使用 Optimum Habana 预训练 Transformers

从 Transformers 预训练模型，如 BERT，与微调一样简单。模型应该从配置中使用 .from_config 实例化，而不是从预训练的检查点中使用 .from_pretrained 实例化。以下是如何使用 GPT2 的示例

from transformers import AutoConfig, AutoModelForXXX

config = AutoConfig.from_pretrained("gpt2")
model = AutoModelForXXX.from_config(config)

其中 XXX 是要执行的任务，例如 ImageClassification。

以下是一个 BERT 用于掩码语言建模的预训练工作示例

from datasets import load_dataset
from optimum.habana import GaudiTrainer, GaudiTrainingArguments
from transformers import AutoConfig, AutoModelForMaskedLM, AutoTokenizer, DataCollatorForLanguageModeling

# Load the training set (this one has already been preprocessed)
training_set = load_dataset("philschmid/processed_bert_dataset", split="train")
# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained("philschmid/bert-base-uncased-2022-habana")

# Instantiate an untrained model
config = AutoConfig.from_pretrained("bert-base-uncased")
model = AutoModelForMaskedLM.from_config(config)

model.resize_token_embeddings(len(tokenizer))

# The data collator will take care of randomly masking the tokens
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer)

training_args = GaudiTrainingArguments(
    output_dir="/tmp/bert-base-uncased-mlm",
    num_train_epochs=1,
    per_device_train_batch_size=8,
    use_habana=True,
    use_lazy_mode=True,
    gaudi_config_name="Habana/bert-base-uncased",
)

# Initialize our Trainer
trainer = GaudiTrainer(
    model=model,
    args=training_args,
    train_dataset=training_set,
    tokenizer=tokenizer,
    data_collator=data_collator,
)

trainer.train()

您可以在这篇博文中看到另一个预训练示例。

< > 在 GitHub 上更新