image_max_pixels

image_max_pixels， 这里的 area 就是宽*高

python
# 在src/llamafactory/data/mm_plugin.py中定义
# 控制图像处理时的最大像素数量
# 如果图像超过这个像素数，会被调整大小
def get_image_processor_preprocess_params(image_processor):
    params = {}
    if hasattr(image_processor, "crop_size"):  # for CLIP
        params["crop_size"] = {
            "height": image_processor.crop_size["height"],
            "width": image_processor.crop_size["width"]
        }
        params["size"] = max(params["crop_size"]["height"], params["crop_size"]["width"])
    elif hasattr(image_processor, "size"):
        if isinstance(image_processor.size, dict):  # for Qwen
            params["size"] = image_processor.size["max_edge"]
        elif isinstance(image_processor.size, list):  # for InternVL
            params["size"] = image_processor.size
        else:
            params["size"] = image_processor.size
    
    # image_max_pixels用于确保图像不会太大，超出内存
    if params.get("size", None) is not None and hasattr(image_processor, "image_max_pixels"):
        area = image_processor.image_max_pixels  # 来自training_args.image_max_pixels
        size = int(math.sqrt(area))
        params["size"] = size

    return params

LLaMA-Factory 源码信息

register_model_group(
    models={
        "InternVL2.5-1B-MPO": {
            DownloadSource.DEFAULT: "kingsley01/InternVL2_5-1B-MPO-hf",
            DownloadSource.MODELSCOPE: "llamafactory/InternVL2_5-1B-MPO-hf",
        },
        "InternVL2.5-2B-MPO": {
            DownloadSource.DEFAULT: "kingsley01/InternVL2_5-2B-MPO-hf",
            DownloadSource.MODELSCOPE: "llamafactory/InternVL2_5-2B-MPO-hf",
        },
        "InternVL2.5-4B-MPO": {
            DownloadSource.DEFAULT: "kingsley01/InternVL2_5-4B-MPO-hf",
            DownloadSource.MODELSCOPE: "llamafactory/InternVL2_5-4B-MPO-hf",
        },
        "InternVL2.5-8B-MPO": {
            DownloadSource.DEFAULT: "kingsley01/InternVL2_5-8B-MPO-hf",
            DownloadSource.MODELSCOPE: "llamafactory/InternVL2_5-8B-MPO-hf",
        },
        "InternVL3-1B-hf": {
            DownloadSource.DEFAULT: "kingsley01/InternVL3-1B-hf",
            DownloadSource.MODELSCOPE: "llamafactory/InternVL3-1B-hf",
        },
        "InternVL3-2B-hf": {
            DownloadSource.DEFAULT: "kingsley01/InternVL3-2B-hf",
            DownloadSource.MODELSCOPE: "llamafactory/InternVL3-2B-hf",
        },
        "InternVL3-8B-hf": {
            DownloadSource.DEFAULT: "kingsley01/InternVL3-8B-hf",
            DownloadSource.MODELSCOPE: "llamafactory/InternVL3-8B-hf",
        },
    },
    template="intern_vl",
    multimodal=True,
)

InternVL 1技术深度分析

1. 引言

InternVL（Internal Vision-Language model）是一个开源的多模态大型模型项目，由上海人工智能实验室（OpenGVLab）开发。InternVL 1是该项目的第一个主要版本，它通过创新的视觉-语言融合方法，实现了强大的图像理解和多模态对话能力。本文将深入分析InternVL 1的技术架构、关键特性和创新点，以提供对该模型的全面了解。

该文章已加密，点击 阅读全文 并输入密码后方可查看。

Spatial Layout Projector (SLP)

1. InternVL采用了一种称为"Spatial Layout Projector (SLP)"的方法，将四维的空间坐标[x1,y1,x2,y2]（一个bounding box）转换为单个token嵌入：

   "A key innovation in LayTextLLM is the Spatial Layout Projector (SLP), which transforms a spatial layout into a singular bounding box token. This enhancement enables the model to process both spatial layouts and textual inputs simultaneously. To be specifically, each OCR-derived spatial layout is represented by a bounding box defined by four-dimensional coordinates [x1,y1,x2,y2]..."

2. 这种方法确实将每个边界框（box）表示为一个token，不同于之前的"coordinate-as-tokens"方案（这种方案会将坐标转换为多个token）：

   "Compared to the coordinate-as-tokens scheme, the SLP represents each bounding box with a single token. This approach significantly reduces the number of input tokens..."

3. 这种单token表示法的计算方式是通过将坐标映射到高维空间来实现的：

   "The process can be computed as z=W⋅c+b, where c∈ℝ^4 is the vector of the bounding box coordinates. W∈ℝ^(d×4) is a weight matrix with d represents the dimension of the embedding, b∈ℝ^(d×1) is a bias vector, z is the resulting bounding box token represented as an d-dimensional embedding."