Cannot deserialize model: mismatched structure

[zebrassimo]

============================== Info about spaCy ==============================

spaCy version    3.1.3                         
Location         /opt/homebrew/lib/python3.9/site-packages/spacy
Platform         macOS-12.0.1-arm64-arm-64bit  
Python version   3.9.7                         
Pipelines        en_core_web_md (3.1.0) 
================= Installed pipeline packages (spaCy v3.1.3) =================
ℹ spaCy installation:
/opt/homebrew/lib/python3.9/site-packages/spacy

NAME             SPACY            VERSION                            
en_core_web_md   >=3.1.0,<3.2.0   3.1.0   ✔

Following @DuyguA 's book mastering spacey (with some 3.0 tweaks):

import random
import spacy
from spacy.training import Example
nlp=spacy.load("en_core_web_md")
trainset = [
    ("navigate home",             [(9,13,"GPE")]),
    ("navigate to office",        [(12,18,"GPE")]),
    ("navigate",                  []),
    ("navigate to Oxford Street", [(12,25,"GPE")])
]
epochs = 20
other_pipes = [pipe for pipe in nlp.pipe_names if pipe != "ner"]
with nlp.disable_pipes(*other_pipes):
    optimizer = nlp.create_optimizer()
    for i in range(epochs):
        random.shuffle(trainset)
        for text, annotation in trainset:
            doc = nlp(text)
            gold_dict = {"entities": annotation}
            example = Example.from_dict(doc, gold_dict)
            losses = nlp.update([example], sgd=optimizer)
ner = nlp.get_pipe("ner")
ner.to_disk("./navi_ner")

Having saved the EntityRecognizer with it's own (from TrainablePipe inherited) to_disk, I assume it should be possible to load it as well:

import spacy
from spacy.lang.en import English
nlp = spacy.load("en_core_web_md", disable=["ner"])
ner = nlp.create_pipe("ner")
ner.from_disk("./navi_ner")

however fails:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/var/folders/pd/_n__n8h507n14my7cvg904n40000gp/T/ipykernel_7898/2280068043.py in <module>
      3 nlp = spacy.load("en_core_web_md", disable=["ner"])
      4 ner = nlp.create_pipe("ner")
----> 5 ner.from_disk("./navi_ner")

/opt/homebrew/lib/python3.9/site-packages/spacy/pipeline/transition_parser.pyx in spacy.pipeline.transition_parser.Parser.from_disk()

/opt/homebrew/lib/python3.9/site-packages/thinc/model.py in from_bytes(self, bytes_data)
    591         msg = srsly.msgpack_loads(bytes_data)
    592         msg = convert_recursive(is_xp_array, self.ops.asarray, msg)
--> 593         return self.from_dict(msg)
    594 
    595     def from_disk(self, path: Union[Path, str]) -> "Model":

/opt/homebrew/lib/python3.9/site-packages/thinc/model.py in from_dict(self, msg)
    608         nodes = list(self.walk())
    609         if len(msg["nodes"]) != len(nodes):
--> 610             raise ValueError("Cannot deserialize model: mismatched structure")
    611         for i, node in enumerate(nodes):
    612             info = msg["nodes"][i]

ValueError: Cannot deserialize model: mismatched structure

I'm rather sure the issue lies with thinc rather than with TrainablePipe, but I'm too much of a newbie.

Any suggestions are highly appreceated.

[taghouti-ghofrane]

any answer

[honnibal]

Hi sorry for leaving this unreplied!

I think this is really a spaCy issue rather than a Thinc issue. I think it's a matter of nlp.create_pipe() creating the component with different config from what the en_core_web_md saved model has. But you should take this up on the spaCy discussions (as I do think it's a usage question rather than a bug) to be sure.

[adrianeboyd]

Yes, this is due to mismatched configs. To load in the ner component from another pipeline, save the full pipeline using nlp.to_disk and use the source option to load config+model together:

nlp = spacy.load("en_core_web_md", exclude=["ner"])
nlp.add_pipe("ner", source=spacy.load("my_custom_pipeline"))