src/arguments_multi_layer_distill.py

from dataclasses import dataclass, field
from transformers import TrainingArguments
from typing import List


@dataclass
class ModelArguments:
    model_name: str = field(metadata={"help": "huggingface model name or path"})
    model_type: str = field(default=None, metadata={"help": "model type, typically includes in config file, but sometimes needs mannually add"})
    processor_name: str = field(default=None, metadata={"help": "processor_name, huggingface model name or path"})
    model_backbone: str = field(default=None, metadata={"help": "HF model type"})
    checkpoint_path: str = field(default=None, metadata={"help": "a local model path, could be a LoRA version"})
    pooling: str = field(default='last', metadata={"help": "pooling method for encoder"})
    normalize: bool = field(default=False, metadata={"help": "normalize query and passage representations"})
    temperature: float = field(default=0.02, metadata={"help": "temperature for softmax"})
    lora: bool = field(default=False, metadata={"help": "do parameter-efficient fine-tuning with lora"})
    lora_r: int = field(default=16, metadata={"help": "lora r"})
    lora_alpha: int = field(default=64, metadata={"help": "lora alpha"})
    lora_dropout: float = field(default=0.1, metadata={"help": "lora dropout"})
    lora_target_modules: str = field(default="qkv_proj,o_proj,gate_up_proj,down_proj,k_proj,q_proj,out_proj,v_proj", metadata={"help": "lora target modules"})
    num_crops: int = field(default=16, metadata={"help": "number of crops used in image encoder"})
    uigraph_use: bool = field(default=False, metadata={"help": "Enable ui graph for token selection"})
    uigraph_diff: int = field(default=1, metadata={"help": "Pixel difference used for constructing ui graph for token selection"})
    uigraph_rand: bool = field(default=False, metadata={"help": "Enable random graph construction for token selection"})
    uimask_ratio: float = field(default=0.5, metadata={"help": "Specify the percentage of patch tokens to skip per component for token selection"})
    uimask_rand: bool = field(default=False, metadata={"help": "Enable random token selection instead of uniform selection"})
    lm_skip_layer: str = field(default='[1,28,0]', metadata={"help": "Specify the layers of the language model to skip for token selection"})
    vis_skip_layer: str = field(default='[1,32,0]', metadata={"help": "Specify the layers of the vision model to skip for token selection"})
    supervise_layers: str = field(default=None, metadata={"help": "Comma-separated layer indices, e.g., '20,-1'"})
    supervise_weights: str = field(default=None, metadata={"help": "Comma-separated weights, e.g., '0.1,0.9'"})
    # 新增：按是否含图像做门控（对齐=两侧同为含图或同为不含图）
    gate_by_image: bool = field(default=True, metadata={"help": "Gate middle-layer contrastive by image presence alignment"})
    # 新增：非对齐样本是否保留一小部分中间层对比（0表示不做；可设0.05小试）
    misalign_mid_ce: float = field(default=0.0, metadata={"help": "Weight for middle-layer contrastive on misaligned pairs"})
    # 新增：自蒸馏系数（misaligned 上启用；若 distill_on_aligned=True 则对齐样本也启用）
    distill_beta: float = field(default=1.0, metadata={"help": "Weight for single-sample self-distillation on middle layers"})
    # 新增：对齐样本是否也做中间层自蒸馏（默认False）
    distill_on_aligned: bool = field(default=False, metadata={"help": "Also apply mid-layer self-distillation on aligned pairs"})
    # 数据集级：中间层对比（InfoNCE）的样本权重（默认1.0）
    ds_mid_ce: str = field(
        default="",
        metadata={"help": "Dataset->mid-layer CE weight mapping, e.g., 'WebQA:0.2,MSCOCO_t2i:0.15'"}
    )
    # 数据集级：中间层自蒸馏的样本权重/系数（默认1.0，作为 distill_beta 的样本级乘子）
    ds_distill_beta: str = field(
        default="",
        metadata={"help": "Dataset->mid-layer distill beta multiplier, e.g., 'FashionIQ:1.5,CIRR:1.2'"}
    )
    ce_stage1: float = field(default=0.3, metadata={"help": "progress < s1: only distill"})
    ce_stage2: float = field(default=0.7, metadata={"help": "progress > s2: CE fixed at ce_final"})
    schedule_mode: str = field(default="cosine", metadata={"help": "linear | cosine"})
    distill_floor: float = field(default=0.2, metadata={"help": "distill scale after s2"})
    ce_final: float = field(default=0.7, metadata={"help": "final CE scale after s2 (<1.0 keeps CE softer)"})
    # 固定权重混合（开启后覆盖时间日程）
    fixed_mix: bool = field(default=False, metadata={"help": "Use fixed mixture of mid CE and distill"})
    mid_ce_coef: float = field(default=1.0, metadata={"help": "Weight for middle-layer CE when fixed_mix=True"})
    mid_distill_coef: float = field(default=0.0, metadata={"help": "Weight for middle-layer distill when fixed_mix=True"})

@dataclass
class DataArguments:
    dataset_config: str = field(default=None, metadata={"help": "yaml file with dataset configuration"})
    data_basedir: str = field(default=None, metadata={"help": "Expect an absolute path to the base directory of all datasets. If set, it will be prepended to each dataset path"})
    dataset_name: str = field(default=None, metadata={"help": "huggingface dataset name"})
    subset_name: List[str] = field(default=None, metadata={"help": "Useful for datasets with subsets"})
    dataset_split: str = field(default='train', metadata={"help": "dataset split"})
    num_sample_per_subset: int = field(default=None, metadata={"help": "number of training samples per subset"})
    image_dir: str = field(default=None, metadata={"help": "Image directory path"})
    encode_output_path: str = field(default=None, metadata={"help": "encode output path"})
    max_len: int = field(default=None, metadata={"help": "The maximum total input sequence length after tokenization. Use with caution, since it may truncate text prompts due to large image lengths."},)
    embedding_type: str = field(default="", metadata={"help": "embedding type"})
    image_resolution: str = field(default=None, metadata={"help": "for models i.e. LLaVA-next and Qwen, resize images first, none means using original image resolution. This is only works when `--resize_use_processor false`."})
    resize_use_processor: bool = field(default=True, metadata={"help": "Resize visual inputs insides processor, e.g. Qwen2VLImageProcessor, instead of by our code."})
    resize_min_pixels: int = field(default=28*28*4, metadata={"help": "The min pixels of the image to resize the image. This is only works when `--resize_use_processor true`."})
    resize_max_pixels: int = field(default=28*28*1280, metadata={"help": "The max pixels of the image to resize the image. This is only works when `--resize_use_processor true`."})
    image_decay_factor: float = field(default=None, metadata={"help": "The image decay factor for resizing temporal images"})
    num_hardneg: int = field(default=0, metadata={"help": "hard negative number"})


@dataclass
class TrainingArguments(TrainingArguments):
    image_encoder_freeze: bool = field(default=False, metadata={"help": "huggingface model name"})
    output_dir: str = field(default=None, metadata={"help": "directory for saving trained models"})
    resume_from: str = field(default="none", metadata={"help": "`auto` will detect if any previous checkpoints should be resumed. or specify specific step of the checkpoint."})
    project_name: str = field(default=None, metadata={"help": "project name"})
    logging_steps: int = field(default=1, metadata={"help": "logging steps"})
    num_train_epochs: int = field(default=1, metadata={"help": "number of training epochs"})
    grad_cache: bool = field(default=False, metadata={"help": "Use gradient cache update"})
    gc_q_chunk_size: int = field(default=2, metadata={"help": "query side subset size"})
    gc_p_chunk_size: int = field(default=2, metadata={"help": "target side subset size"})
    interleave_stopping_strategy: str = field(default="all_exhausted", metadata={"help": "all_exhausted or first_exhausted"})
    interleave_batch_size: float = field(default=0, metadata={"help": "Specify mini-batch size to interleave data from multi-sources, 0/None means random sampling by examples, 1 means full batch."})

@dataclass
class MTEBArguments:
    device: str = field(default="cuda", metadata={"help": "use cuda for single GPU inference, if multiple GPUs are available it will use DP automatically"})
    batch_size_per_device: int = field(default=16, metadata={"help": ""})
    max_length: int = field(default=512, metadata={"help": ""})
    eval_output_dir: str = field(default=None, metadata={"help": "directory for saving trained models"})
    task_types: List[str] = field(default=None, metadata={"help": ""})
    tasks: List[str] = field(default=None, metadata={"help": ""})
    prompt_family: List[str] = field(default=None, metadata={"help": ""})