univa/training/configuration_lvlm.py

from transformers import TrainingArguments
from dataclasses import dataclass


@dataclass
class TrainingConfig(TrainingArguments): ...


@dataclass
class DatasetConfig:
    data_txt: str


@dataclass
class ModelConfig:
    pretrained_model_path_or_name: str
    image_processor_path: str

    train_vision_tower: bool = False
    train_mm_projector: bool = True
    train_llm: bool = True
    train_lm_head: bool = True


@dataclass
class UnivaTrainingConfig:
    training_config: TrainingConfig
    dataset_config: DatasetConfig
    model_config: ModelConfig

    @classmethod
    def from_dict(cls, training_config: dict, dataset_config: dict, model_config: dict):
        return cls(
            training_config=TrainingConfig(**training_config),
            dataset_config=DatasetConfig(**dataset_config),
            model_config=ModelConfig(**model_config),
        )