InferenceApi for TokenClassification does not work

[KoichiYasuoka]

I've tried InferenceApi for the pipeline of token-classification but failed:

>>> from huggingface_hub.inference_api import InferenceApi
>>> api=InferenceApi("KoichiYasuoka/bert-base-japanese-upos")
>>> r=api(inputs="國境の長いトンネルを拔けると雪國であつた。")
>>> import json
>>> print(json.dumps(r,indent=2,ensure_ascii=False))
{
  "error": "maximum recursion depth exceeded while calling a Python object"
}

I remember that it worked last month and the result still remains in the cache:

>>> from huggingface_hub.inference_api import InferenceApi
>>> api=InferenceApi("KoichiYasuoka/bert-base-japanese-upos")
>>> r=api(inputs="難儀な難儀は難儀する")
>>> import json
>>> print(json.dumps(r,indent=2,ensure_ascii=False))
[
  {
    "entity_group": "ADJ",
    "score": 0.6918697953224182,
    "word": "難 儀",
    "start": null,
    "end": null
  },
  {
    "entity_group": "AUX",
    "score": 0.995212972164154,
    "word": "な",
    "start": null,
    "end": null
  },
  {
    "entity_group": "NOUN",
    "score": 0.9972944259643555,
    "word": "難 儀",
    "start": null,
    "end": null
  },
  {
    "entity_group": "ADP",
    "score": 0.9995647072792053,
    "word": "は",
    "start": null,
    "end": null
  },
  {
    "entity_group": "VERB",
    "score": 0.6875597238540649,
    "word": "難 儀",
    "start": null,
    "end": null
  },
  {
    "entity_group": "AUX",
    "score": 0.9953439235687256,
    "word": "す る",
    "start": null,
    "end": null
  }
]

Something changed on InferenceApi or TokenClassification?

[julien-c]

cc'ing @Narsil for visibility

[KoichiYasuoka]

Thank you @julien-c for your cc with @Narsil, and I've just found a very big PR huggingface/transformers#13308 and a smaller PR huggingface/transformers#13493

[Narsil]

Ok, it's fixed now.

It seemed linked to recursive import, which is odd, we're looking into it.

[KoichiYasuoka]

Thank you @Narsil and I've just confirmed that it works well:

>>> from huggingface_hub.inference_api import InferenceApi
>>> api=InferenceApi("KoichiYasuoka/bert-base-japanese-upos")
>>> r=api(inputs="國境の長いトンネルを拔けると雪國であつた。")
>>> import json
>>> print(json.dumps(r,indent=2,ensure_ascii=False))
[
  {
    "entity_group": "NOUN",
    "score": 0.9991488456726074,
    "word": "國境",
    "start": 0,
    "end": 2
  },
  {
    "entity_group": "ADP",
    "score": 0.99976646900177,
    "word": "の",
    "start": 2,
    "end": 3
  },
  {
    "entity_group": "ADJ",
    "score": 0.9954113960266113,
    "word": "長い",
    "start": 3,
    "end": 5
  },
  {
    "entity_group": "NOUN",
    "score": 0.9995394349098206,
    "word": "トンネル",
    "start": 5,
    "end": 9
  },
  {
    "entity_group": "ADP",
    "score": 0.9997807145118713,
    "word": "を",
    "start": 9,
    "end": 10
  },
  {
    "entity_group": "VERB",
    "score": 0.9993178844451904,
    "word": "拔ける",
    "start": 10,
    "end": 13
  },
  {
    "entity_group": "CCONJ",
    "score": 0.9845631718635559,
    "word": "と",
    "start": 13,
    "end": 14
  },
  {
    "entity_group": "NOUN",
    "score": 0.9390370845794678,
    "word": "雪國",
    "start": 14,
    "end": 16
  },
  {
    "entity_group": "AUX",
    "score": 0.9977393746376038,
    "word": "で",
    "start": 16,
    "end": 17
  },
  {
    "entity_group": "AUX",
    "score": 0.9874789118766785,
    "word": "あつた",
    "start": 17,
    "end": 20
  },
  {
    "entity_group": "PUNCT",
    "score": 0.9998145699501038,
    "word": "。",
    "start": 20,
    "end": 21
  }
]

Thank you again and I close this issue.

[Narsil]

Found the culprit which was linked to bad merge between master and the API fork (Turns out GPT2ForTokenClassification was added relatively recently).