--- title: ofa_pretrain_base_zh canonical_url: "https://www.modelscope.cn/models/damo/ofa_pretrain_base_zh" md_url: "https://www.modelscope.cn/models/damo/ofa_pretrain_base_zh.md" repository: damo/ofa_pretrain_base_zh chinese_name: "OFA预训练模型-中文-通用领域-base" last_updated: 2025-06-12 license: "Apache License 2.0" model_type: - ofa architectures: - OFAModel library_name: - pytorch frameworks: - pytorch domain: - multi-modal downloads: 5310 stars: 11 tags: - Alibaba - OFA - M6 - ICML2022 - "arxiv:2202.03052" - "预训练模型" - "多模态" --- # ofa_pretrain_base_zh > ofa_pretrain_base_zh - damo 在 ModelScope 开源的模型。OFA的预训练ckpt，能够在完全不改变模型结构的情况下进行下游任务的finetune，是finetune的基础ckpt。 damo/ofa_pretrain_base_zh 是 ModelScope 魔搭社区上的机器学习模型，采用 Apache License 2.0 许可。 - **Repository**: damo/ofa_pretrain_base_zh - **License**: Apache License 2.0 - **Tags**: Alibaba, OFA, M6, ICML2022, arxiv:2202.03052, 预训练模型, 多模态 - **Downloads**: 5310 - **Stars**: 11 - **Last updated**: 2025-06-12 Source: https://www.modelscope.cn/models/damo/ofa_pretrain_base_zh --- # OFA预训练模型 ## News - 2023年1月： - 优化了finetune流程，支持参数更新、自定义数据及脚本分布式训练等，见finetune示例。 - 2022年11月： - 发布ModelScope 1.0版本，以下能力请使用1.0.2及以上版本。 - 新增[OFA Tutorial](https://modelscope.cn/docs/OFA_Tutorial#1.4%20%E5%A6%82%E4%BD%95%E8%AE%AD%E7%BB%83)。 OFA预训练模型是OFA在8个预训练任务（具体参见论文）上得到ckpt，是finetune下游任务的基础。 ## 如何使用OFA预训练模型 ### Finetune 模型细节参考文档：[OFA Tutorial](https://modelscope.cn/docs/OFA_Tutorial#1.4%20%E5%A6%82%E4%BD%95%E8%AE%AD%E7%BB%83) 1.4节。 这里直接使用预训练模型在caption任务上进行实例演示，要求 ModelScope Library >= 1.2.0 ```python import tempfile from modelscope.msdatasets import MsDataset from modelscope.metainfo import Trainers from modelscope.trainers import build_trainer from modelscope.utils.constant import DownloadMode from modelscope.utils.hub import snapshot_download train_dataset = MsDataset( MsDataset.load( "coco_2014_caption", namespace="modelscope", split="train[:100]", download_mode=DownloadMode.REUSE_DATASET_IF_EXISTS).remap_columns({ 'image': 'image', 'caption': 'text' })) test_dataset = MsDataset( MsDataset.load( "coco_2014_caption", namespace="modelscope", split="validation[:20]", download_mode=DownloadMode.REUSE_DATASET_IF_EXISTS).remap_columns({ 'image': 'image', 'caption': 'text' })) def cfg_modify_fn(cfg): cfg.train.hooks = [{ 'type': 'CheckpointHook', 'interval': 2 }, { 'type': 'TextLoggerHook', 'interval': 1 }, { 'type': 'IterTimerHook' }] cfg.train.max_epochs=2 return cfg pretrained_model = 'damo/ofa_pretrain_base_zh' pretrain_path = snapshot_download(pretrained_model, revision='v1.0.2') args = dict( model=pretrain_path, train_dataset=train_dataset, eval_dataset=test_dataset, cfg_modify_fn=cfg_modify_fn, work_dir = tempfile.TemporaryDirectory().name) trainer = build_trainer(name=Trainers.ofa, default_args=args) trainer.train() ``` ### ZeroShot OFA预训练模型可以通过以下代码进行下载： 注：modelscope V1.0.0开始使用tag进行兼容性管理，目前策略是不指定tag的情况下使用modelscope版本发布前的最近的tag。因为预训练模型在modelscope==1.0.2版本后发布的，所以使用1.0.2版本的同学需要加上revision信息。 ```python from modelscope.utils.hub import snapshot_download pretrained_model = 'damo/ofa_pretrain_base_zh' pretrain_path = snapshot_download(pretrained_model, revision='master') ``` 有了预训练模型，可以OFA自身的能力特点（One For All）利用预训练模型测试其在下游任务的效果，替换configuration文件示例： ```python import os import shutil from modelscope.utils.hub import snapshot_download from modelscope.pipelines import pipeline from modelscope.utils.constant import Tasks from modelscope.outputs import OutputKeys from modelscope.utils.constant import ModelFile pretrained_model = 'damo/ofa_pretrain_base_zh' pretrain_path = snapshot_download(pretrained_model, revision='master') task_model = 'damo/ofa_image-caption_muge_base_zh' task_path = snapshot_download(task_model, revision='master') shutil.copy(os.path.join(task_path, ModelFile.CONFIGURATION), # 将任务的配置覆盖预训练模型的配置 os.path.join(pretrain_path, ModelFile.CONFIGURATION)) ofa_pipe = pipeline(Tasks.image_captioning, model=pretrain_path) result = ofa_pipe('https://modelscope.oss-cn-beijing.aliyuncs.com/demo/image-captioning/donuts.jpeg') print(result[OutputKeys.CAPTION]) ``` ## OFA是什么？ OFA(One-For-All)是通用多模态预训练模型，使用简单的序列到序列的学习框架统一模态（跨模态、视觉、语言等模态）和任务（如图片生成、视觉定位、图片描述、图片分类、文本生成等），详见我们发表于ICML 2022的论文：[OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework](https://arxiv.org/abs/2202.03052)，以及我们的官方Github仓库[https://github.com/OFA-Sys/OFA](https://github.com/OFA-Sys/OFA)。




Github  ｜  Paper   ｜  Blog



## OFA模型规模
ModelParams-enParams-zhBackboneHidden sizeIntermediate sizeNum. of headsEnc layersDec layers
OFATiny33M-ResNet502561024444
OFAMedium93M-ResNet1015122048844
OFABase180M160MResNet10176830721266
OFALarge470M440MResNet15210244096161212
OFAHuge930M-ResNet15212805120162412

## OFA 模型任务矩阵 目前ModelScope上面所有已经上传的模型和任务可以在下面导航表格看到，点击链接可以跳转到相应modelcard。 | 模型规模 | 预训练 | 图像描述 | 视觉问答 | 视觉定位 | 视觉蕴含 | 文生图 | 图像分类 | 文字识别 | 文本摘要 | 文本分类 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | OFATiny | [英文](https://modelscope.cn/models/damo/ofa_pretrain_tiny_en/summary) | [英文](https://modelscope.cn/models/damo/ofa_image-caption_coco_distilled_en/summary) | - | [英文](https://modelscope.cn/models/damo/ofa_visual-grounding_refcoco_distilled_en) | [英文](https://modelscope.cn/models/damo/ofa_visual-entailment_snli-ve_distilled_v2_en/summary) | - | - | - | - | - | | OFAMedium | [英文](https://modelscope.cn/models/damo/ofa_pretrain_medium_en/summary) | - | - | - | - | - | - | - | - | - | | OFABase | [中文](https://modelscope.cn/models/damo/ofa_pretrain_base_zh/summary)/[英文](https://modelscope.cn/models/damo/ofa_pretrain_base_en/summary) | [中文电商](https://modelscope.cn/models/damo/ofa_image-caption_muge_base_zh/summary) | - | - | - | - | - | [场景中文](https://modelscope.cn/models/damo/ofa_ocr-recognition_scene_base_zh/summary) | - | - | | OFALarge | [中文](https://modelscope.cn/models/damo/ofa_pretrain_large_zh/summary)/[英文](https://modelscope.cn/models/damo/ofa_pretrain_large_en/summary) | [英文](https://modelscope.cn/models/damo/ofa_image-caption_coco_large_en/summary) | [英文](https://modelscope.cn/models/damo/ofa_visual-question-answering_pretrain_large_en/summary) | [中文](https://modelscope.cn/models/damo/ofa_visual-grounding_refcoco_large_zh/summary)/[英文](https://modelscope.cn/models/damo/ofa_visual-grounding_refcoco_large_en/summary) | [英文](https://modelscope.cn/models/damo/ofa_visual-entailment_snli-ve_large_en/summary) | [英文](https://modelscope.cn/models/damo/ofa_text-to-image-synthesis_coco_large_en/summary) | [英文](https://modelscope.cn/models/damo/ofa_image-classification_imagenet_large_en/summary) | - | [英文](https://modelscope.cn/models/damo/ofa_summarization_gigaword_large_en/summary) | [英文](https://modelscope.cn/models/damo/ofa_text-classification_mnli_large_en/summary) | | OFAHuge | [英文](https://modelscope.cn/models/damo/ofa_pretrain_huge_en/summary) | [英文](https://modelscope.cn/models/damo/ofa_image-caption_coco_huge_en/summary) | [英文](https://modelscope.cn/models/damo/ofa_visual-question-answering_pretrain_huge_en/summary) | - | - | - | - | - | - | - | | OFA6B | - | [英文](https://modelscope.cn/models/damo/ofa_image-caption_coco_6b_en/summary) | - | - | - | - | - | - | - | - | ## 相关论文以及引用 如果你觉得OFA好用，喜欢我们的工作，欢迎引用： ``` @article{wang2022ofa, author = {Peng Wang and An Yang and Rui Men and Junyang Lin and Shuai Bai and Zhikang Li and Jianxin Ma and Chang Zhou and Jingren Zhou and Hongxia Yang}, title = {OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework}, journal = {CoRR}, volume = {abs/2202.03052}, year = {2022} } ```