[Question]: SequenceTagger.load fails with TypeError: not a string when loading a fine-tuned model

[pganesh]

Question

For fine-tuning, I used the google/flan-t5-large as the transformer like this:

4. initialize fine-tuneable transformer embeddings WITH document context

embeddings = TransformerWordEmbeddings(model='google/flan-t5-large',
layers="-1",
subtoken_pooling="first",
fine_tune=True,
use_context=True,
)
The training was succesful and the model was saved correctly as a final-model.pt
However, when I try to use it as follows:

load the trained model

model = SequenceTagger.load('/content/resources/taggers/sota-ner-flert/final-model.pt')

I get an error saying that "TypeError: not a string"
I suspect this is due to using the flan-t5-large - everything works great if I used xlm-roberta-large for example
I need some help on how to use the saved/fine-tuned model if the original embedding was based on flan-t5
thanks!

[pganesh]

HEre's the error trace:

TypeError Traceback (most recent call last)
in <cell line: 2>()
1 # Model name points to the best-model.pt or final-model.pt path
----> 2 model = SequenceTagger.load(final_model_name)
3
4 # Use corpus.dev or corpus.test
5 result = model.evaluate(corpus.dev, gold_label_type="ner", mini_batch_size=8)

19 frames
/usr/local/lib/python3.10/dist-packages/flair/models/sequence_tagger_model.py in load(cls, model_path)
923 from typing import cast
924
--> 925 return cast("SequenceTagger", super().load(model_path=model_path))

/usr/local/lib/python3.10/dist-packages/flair/nn/model.py in load(cls, model_path)
562 from typing import cast
563
--> 564 return cast("Classifier", super().load(model_path=model_path))
565
566

/usr/local/lib/python3.10/dist-packages/flair/nn/model.py in load(cls, model_path)
195 state.pop("cls")
196
--> 197 model = cls._init_model_with_state_dict(state)
198
199 if "model_card" in state:

/usr/local/lib/python3.10/dist-packages/flair/models/sequence_tagger_model.py in _init_model_with_state_dict(cls, state, **kwargs)
622 del state["state_dict"]["transitions"]
623
--> 624 return super()._init_model_with_state_dict(
625 state,
626 embeddings=state.get("embeddings"),

/usr/local/lib/python3.10/dist-packages/flair/nn/model.py in _init_model_with_state_dict(cls, state, **kwargs)
98 embeddings = kwargs.pop("embeddings")
99 if isinstance(embeddings, dict):
--> 100 embeddings = load_embeddings(embeddings)
101 kwargs["embeddings"] = embeddings
102

/usr/local/lib/python3.10/dist-packages/flair/embeddings/base.py in load_embeddings(params)
229 cls_name = params.pop("cls")
230 cls = EMBEDDING_CLASSES[cls_name]
--> 231 return cls.load_embedding(params)

/usr/local/lib/python3.10/dist-packages/flair/embeddings/base.py in load_embedding(cls, params)
95 state_dict = params.pop("state_dict", None)
96
---> 97 embedding = cls.from_params(params)
98 if state_dict is not None:
99 embedding.load_state_dict(state_dict)

/usr/local/lib/python3.10/dist-packages/flair/embeddings/transformer.py in from_params(cls, params)
1326 config_class = CONFIG_MAPPING[model_type]
1327 config = config_class.from_dict(config_state_dict)
-> 1328 return cls.create_from_state(saved_config=config, **params)
1329
1330 def to_params(self):

/usr/local/lib/python3.10/dist-packages/flair/embeddings/token.py in create_from_state(cls, **state)
58 # this parameter is fixed
59 del state["is_token_embedding"]
---> 60 return cls(**state)
61
62

/usr/local/lib/python3.10/dist-packages/flair/embeddings/token.py in init(self, model, is_document_embedding, allow_long_sentences, **kwargs)
45 **kwargs: Arguments propagated to :meth:flair.embeddings.transformer.TransformerEmbeddings.__init__
46 """
---> 47 TransformerEmbeddings.init(
48 self,
49 model=model,

/usr/local/lib/python3.10/dist-packages/flair/embeddings/transformer.py in init(self, model, fine_tune, layers, layer_mean, subtoken_pooling, cls_pooling, is_token_embedding, is_document_embedding, allow_long_sentences, use_context, respect_document_boundaries, context_dropout, saved_config, tokenizer_data, feature_extractor_data, name, force_max_length, needs_manual_ocr, use_context_separator, transformers_tokenizer_kwargs, transformers_config_kwargs, transformers_model_kwargs, peft_config, peft_gradient_checkpointing_kwargs, **kwargs)
1062 else:
1063 # load tokenizer from inmemory zip-file
-> 1064 self.tokenizer = self._tokenizer_from_bytes(tokenizer_data)
1065 if feature_extractor_data is not None:
1066 self.feature_extractor = self._feature_extractor_from_bytes(feature_extractor_data)

/usr/local/lib/python3.10/dist-packages/flair/embeddings/transformer.py in _tokenizer_from_bytes(cls, zip_data)
428 with tempfile.TemporaryDirectory() as temp_dir:
429 zip_obj.extractall(temp_dir)
--> 430 return AutoTokenizer.from_pretrained(temp_dir, add_prefix_space=True)
431
432 @classmethod

/usr/local/lib/python3.10/dist-packages/transformers/models/auto/tokenization_auto.py in from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs)
918 f"Tokenizer class {tokenizer_class_candidate} does not exist or is not currently imported."
919 )
--> 920 return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
921
922 # Otherwise we have to be creative.

/usr/local/lib/python3.10/dist-packages/transformers/tokenization_utils_base.py in from_pretrained(cls, pretrained_model_name_or_path, cache_dir, force_download, local_files_only, token, revision, trust_remote_code, *init_inputs, **kwargs)
2211 logger.info(f"loading file {file_path} from cache at {resolved_vocab_files[file_id]}")
2212
-> 2213 return cls._from_pretrained(
2214 resolved_vocab_files,
2215 pretrained_model_name_or_path,

/usr/local/lib/python3.10/dist-packages/transformers/tokenization_utils_base.py in _from_pretrained(cls, resolved_vocab_files, pretrained_model_name_or_path, init_configuration, token, cache_dir, local_files_only, _commit_hash, _is_local, trust_remote_code, *init_inputs, **kwargs)
2445 # Instantiate the tokenizer.
2446 try:
-> 2447 tokenizer = cls(*init_inputs, **init_kwargs)
2448 except import_protobuf_decode_error():
2449 logger.info(

/usr/local/lib/python3.10/dist-packages/transformers/models/t5/tokenization_t5_fast.py in init(self, vocab_file, tokenizer_file, eos_token, unk_token, pad_token, extra_ids, additional_special_tokens, add_prefix_space, **kwargs)
117 kwargs["from_slow"] = True
118
--> 119 super().init(
120 vocab_file,
121 tokenizer_file=tokenizer_file,

/usr/local/lib/python3.10/dist-packages/transformers/tokenization_utils_fast.py in init(self, *args, **kwargs)
130 elif self.slow_tokenizer_class is not None and slow_tokenizer is not False:
131 # We need to create and convert a slow tokenizer to build the backend
--> 132 slow_tokenizer = self.slow_tokenizer_class(*args, **kwargs)
133 fast_tokenizer = convert_slow_tokenizer(slow_tokenizer)
134 elif not slow_tokenizer:

/usr/local/lib/python3.10/dist-packages/transformers/models/t5/tokenization_t5.py in init(self, vocab_file, eos_token, unk_token, pad_token, extra_ids, additional_special_tokens, sp_model_kwargs, legacy, add_prefix_space, **kwargs)
148
149 self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
--> 150 self.sp_model.Load(vocab_file)
151
152 if additional_special_tokens is not None:

/usr/local/lib/python3.10/dist-packages/sentencepiece/init.py in Load(self, model_file, model_proto)
959 if model_proto:
960 return self.LoadFromSerializedProto(model_proto)
--> 961 return self.LoadFromFile(model_file)
962
963

/usr/local/lib/python3.10/dist-packages/sentencepiece/init.py in LoadFromFile(self, arg)
314
315 def LoadFromFile(self, arg):
--> 316 return _sentencepiece.SentencePieceProcessor_LoadFromFile(self, arg)
317
318 def _EncodeAsIds(self, text, enable_sampling, nbest_size, alpha, add_bos, add_eos, reverse, emit_unk_piece):

TypeError: not a string

[helpmefindaname]

Hi @pganesh
I am sorry, but I don't want to read so much text with bad formatting. Besides that, I think there is some default information missing, as this isn't a question but a bug report.

I am closing this for now, please create an issue using the bug-report template, fill out all fields there and verify afterwards that the markdown is readable.

[pganesh]

OK understood - will open a bug report

…

On Fri, Dec 6, 2024 at 7:58 AM Benedikt Fuchs ***@***.***> wrote: Hi @pganesh <https://github.com/pganesh> I am sorry, but I don't want to read so much text with bad formatting. Besides that, I think there is some default information missing, as this isn't a question but a bug report. I am closing this for now, please create an issue using the bug-report template, fill out all fields there and verify afterwards that the markdown is readable. — Reply to this email directly, view it on GitHub <#3574 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABOOQI3QA4V62K74JZUNDZT2EHCQPAVCNFSM6AAAAABS4VM6PWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKMRTGU3DKOBWGE> . You are receiving this because you were mentioned.Message ID: ***@***.***>