BIDINLTK/dev/bert/BIDIBert_train.py

import sys

sys.path.insert(0,'../..')


from transformers import ElectraTokenizer

base_model_name = "chinese-electra-180g-small-discriminator"
model_name = "chinese-electra-180g-small-discriminator"
save_name = "BIDIBert"

tokenizer = ElectraTokenizer.from_pretrained(base_model_name)



from dev.bert.BIDIBert import BIDIBert,BIDIBertForMaskedLM
lm_model = BIDIBertForMaskedLM(len(tokenizer.get_vocab()),d_word_vec=128, n_layers=3, n_head=3, d_k=128, d_v=128,
                    d_model=128, d_inner=128,pad_idx=tokenizer.pad_token_id)
model_pretrain_path = "/data2/luojiehua/BIDINLTK/dev/bert/BIDIBert-bidi/checkpoint-4802000/pytorch_model.bin"
import torch
lm_model.load_state_dict(torch.load(model_pretrain_path))

# #freeze the backbone
# for param in lm_model.parameters():
#     param.requires_grad = False


from transformers import LineByLineTextDataset


# inputfile = r"G:\NLPDatasets\bidi_industry\bidi_industry_knowledge.txt"
inputfile = r"/data2/luojiehua/BIDINLTK/dev/data/bidi_industry_knowledge.txt"
# inputfile = r'/data2/luojiehua/BIDINLTK/dev/data/wikiouta.txt'

dataset = LineByLineTextDataset(
    tokenizer=tokenizer,
    file_path=inputfile,
    block_size=512,
)


print("dataset loaded")

from transformers import DataCollatorForLanguageModeling

data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer, mlm=True, mlm_probability=0.15
)

from transformers import Trainer, TrainingArguments

training_args = TrainingArguments(
    output_dir="./%s-bidi"%save_name,
    overwrite_output_dir=True,
    num_train_epochs=40,
    per_gpu_train_batch_size=3,
    save_steps=2000,
    learning_rate=5e-6,
    save_total_limit=2,
)

trainer = Trainer(
    model=lm_model,
    args=training_args,
    data_collator=data_collator,
    train_dataset=dataset,
)

trainer.train()