optimise training.iob_utils.offsets_to_biluo_tags method

[deepcode-debug]

How to reproduce the behaviour

run space.training.Example.from_dict

the method offsets_to_biluo_tags in training.iob_utils is not optimised, it is taking quite significant time while running Example.from_dict

Here is a profile snapshot

ncalls tottime percall cumtime percall filename:lineno(function)
14601 24.677 0.002 90.797 0.006 {built-in method from_dict}
112839144 19.502 0.000 33.350 0.000 doc.pyx:484(iter)
2 15.860 7.930 15.860 7.930 {built-in method _pickle.load}
112702047 13.855 0.000 13.855 0.000 token.pxd:21(cinit)
15276 9.896 0.001 22.981 0.002 doc.pyx:1327(to_dict)
18106173 7.784 0.000 11.133 0.000 vocab.pyx:142(get)
1 7.507 7.507 41.008 41.008 {built-in method _pickle.dump}
537703 7.312 0.000 7.312 0.000 decoder.py:343(raw_decode)
11180 6.963 0.001 24.228 0.002 iob_utils.py:63(offsets_to_biluo_tags)
31407 6.523 0.000 15.889 0.001 doc.pyx:180(init)
1 5.164 5.164 181.599 181.599 Ned_process_deep.py:1()
7968792 4.563 0.000 7.044 0.000 doc.pyx:694(genexpr)
20025 4.461 0.000 7.149 0.000 init.py:100(get_all)

Your Environment

• Operating System: Mac
• Python Version Used: 3.9
• spaCy Version Used: 3.3.1
• Environment Information: condo

[rmitsch]

Converting this to a discussion, as I don't think this is a concrete issue and a discussion might be a better format.

[deepcode-debug]

the problem is with the large loop in line 130:
an alternative could be to initilialize all with "O" tokens and then assign only the entity values.

[adrianeboyd]

Example.from_dict is mainly intended as a transition method to convert the format used in v2 demo examples to the v3 Doc-based training data format DocBin/.spacy. I don't have enough context from this thread, but my guess is that you're concerned about the speed because you're using Example.from_dict within a training loop?

I would recommend converting your data once, storing it in a DocBin, saving it as a .spacy file to use with spacy train (or you can also read from it with DocBin or Corpus in custom training scripts, but we strongly recommend using a config with spacy train).

An example of a conversion script for NER data is here:

• data: https://github.com/explosion/projects/blob/v3/pipelines/ner_demo/assets/train.json
• conversion script: https://github.com/explosion/projects/blob/v3/pipelines/ner_demo/scripts/convert.py

A PR that improves the speed would always be welcome!