aurora

1
# Inspired by: https://github.com/huggingface/transformers/blob/v4.34.1/examples/pytorch/language-modeling/run_clm.py
2

3
import math
4
from typing import TYPE_CHECKING, Optional, List
5
from transformers import DataCollatorForLanguageModeling, Trainer
6

7
from llmtuner.data import get_dataset, preprocess_dataset, split_dataset
8
from llmtuner.extras.ploting import plot_loss
9
from llmtuner.model import load_model_and_tokenizer
10
from llmtuner.train.utils import create_modelcard_and_push
11

12
if TYPE_CHECKING:
13
    from transformers import Seq2SeqTrainingArguments, TrainerCallback
14
    from llmtuner.hparams import ModelArguments, DataArguments, FinetuningArguments
15

16

17
def run_pt(
18
    model_args: "ModelArguments",
19
    data_args: "DataArguments",
20
    training_args: "Seq2SeqTrainingArguments",
21
    finetuning_args: "FinetuningArguments",
22
    callbacks: Optional[List["TrainerCallback"]] = None
23
):
24
    dataset = get_dataset(model_args, data_args)
25
    model, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, training_args.do_train)
26
    dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, stage="pt")
27
    data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
28

29
    # Initialize our Trainer
30
    trainer = Trainer(
31
        model=model,
32
        args=training_args,
33
        tokenizer=tokenizer,
34
        data_collator=data_collator,
35
        callbacks=callbacks,
36
        **split_dataset(dataset, data_args, training_args)
37
    )
38

39
    # Training
40
    if training_args.do_train:
41
        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
42
        trainer.save_model()
43
        trainer.log_metrics("train", train_result.metrics)
44
        trainer.save_metrics("train", train_result.metrics)
45
        trainer.save_state()
46
        if trainer.is_world_process_zero() and finetuning_args.plot_loss:
47
            plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
48

49
    # Evaluation
50
    if training_args.do_eval:
51
        metrics = trainer.evaluate(metric_key_prefix="eval")
52
        try:
53
            perplexity = math.exp(metrics["eval_loss"])
54
        except OverflowError:
55
            perplexity = float("inf")
56

57
        metrics["perplexity"] = perplexity
58
        trainer.log_metrics("eval", metrics)
59
        trainer.save_metrics("eval", metrics)
60

61
    # Create model card
62
    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)
63