model_path, meta, vocab, disable, enable, exclude, config)
    512         meta=meta,
    513     )
--> 514     return nlp.from_disk(model_path, exclude=exclude, overrides=overrides)
    515 
    516 

/opt/conda/lib/python3.7/site-packages/spacy/language.py in from_disk(self, path, exclude, overrides)
   2123             # Convert to list here in case exclude is (default) tuple
   2124             exclude = list(exclude) + ["vocab"]
-> 2125         util.from_disk(path, deserializers, exclude)  # type: ignore[arg-type]
   2126         self._path = path  # type: ignore[assignment]
   2127         self._link_components()

/opt/conda/lib/python3.7/site-packages/spacy/util.py in from_disk(path, readers, exclude)
   1350         # Split to support file names like meta.json
   1351         if key.split(".")[0] not in exclude:
-> 1352             reader(path / key)
   1353     return path
   1354 

/opt/conda/lib/python3.7/site-packages/spacy/language.py in <lambda>(p, proc)
   2118                 continue
   2119             deserializers[name] = lambda p, proc=proc: proc.from_disk(  # type: ignore[misc]
-> 2120                 p, exclude=["vocab"]
   2121             )
   2122         if not (path / "vocab").exists() and "vocab" not in exclude:  # type: ignore[operator]

/opt/conda/lib/python3.7/site-packages/spacy/pipeline/transition_parser.pyx in spacy.pipeline.transition_parser.Parser.from_disk()

/opt/conda/lib/python3.7/site-packages/thinc/model.py in from_bytes(self, bytes_data)
    617         msg = srsly.msgpack_loads(bytes_data)
    618         msg = convert_recursive(is_xp_array, self.ops.asarray, msg)
--> 619         return self.from_dict(msg)
    620 
    621     def from_disk(self, path: Union[Path, str]) -> "Model":

/opt/conda/lib/python3.7/site-packages/thinc/model.py in from_dict(self, msg)
    640             for dim, value in info["dims"].items():
    641                 if value is not None:
--> 642                     node.set_dim(dim, value)
    643             for ref, ref_index in info["refs"].items():
    644                 if ref_index is None:

/opt/conda/lib/python3.7/site-packages/thinc/model.py in set_dim(self, name, value, force)
    190         if invalid_change:
    191             err = f"Attempt to change dimension '{name}' for model '{self.name}' from {old_value} to {value}"
--> 192             raise ValueError(err)
    193         self._dims[name] = value
    194 `

[danieldk]

I am sorry that you are running into this issue. nV is one of the dimensions used by the HashEmbed layer. It is the number of vectors in the hash embedding table. The error message hints that the value for nV derived from the model configuration is not the same as the serialized model. You could check your config.cfg inside the model directory. One of the row sizes for the tok2vec embedder is probably 2000, whereas the model was trained with 5000:

[components.senter.model.tok2vec.embed]
@architectures = "spacy.MultiHashEmbed.v2"
# One or more of these is probably set to 2000:
rows = [1000,500,500,500,50]

How this could have happened is hard to tell. Could you give us a minimal reproducer that results in this error?

[vrunm]

I took this as my training data

TRAIN_DATA = [    ("Amazon", {"entities": [(0, 6, "AMZN")]}),
    ("Apple", {"entities": [(0, 5, "AAPL")]}),
    ("Google", {"entities": [(0, 6, "GOOGL")]}),
    ("Intel", {"entities": [(0, 5, "INTC")]}),
    ("Micron", {"entities": [(0, 5, "MU")]}),
    ("AMD", {"entities": [(0, 3, "AMD")]}),
    ("Amazon", {"entities": [(0, 6, "AMZN")]}),
    ("Cisco", {"entities": [(0, 4, "CSCO")]}),
    ("Nvidia", {"entities": [(0, 6, "NVDA")]}),
    ("ASML", {"entities": [(0, 4, "ASML")]}),
    ("Microsoft", {"entities": [(0, 9, "MSFT")]}),
]

Loaded the spacy model

import spacy

nlp = spacy.load("en_core_web_sm")
ner = nlp.create_pipe("ner")

Created a training loop

from spacy.util import minibatch, compounding
from spacy.training import Example
import random
n_iter = 10
for i in range(n_iter):
    random.shuffle(TRAIN_DATA)
    losses = {}
    batches = minibatch(TRAIN_DATA, size=compounding(4.0, 32.0, 1.001))
    for batch in batches:
        texts, annotations = zip(*batch)
        examples = [Example.from_dict(nlp.make_doc(text), annotation) for text, annotation in zip(texts, annotations)]
        nlp.update(examples, losses=losses)
    print("Losses:", losses)

Saved the model

nlp.to_disk("my_custom_model")

Loaded the model for predictions

nlp = spacy.load("my_custom_model")
doc = nlp("Tim is the CEO of Apple")
print([(ent.text, ent.label_) for ent in doc.ents])

After running this I got the error as above.

[danieldk]

Thanks. The issue is that there already is a ner pipe in en_core_web_sm. The language class should reject adding a pipe with the same name, but doesn't, I'll investigate this further. You could solve your issue by giving your NER pipe a different name:

nlp = spacy.load("en_core_web_sm")
ner = nlp.create_pipe("ner", name="my_ner")

You could also remove the existing ner pipe:

nlp = spacy.load("en_core_web_sm")
nlp.remove_pipe("ner")
ner = nlp.create_pipe("ner")

[danieldk]

Getting back to this part:

The language class should reject adding a pipe with the same name, but doesn't, I'll investigate this further.

I read over this in your code fragment, but you should use add_pipe, create_pipe is now for internal use. Using ner = nlp.add_pipe("ner") actually returns the expected error:

ValueError: [E007] 'ner' already exists in pipeline. Existing names: ['tok2vec', 'tagger', 'parser', 'senter', 'attribute_ruler', 'lemmatizer', 'ner']

[vrunm]

@danieldk Sure I will use that.

[vrunm]

@danieldk I tried this and it solves the issue

[vrunm]

Using this

nlp = spacy.load("en_core_web_sm")
ner_pipe = nlp.add_pipe("ner", name="my_ner")

I ran into this error:

KeyError                                  Traceback (most recent call last)
/tmp/ipykernel_27/1755169283.py in <module>
     10         texts, annotations = zip(*batch)
     11         examples = [Example.from_dict(nlp.make_doc(text), annotation) for text, annotation in zip(texts, annotations)]
---> 12         nlp.update(examples, losses=losses)
     13     print("Losses:", losses)

/opt/conda/lib/python3.7/site-packages/spacy/language.py in update(self, examples, _, drop, sgd, losses, component_cfg, exclude, annotates)
   1153             # ignore statements are used here because mypy ignores hasattr
   1154             if name not in exclude and hasattr(proc, "update"):
-> 1155                 proc.update(examples, sgd=None, losses=losses, **component_cfg[name])  # type: ignore
   1156             if sgd not in (None, False):
   1157                 if (

/opt/conda/lib/python3.7/site-packages/spacy/pipeline/transition_parser.pyx in spacy.pipeline.transition_parser.Parser.update()

/opt/conda/lib/python3.7/site-packages/thinc/model.py in begin_update(self, X)
    307         and returns the gradient with respect to the input.
    308         """
--> 309         return self._func(self, X, is_train=True)
    310 
    311     def predict(self, X: InT) -> OutT:

/opt/conda/lib/python3.7/site-packages/spacy/ml/tb_framework.py in forward(model, X, is_train)
     36         unseen_classes=model.attrs["unseen_classes"],
     37         train=is_train,
---> 38         has_upper=model.attrs["has_upper"],
     39     )
     40 

/opt/conda/lib/python3.7/site-packages/spacy/ml/parser_model.pyx in spacy.ml.parser_model.ParserStepModel.__init__()

/opt/conda/lib/python3.7/site-packages/thinc/model.py in __call__(self, X, is_train)
    289         """Call the model's `forward` function, returning the output and a
    290         callback to compute the gradients via backpropagation."""
--> 291         return self._func(self, X, is_train=is_train)
    292 
    293     def initialize(self, X: Optional[InT] = None, Y: Optional[OutT] = None) -> "Model":

/opt/conda/lib/python3.7/site-packages/thinc/layers/chain.py in forward(model, X, is_train)
     53     callbacks = []
     54     for layer in model.layers:
---> 55         Y, inc_layer_grad = layer(X, is_train=is_train)
     56         callbacks.append(inc_layer_grad)
     57         X = Y

/opt/conda/lib/python3.7/site-packages/thinc/model.py in __call__(self, X, is_train)
    289         """Call the model's `forward` function, returning the output and a
    290         callback to compute the gradients via backpropagation."""
--> 291         return self._func(self, X, is_train=is_train)
    292 
    293     def initialize(self, X: Optional[InT] = None, Y: Optional[OutT] = None) -> "Model":

/opt/conda/lib/python3.7/site-packages/thinc/layers/chain.py in forward(model, X, is_train)
     53     callbacks = []
     54     for layer in model.layers:
---> 55         Y, inc_layer_grad = layer(X, is_train=is_train)
     56         callbacks.append(inc_layer_grad)
     57         X = Y

/opt/conda/lib/python3.7/site-packages/thinc/model.py in __call__(self, X, is_train)
    289         """Call the model's `forward` function, returning the output and a
    290         callback to compute the gradients via backpropagation."""
--> 291         return self._func(self, X, is_train=is_train)
    292 
    293     def initialize(self, X: Optional[InT] = None, Y: Optional[OutT] = None) -> "Model":

/opt/conda/lib/python3.7/site-packages/thinc/layers/chain.py in forward(model, X, is_train)
     53     callbacks = []
     54     for layer in model.layers:
---> 55         Y, inc_layer_grad = layer(X, is_train=is_train)
     56         callbacks.append(inc_layer_grad)
     57         X = Y

/opt/conda/lib/python3.7/site-packages/thinc/model.py in __call__(self, X, is_train)
    289         """Call the model's `forward` function, returning the output and a
    290         callback to compute the gradients via backpropagation."""
--> 291         return self._func(self, X, is_train=is_train)
    292 
    293     def initialize(self, X: Optional[InT] = None, Y: Optional[OutT] = None) -> "Model":

/opt/conda/lib/python3.7/site-packages/thinc/layers/with_array.py in forward(model, Xseq, is_train)
     30 ) -> Tuple[SeqT, Callable]:
     31     if isinstance(Xseq, Ragged):
---> 32         return cast(Tuple[SeqT, Callable], _ragged_forward(model, Xseq, is_train))
     33     elif isinstance(Xseq, Padded):
     34         return cast(Tuple[SeqT, Callable], _padded_forward(model, Xseq, is_train))

/opt/conda/lib/python3.7/site-packages/thinc/layers/with_array.py in _ragged_forward(model, Xr, is_train)
     85 ) -> Tuple[Ragged, Callable]:
     86     layer: Model[ArrayXd, ArrayXd] = model.layers[0]
---> 87     Y, get_dX = layer(Xr.dataXd, is_train)
     88 
     89     def backprop(dYr: Ragged) -> Ragged:

/opt/conda/lib/python3.7/site-packages/thinc/model.py in __call__(self, X, is_train)
    289         """Call the model's `forward` function, returning the output and a
    290         callback to compute the gradients via backpropagation."""
--> 291         return self._func(self, X, is_train=is_train)
    292 
    293     def initialize(self, X: Optional[InT] = None, Y: Optional[OutT] = None) -> "Model":

/opt/conda/lib/python3.7/site-packages/thinc/layers/concatenate.py in forward(model, X, is_train)
     42 
     43 def forward(model: Model[InT, OutT], X: InT, is_train: bool) -> Tuple[OutT, Callable]:
---> 44     Ys, callbacks = zip(*[layer(X, is_train=is_train) for layer in model.layers])
     45     if isinstance(Ys[0], list):
     46         data_l, backprop = _list_forward(model, X, Ys, callbacks, is_train)

/opt/conda/lib/python3.7/site-packages/thinc/layers/concatenate.py in <listcomp>(.0)
     42 
     43 def forward(model: Model[InT, OutT], X: InT, is_train: bool) -> Tuple[OutT, Callable]:
---> 44     Ys, callbacks = zip(*[layer(X, is_train=is_train) for layer in model.layers])
     45     if isinstance(Ys[0], list):
     46         data_l, backprop = _list_forward(model, X, Ys, callbacks, is_train)

/opt/conda/lib/python3.7/site-packages/thinc/model.py in __call__(self, X, is_train)
    289         """Call the model's `forward` function, returning the output and a
    290         callback to compute the gradients via backpropagation."""
--> 291         return self._func(self, X, is_train=is_train)
    292 
    293     def initialize(self, X: Optional[InT] = None, Y: Optional[OutT] = None) -> "Model":

/opt/conda/lib/python3.7/site-packages/thinc/layers/chain.py in forward(model, X, is_train)
     53     callbacks = []
     54     for layer in model.layers:
---> 55         Y, inc_layer_grad = layer(X, is_train=is_train)
     56         callbacks.append(inc_layer_grad)
     57         X = Y

/opt/conda/lib/python3.7/site-packages/thinc/model.py in __call__(self, X, is_train)
    289         """Call the model's `forward` function, returning the output and a
    290         callback to compute the gradients via backpropagation."""
--> 291         return self._func(self, X, is_train=is_train)
    292 
    293     def initialize(self, X: Optional[InT] = None, Y: Optional[OutT] = None) -> "Model":

/opt/conda/lib/python3.7/site-packages/thinc/layers/hashembed.py in forward(model, ids, is_train)
     61     model: Model[Ints1d, OutT], ids: Ints1d, is_train: bool
     62 ) -> Tuple[OutT, Callable]:
---> 63     vectors = cast(Floats2d, model.get_param("E"))
     64     nV = vectors.shape[0]
     65     nO = vectors.shape[1]

/opt/conda/lib/python3.7/site-packages/thinc/model.py in get_param(self, name)
    215         if not self._params.has_param(self.id, name):
    216             raise KeyError(
--> 217                 f"Parameter '{name}' for model '{self.name}' has not been allocated yet."
    218             )
    219         return self._params.get_param(self.id, name)

KeyError: "Parameter 'E' for model 'hashembed' has not been allocated yet."

@danieldk Is there any way around this?

[danieldk]

You need to initialize a fresh pipeline component with a representative sample. E.g.:

ner = nlp.add_pipe("ner", name="my_ner")
examples = [Example.from_dict(nlp.make_doc(text), annotation) for text, annotation in TRAIN_DATA]
ner.initialize(lambda: examples)

The reason is that the component's model can only be initialized when e.g. the number of labels is known. The initialize method gathers the labels (among other things) from the provided sample.

[vrunm]

Sure I used a fresh pipeline component with a training sample and it works.