{"id":19320175,"url":"https://github.com/invictus717/mico","last_synced_at":"2025-03-15T14:09:37.457Z","repository":{"id":244259054,"uuid":"813507636","full_name":"invictus717/MiCo","owner":"invictus717","description":"Explore the Limits of Omni-modal Pretraining at Scale","archived":false,"fork":false,"pushed_at":"2024-09-02T09:22:37.000Z","size":12152,"stargazers_count":96,"open_issues_count":5,"forks_count":4,"subscribers_count":2,"default_branch":"master","last_synced_at":"2025-02-28T19:54:39.953Z","etag":null,"topics":["deep-learning","multimodal","multimodal-large-language-models","omnimodal","pretraining","scale-up"],"latest_commit_sha":null,"homepage":"https://invictus717.github.io/MiCo/","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"apache-2.0","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/invictus717.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null}},"created_at":"2024-06-11T08:14:04.000Z","updated_at":"2024-12-18T09:00:42.000Z","dependencies_parsed_at":"2024-06-28T09:33:31.419Z","dependency_job_id":"e1e3e82a-ba7c-43d5-801b-1f99c7e746b2","html_url":"https://github.com/invictus717/MiCo","commit_stats":null,"previous_names":["invictus717/mico"],"tags_count":0,"template":false,"template_full_name":null,"repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/invictus717%2FMiCo","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/invictus717%2FMiCo/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/invictus717%2FMiCo/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/invictus717%2FMiCo/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/invictus717","download_url":"https://codeload.github.com/invictus717/MiCo/tar.gz/refs/heads/master","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":243738985,"owners_count":20340002,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["deep-learning","multimodal","multimodal-large-language-models","omnimodal","pretraining","scale-up"],"created_at":"2024-11-10T01:27:24.316Z","updated_at":"2025-03-15T14:09:37.439Z","avatar_url":"https://github.com/invictus717.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"\u003cp align=\"center\" width=\"100%\"\u003e\n\u003cimg src=\"assets/banner.png\"  width=\"100%\" height=\"80%\"\u003e\n\u003c/p\u003e\n\n[![arXiv](https://img.shields.io/badge/arxiv-2406.09412-b31b1b?style=plastic\u0026color=b31b1b\u0026link=https%3A%2F%2Farxiv.org%2Fabs%2F2406.09412)](https://arxiv.org/abs/2406.09412)\n[![website](https://img.shields.io/badge/Project-Website-purple)](https://invictus717.github.io/MiCo/)\n[![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/Yiyuan/MiCo-ViT-g-14-omnimodal-300k-b64K)\n\u003ca href=\"#LICENSE--citation\"\u003e\n  \u003cimg alt=\"License: Apache2.0\" src=\"https://img.shields.io/badge/LICENSE-Apache%202.0-blue.svg\"/\u003e\n\u003c/a\u003e\n\u003c!-- \u003ca href=\"https://twitter.com/_akhaliq/status/1682248055637041152\"\u003e\u003cimg src=\"https://img.shields.io/twitter/url/https/twitter.com/cloudposse.svg?style=social\"\u003e\u003c/a\u003e\n\u003ca href=\"https://www.youtube.com/watch?v=V8L8xbsTyls\u0026ab_channel=CSBoard\"\u003e\u003cimg src=\"https://img.icons8.com/color/48/000000/youtube-play.png\" width=\"28\" height=\"23\"\u003e\u003c/a\u003e  --\u003e\n\n\n\n### ✨ Inspiration of Multimodal Context: Multimedia Brain Cognition\n\n\u003cp align=\"center\" width=\"100%\"\u003e\n\u003cimg src=\"assets/brain.png\"  width=\"100%\" height=\"60%\"\u003e\n\u003c/p\u003e\n\n***How the human brain performs coherent multimodal cognition?***\n\nAs outlined in Richard Mayer's Cognitive Theory of Multimedia Learning,our brain processes multimedia signals through two distinct channels—auditory and visual—in sensory memory, as depicted in Figure(a). The sensory memory integrates these signals with prior knowledge through words, transforming new multimedia information into long-term memory. Notably, **1**) multimedia signals in the brain share channels, and **2**) words function as the reasoning interface in our brain.\n\nInspired by these insights, we categorize diverse modalities into two types: ``knowledge modality`` and ``interface modality``. *Knowledge modalities*, primarily derived from raw sensors, contribute knowledge in diverse formats. For example, images and depth maps offer visual knowledge, while audio and video provide auditory and spatiotemporal knowledge. The language modality, developed by humans, is inherently more abstract and naturally functions as the *interface modality*, facilitating learning, reasoning, and the coordination of knowledge. To this end, we design an omni-modal learning architecture, illustrated in Figure (b), with two distinct branches: one for knowledge modalities and one for the interface modality, *i.e.* natural language. The knowledge and interface modalities are aligned through a novel generative reasoning method.\n\n### 🚀 MiCo, An omni-modal and scalable pretraining paradigm\n\n\u003cp align=\"center\" width=\"100%\"\u003e\n\u003cimg src=\"assets/omnimodal_pretraining.png\"  width=\"100%\" height=\"60%\"\u003e\n\u003c/p\u003e\n\nWe propose collecting large-scale omni-modal paired data, including text,\nimage, video, depth, and normal maps, to learn universal representations.\n\n\u003cp align=\"center\" width=\"100%\"\u003e\n\u003cimg src=\"assets/paradigm.png\"  width=\"100%\" height=\"60%\"\u003e\n\u003c/p\u003e\n\n**🚀 Evolution of Pretraining Paradigms**. Masked modeling (a) has shown great success in single modality, general-purpose understanding. Contrastive learning (b) distinguishes transferable features with modality tuples (such as text-image, text-video, text-audio, etc). \n\n*🚀🚀🚀 We aim to achieve general-purpose omni-modal understanding and learn transferable, universal representations in (c).*\n\n### 🌟🌟🌟 The Multimodal Scaling Laws with MiCo: Modalities Help Modalies!\n\n\u003cp align=\"center\" width=\"100%\"\u003e\n\u003cimg src=\"assets/scaling_laws.png\"  width=\"100%\" height=\"60%\"\u003e\n\u003c/p\u003e\n\n### 🔓 Pretrained Omni-Modal Models\n\u003c!-- \u003cdetails\u003e --\u003e\n**We will continue to update this model zoo including all scales of ViTs and highly-efficient ConvNets with the MiCo pretraining paradigm**\n\n\u003csummary\u003e Current Checkpoints \u003c/summary\u003e\n\u003cbr\u003e\n\u003cdiv\u003e\n\n|      Model      |   Pretraining   | Scale |  Modality | #Param |                                               Google Drive | Hugging Face\n| :------------: | :----------: | :----------------------: | :----: | :---------------------------------------------------------------------------------------------------: |:----: | :----: | \n| MiCo  | 300k steps |         ViT-g          | Omni-modal |  1.3B  |   [ckpt](https://drive.google.com/drive/folders/1AIQjV1KU8K4OXiO-4gFirxkoxt3twWIq?usp=sharing)    | [ckpt](https://huggingface.co/Yiyuan/MiCo-ViT-g-14-omnimodal-300k-b64K)\n\n\n\u003c/div\u003e\n\n### 🔓 Omni-Modal Dataset Collection\n\nWe provdie a detailed [doc](data/README.md) for preparing the omni-modal dataset step-by-step\n\n### ⚡ Quick Start\n```bash\npip install gdown \ngdown 1AIQjV1KU8K4OXiO-4gFirxkoxt3twWIq --folder\npython inference_demo.py\n```\n# Citation\nIf the code and paper help your research, please kindly cite:\n```\n@article{zhang2024explore,\n  title={Explore the Limits of Omni-modal Pretraining at Scale},\n  author={Zhang, Yiyuan and Li, Handong and Liu, Jing and Yue, Xiangyu},\n  journal={arXiv preprint arXiv:2406.09412},\n  year={2024}\n}\n```\n# License\nThis project is released under the [Apache 2.0 license](LICENSE).\n# Acknowledgement\nWe appreciate [Dr. Xiaohan Ding](https://dingxiaohan.xyz/) for the valuable discussion and suggestions.This code is developed based [Meta-Transformer](https://github.com/invictus717/MetaTransformer), [VAST](https://github.com/TXH-mercury/VAST), [DPT](https://github.com/EPFL-VILAB/omnidata), and [GeoWizard](https://github.com/fuxiao0719/GeoWizard).","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Finvictus717%2Fmico","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Finvictus717%2Fmico","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Finvictus717%2Fmico/lists"}