Fail to train openai-community / gpt2 model for custom NER on SpaCy framework

[Project-NSE]

I want to create a custom NER tag using GPT2. I want to use this model. I am familiar with SpaCy custom training framework. I formatted the config.cfg file as per the requirement. The config.cfg file is as follows

[paths]
train = null
dev = null
vectors = null
init_tok2vec = null

[system]
gpu_allocator = "pytorch"
seed = 0

[nlp]
lang = "en"
pipeline = ["transformer","ner"]
batch_size = 128
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
tokenizer = {"@tokenizers":"spacy.Tokenizer.v1"}
vectors = {"@vectors":"spacy.Vectors.v1"}
tokenizer.pad_token = tokenizer.eos_token

[components]

[components.ner]
factory = "ner"
incorrect_spans_key = null
moves = null
scorer = {"@scorers":"spacy.ner_scorer.v1"}
update_with_oracle_cut_size = 100

[components.ner.model]
@architectures = "spacy.TransitionBasedParser.v2"
state_type = "ner"
extra_state_tokens = false
hidden_width = 64
maxout_pieces = 2
use_upper = false
nO = null

[components.ner.model.tok2vec]
@architectures = "spacy-transformers.TransformerListener.v1"
grad_factor = 1.0
pooling = {"@layers":"reduce_mean.v1"}
upstream = "*"

[components.transformer]
factory = "transformer"
max_batch_items = 4096
set_extra_annotations = {"@annotation_setters":"spacy-transformers.null_annotation_setter.v1"}

[components.transformer.model]
@architectures = "spacy-transformers.TransformerModel.v3"
name = "openai-community/gpt2"
mixed_precision = false

[components.transformer.model.get_spans]
@span_getters = "spacy-transformers.strided_spans.v1"
window = 128
stride = 96

[components.transformer.model.grad_scaler_config]

[components.transformer.model.tokenizer_config]
use_fast = true

[components.transformer.model.transformer_config]

[corpora]

[corpora.dev]
@readers = "spacy.Corpus.v1"
path = ${paths.dev}
max_length = 0
gold_preproc = false
limit = 0
augmenter = null

[corpora.train]
@readers = "spacy.Corpus.v1"
path = ${paths.train}
max_length = 0
gold_preproc = false
limit = 0
augmenter = null

[training]
accumulate_gradient = 3
dev_corpus = "corpora.dev"
train_corpus = "corpora.train"
seed = ${system.seed}
gpu_allocator = ${system.gpu_allocator}
dropout = 0.1
patience = 1600
max_epochs = 0
max_steps = 20000
eval_frequency = 200
frozen_components = []
annotating_components = []
before_to_disk = null
before_update = null

[training.batcher]
@batchers = "spacy.batch_by_padded.v1"
discard_oversize = true
size = 2000
buffer = 256
get_length = null

[training.logger]
@loggers = "spacy.ConsoleLogger.v1"
progress_bar = false

[training.optimizer]
@optimizers = "Adam.v1"
beta1 = 0.9
beta2 = 0.999
L2_is_weight_decay = true
L2 = 0.01
grad_clip = 1.0
use_averages = false
eps = 0.00000001

[training.optimizer.learn_rate]
@schedules = "warmup_linear.v1"
warmup_steps = 250
total_steps = 20000
initial_rate = 0.00005

[training.score_weights]
ents_f = 1.0
ents_p = 0.0
ents_r = 0.0
ents_per_type = null

[pretraining]

[initialize]
vectors = ${paths.vectors}
init_tok2vec = ${paths.init_tok2vec}
vocab_data = null
lookups = null
before_init = null
after_init = null

[initialize.components]

[initialize.tokenizer]

I received an error "ValueError: Asking to pad but the tokenizer does not have a padding token. Please select a token to use as pad_token (tokenizer.pad_token = tokenizer.eos_token e.g.) or add a new pad token via tokenizer.add_special_tokens({'pad_token': '[PAD]'})."

I posted this issue to the OpenAi forum and came to know from @younesbelkada that I need to call tokenizer.pad_token = tokenizer.eos_token before launching the training.

I modified the config.cfg file again. The modification portion is as follows

[nlp]
lang = "en"
pipeline = ["transformer","ner"]
batch_size = 128
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
tokenizer = {"@tokenizers":"spacy.Tokenizer.v1"}
vectors = {"@vectors":"spacy.Vectors.v1"}
tokenizer.pad_token = tokenizer.eos_token

Now I received the error "✘ Config validation error
nlp -> tokenizer.pad_token extra fields not permitted
{'lang': 'en', 'pipeline': ['transformer', 'ner'], 'batch_size': 128, 'disabled': [], 'before_creation': None, 'after_creation': None, 'after_pipeline_creation': None, 'tokenizer': {'@Tokenizers': 'spacy.Tokenizer.v1'}, 'vectors': {'@vectors': 'spacy.Vectors.v1'}, 'tokenizer.pad_token': 'tokenizer.eos_token'}"

Can you please let me know how can I fix this issue and train the GPT2 model.

Thank you in advance.

[svlandeg]

Hi!

There's a bit of a confusion here between spaCy's tokenizer and the HuggingFace transformer tokenizer.

spaCy's tokenizer is really just used to determine token boundaries in the text. It's defined in the [nlp] section in your config and can stay as is.

The error message that you're getting:

ValueError: Asking to pad but the tokenizer does not have a padding token. Please select a token to use as pad_token (tokenizer.pad_token = tokenizer.eos_token e.g.) or add a new pad token via tokenizer.add_special_tokens({'pad_token': '[PAD]'}).

on the other hand, is thrown by the Huggingface transformers library, and is about the tokenizer that you're loading by specifying

[components.transformer.model]
@architectures = "spacy-transformers.TransformerModel.v3"
name = "openai-community/gpt2"

The tool that bridges between spaCy and Huggingface here is called spacy-transformers, and it will load the pretrained HF transformer like so:

    tokenizer = tokenizer_cls.from_pretrained(str_path, **tok_config)
    config = config_cls.from_pretrained(str_path, **trf_config)
    transformer = model_cls.from_pretrained(str_path, config=config)

Note in particular the tok_config and trf_config variables - these can be set in your spaCy config and will be passed onwards to Huggingface for the tokenizer & transformer construction. You already have these:

[components.transformer.model.tokenizer_config]
use_fast = true

[components.transformer.model.transformer_config]

Now you can extend the tokenizer config to define the pad_token, for instance:

[components.transformer.model.tokenizer_config]
use_fast = true
pad_token = "<|endoftext|>"

Let us know if that works!