[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-fastai--fastai":3,"tool-fastai--fastai":61},[4,18,26,36,44,53],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":17},4358,"openclaw","openclaw\u002Fopenclaw","OpenClaw 是一款专为个人打造的本地化 AI 助手，旨在让你在自己的设备上拥有完全可控的智能伙伴。它打破了传统 AI 助手局限于特定网页或应用的束缚，能够直接接入你日常使用的各类通讯渠道，包括微信、WhatsApp、Telegram、Discord、iMessage 等数十种平台。无论你在哪个聊天软件中发送消息，OpenClaw 都能即时响应，甚至支持在 macOS、iOS 和 Android 设备上进行语音交互，并提供实时的画布渲染功能供你操控。\n\n这款工具主要解决了用户对数据隐私、响应速度以及“始终在线”体验的需求。通过将 AI 部署在本地，用户无需依赖云端服务即可享受快速、私密的智能辅助，真正实现了“你的数据，你做主”。其独特的技术亮点在于强大的网关架构，将控制平面与核心助手分离，确保跨平台通信的流畅性与扩展性。\n\nOpenClaw 非常适合希望构建个性化工作流的技术爱好者、开发者，以及注重隐私保护且不愿被单一生态绑定的普通用户。只要具备基础的终端操作能力（支持 macOS、Linux 及 Windows WSL2），即可通过简单的命令行引导完成部署。如果你渴望拥有一个懂你",349277,3,"2026-04-06T06:32:30",[13,14,15,16],"Agent","开发框架","图像","数据工具","ready",{"id":19,"name":20,"github_repo":21,"description_zh":22,"stars":23,"difficulty_score":10,"last_commit_at":24,"category_tags":25,"status":17},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,"2026-04-05T11:01:52",[14,15,13],{"id":27,"name":28,"github_repo":29,"description_zh":30,"stars":31,"difficulty_score":32,"last_commit_at":33,"category_tags":34,"status":17},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",142651,2,"2026-04-06T23:34:12",[14,13,35],"语言模型",{"id":37,"name":38,"github_repo":39,"description_zh":40,"stars":41,"difficulty_score":32,"last_commit_at":42,"category_tags":43,"status":17},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",107888,"2026-04-06T11:32:50",[14,15,13],{"id":45,"name":46,"github_repo":47,"description_zh":48,"stars":49,"difficulty_score":32,"last_commit_at":50,"category_tags":51,"status":17},4721,"markitdown","microsoft\u002Fmarkitdown","MarkItDown 是一款由微软 AutoGen 团队打造的轻量级 Python 工具，专为将各类文件高效转换为 Markdown 格式而设计。它支持 PDF、Word、Excel、PPT、图片（含 OCR）、音频（含语音转录）、HTML 乃至 YouTube 链接等多种格式的解析，能够精准提取文档中的标题、列表、表格和链接等关键结构信息。\n\n在人工智能应用日益普及的今天，大语言模型（LLM）虽擅长处理文本，却难以直接读取复杂的二进制办公文档。MarkItDown 恰好解决了这一痛点，它将非结构化或半结构化的文件转化为模型“原生理解”且 Token 效率极高的 Markdown 格式，成为连接本地文件与 AI 分析 pipeline 的理想桥梁。此外，它还提供了 MCP（模型上下文协议）服务器，可无缝集成到 Claude Desktop 等 LLM 应用中。\n\n这款工具特别适合开发者、数据科学家及 AI 研究人员使用，尤其是那些需要构建文档检索增强生成（RAG）系统、进行批量文本分析或希望让 AI 助手直接“阅读”本地文件的用户。虽然生成的内容也具备一定可读性，但其核心优势在于为机器",93400,"2026-04-06T19:52:38",[52,14],"插件",{"id":54,"name":55,"github_repo":56,"description_zh":57,"stars":58,"difficulty_score":10,"last_commit_at":59,"category_tags":60,"status":17},4487,"LLMs-from-scratch","rasbt\u002FLLMs-from-scratch","LLMs-from-scratch 是一个基于 PyTorch 的开源教育项目，旨在引导用户从零开始一步步构建一个类似 ChatGPT 的大型语言模型（LLM）。它不仅是同名技术著作的官方代码库，更提供了一套完整的实践方案，涵盖模型开发、预训练及微调的全过程。\n\n该项目主要解决了大模型领域“黑盒化”的学习痛点。许多开发者虽能调用现成模型，却难以深入理解其内部架构与训练机制。通过亲手编写每一行核心代码，用户能够透彻掌握 Transformer 架构、注意力机制等关键原理，从而真正理解大模型是如何“思考”的。此外，项目还包含了加载大型预训练权重进行微调的代码，帮助用户将理论知识延伸至实际应用。\n\nLLMs-from-scratch 特别适合希望深入底层原理的 AI 开发者、研究人员以及计算机专业的学生。对于不满足于仅使用 API，而是渴望探究模型构建细节的技术人员而言，这是极佳的学习资源。其独特的技术亮点在于“循序渐进”的教学设计：将复杂的系统工程拆解为清晰的步骤，配合详细的图表与示例，让构建一个虽小但功能完备的大模型变得触手可及。无论你是想夯实理论基础，还是为未来研发更大规模的模型做准备",90106,"2026-04-06T11:19:32",[35,15,13,14],{"id":62,"github_repo":63,"name":64,"description_en":65,"description_zh":66,"ai_summary_zh":66,"readme_en":67,"readme_zh":68,"quickstart_zh":69,"use_case_zh":70,"hero_image_url":71,"owner_login":64,"owner_name":72,"owner_avatar_url":73,"owner_bio":74,"owner_company":75,"owner_location":75,"owner_email":75,"owner_twitter":76,"owner_website":77,"owner_url":78,"languages":79,"stars":92,"forks":93,"last_commit_at":94,"license":95,"difficulty_score":32,"env_os":96,"env_gpu":97,"env_ram":98,"env_deps":99,"category_tags":104,"github_topics":105,"view_count":32,"oss_zip_url":75,"oss_zip_packed_at":75,"status":17,"created_at":113,"updated_at":114,"faqs":115,"releases":144},4698,"fastai\u002Ffastai","fastai","The fastai deep learning library","fastai 是一个基于 PyTorch 构建的深度学习库，旨在让开发者既能快速获得业界领先的模型效果，又能灵活地进行底层定制。它主要解决了深度学习门槛高、代码重复繁琐以及难以兼顾易用性与灵活性的痛点，让用户仅用几行代码即可完成图像分类、文本情感分析、推荐系统等复杂任务的建模与训练。\n\n无论是刚入门的学生、希望快速验证想法的从业者，还是需要探索新算法的研究人员，fastai 都能提供合适的支持。对于初学者，它提供了高层组件和配套免费课程，帮助快速上手；对于专家，它开放了底层接口，支持自由组合模块以构建创新架构。\n\nfastai 的独特之处在于其精心设计的分层架构：通过新型类型分发系统、双向回调机制（可在训练任意阶段干预数据、模型或优化器）以及高度简化的优化器实现，将复杂的深度学习模式抽象为简洁易用的代码。这种设计既保留了 Python 的动态特性，又充分发挥了 PyTorch 的灵活性，真正实现了“易于上手”与“深度可黑盒”的完美平衡。","# Welcome to fastai\n\n\n\u003C!-- WARNING: THIS FILE WAS AUTOGENERATED! DO NOT EDIT! -->\n\n[![CI](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Factions\u002Fworkflows\u002Fmain.yml\u002Fbadge.svg)](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Factions\u002Fworkflows\u002Fmain.yml)\n[![PyPI](https:\u002F\u002Fimg.shields.io\u002Fpypi\u002Fv\u002Ffastai?color=blue&label=pypi%20version.png)](https:\u002F\u002Fpypi.org\u002Fproject\u002Ffastai\u002F#description)\n[![Conda (channel\nonly)](https:\u002F\u002Fimg.shields.io\u002Fconda\u002Fvn\u002Ffastai\u002Ffastai?color=seagreen&label=conda%20version.png)](https:\u002F\u002Fanaconda.org\u002Ffastai\u002Ffastai)\n\n## Installing\n\nYou can use fastai without any installation by using [Google\nColab](https:\u002F\u002Fcolab.research.google.com\u002F). In fact, every page of this\ndocumentation is also available as an interactive notebook - click “Open\nin colab” at the top of any page to open it (be sure to change the Colab\nruntime to “GPU” to have it run fast!) See the fast.ai documentation on\n[Using Colab](https:\u002F\u002Fcourse19.fast.ai\u002Fstart_colab.html) for more\ninformation.\n\nYou can install fastai on your own machines with: `pip install fastai`.\n\nTo ensure that you have the best available version of PyTorch on your\nmachine, recommend\n[installing](https:\u002F\u002Fpytorch.org\u002Fget-started\u002Flocally\u002F) that first.\n\nIf you plan to develop fastai yourself, or want to be on the cutting\nedge, you can use an editable install (if you do this, you should also\nuse an editable install of\n[fastcore](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastcore) to go with it.) First\ninstall PyTorch, and then:\n\n    git clone https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\n    pip install -e \"fastai[dev]\"\n\n## Learning fastai\n\nThe best way to get started with fastai (and deep learning) is to read\n[the\nbook](https:\u002F\u002Fwww.amazon.com\u002FDeep-Learning-Coders-fastai-PyTorch\u002Fdp\u002F1492045527),\nand complete [the free course](https:\u002F\u002Fcourse.fast.ai).\n\nTo see what’s possible with fastai, take a look at the [Quick\nStart](https:\u002F\u002Fdocs.fast.ai\u002Fquick_start.html), which shows how to use\naround 5 lines of code to build an image classifier, an image\nsegmentation model, a text sentiment model, a recommendation system, and\na tabular model. For each of the applications, the code is much the\nsame.\n\nRead through the [Tutorials](https:\u002F\u002Fdocs.fast.ai\u002Ftutorial.html) to\nlearn how to train your own models on your own datasets. Use the\nnavigation sidebar to look through the fastai documentation. Every\nclass, function, and method is documented here.\n\nTo learn about the design and motivation of the library, read the [peer\nreviewed paper](https:\u002F\u002Fwww.mdpi.com\u002F2078-2489\u002F11\u002F2\u002F108\u002Fhtm).\n\n## About fastai\n\nfastai is a deep learning library which provides practitioners with\nhigh-level components that can quickly and easily provide\nstate-of-the-art results in standard deep learning domains, and provides\nresearchers with low-level components that can be mixed and matched to\nbuild new approaches. It aims to do both things without substantial\ncompromises in ease of use, flexibility, or performance. This is\npossible thanks to a carefully layered architecture, which expresses\ncommon underlying patterns of many deep learning and data processing\ntechniques in terms of decoupled abstractions. These abstractions can be\nexpressed concisely and clearly by leveraging the dynamism of the\nunderlying Python language and the flexibility of the PyTorch library.\nfastai includes:\n\n- A new type dispatch system for Python along with a semantic type\n  hierarchy for tensors\n- A GPU-optimized computer vision library which can be extended in pure\n  Python\n- An optimizer which refactors out the common functionality of modern\n  optimizers into two basic pieces, allowing optimization algorithms to\n  be implemented in 4–5 lines of code\n- A novel 2-way callback system that can access any part of the data,\n  model, or optimizer and change it at any point during training\n- A new data block API\n- And much more…\n\nfastai is organized around two main design goals: to be approachable and\nrapidly productive, while also being deeply hackable and configurable.\nIt is built on top of a hierarchy of lower-level APIs which provide\ncomposable building blocks. This way, a user wanting to rewrite part of\nthe high-level API or add particular behavior to suit their needs does\nnot have to learn how to use the lowest level.\n\n\u003Cimg alt=\"Layered API\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Ffastai_fastai_readme_9ca41d064328.png\" width=\"345\">\n\n## Migrating from other libraries\n\nIt’s very easy to migrate from plain PyTorch, Ignite, or any other\nPyTorch-based library, or even to use fastai in conjunction with other\nlibraries. Generally, you’ll be able to use all your existing data\nprocessing code, but will be able to reduce the amount of code you\nrequire for training, and more easily take advantage of modern best\npractices. Here are migration guides from some popular libraries to help\nyou on your way:\n\n- [Plain PyTorch](https:\u002F\u002Fdocs.fast.ai\u002Fexamples\u002Fmigrating_pytorch.html)\n- [Ignite](https:\u002F\u002Fdocs.fast.ai\u002Fexamples\u002Fmigrating_ignite.html)\n- [Lightning](https:\u002F\u002Fdocs.fast.ai\u002Fexamples\u002Fmigrating_lightning.html)\n- [Catalyst](https:\u002F\u002Fdocs.fast.ai\u002Fexamples\u002Fmigrating_catalyst.html)\n\n## Windows Support\n\nDue to python multiprocessing issues on Jupyter and Windows,\n`num_workers` of `Dataloader` is reset to 0 automatically to avoid\nJupyter hanging. This makes tasks such as computer vision in Jupyter on\nWindows many times slower than on Linux. This limitation doesn’t exist\nif you use fastai from a script.\n\nSee [this\nexample](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fblob\u002Fmaster\u002Fnbs\u002Fexamples\u002Fdataloader_spawn.py)\nto fully leverage the fastai API on Windows.\n\nWe recommend using Windows Subsystem for Linux (WSL) instead – if you do\nthat, you can use the regular Linux installation approach, and you won’t\nhave any issues with `num_workers`.\n\n## Tests\n\nTo run the tests in parallel, launch:\n\n`nbdev_test`\n\nFor all the tests to pass, you’ll need to install the dependencies\nspecified as part of dev_requirements in settings.ini\n\n`pip install -e .[dev]`\n\nTests are written using `nbdev`, for example see the documentation for\n`test_eq`.\n\n## Contributing\n\nAfter you clone this repository, make sure you have run\n`nbdev_install_hooks` in your terminal. This install Jupyter and git\nhooks to automatically clean, trust, and fix merge conflicts in\nnotebooks.\n\nAfter making changes in the repo, you should run `nbdev_prepare` and\nmake additional and necessary changes in order to pass all the tests.\n\n## Docker Containers\n\nFor those interested in official docker containers for this project,\nthey can be found\n[here](https:\u002F\u002Fgithub.com\u002Ffastai\u002Fdocker-containers#fastai).\n","# 欢迎来到 fastai\n\n\n\u003C!-- 警告：此文件由自动化工具生成！请勿编辑！ -->\n\n[![CI](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Factions\u002Fworkflows\u002Fmain.yml\u002Fbadge.svg)](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Factions\u002Fworkflows\u002Fmain.yml)\n[![PyPI](https:\u002F\u002Fimg.shields.io\u002Fpypi\u002Fv\u002Ffastai?color=blue&label=pypi%20version.png)](https:\u002F\u002Fpypi.org\u002Fproject\u002Ffastai\u002F#description)\n[![Conda (仅频道)](https:\u002F\u002Fimg.shields.io\u002Fconda\u002Fvn\u002Ffastai\u002Ffastai?color=seagreen&label=conda%20version.png)](https:\u002F\u002Fanaconda.org\u002Ffastai\u002Ffastai)\n\n## 安装\n\n您无需安装即可使用 fastai，只需通过 [Google Colab](https:\u002F\u002Fcolab.research.google.com\u002F) 即可。事实上，本文档的每一页都以交互式笔记本的形式提供——点击任意页面顶部的“在 Colab 中打开”即可打开（请务必将 Colab 运行时环境设置为“GPU”，以确保快速运行！）。有关更多信息，请参阅 fast.ai 文档中的 [使用 Colab](https:\u002F\u002Fcourse19.fast.ai\u002Fstart_colab.html) 部分。\n\n您也可以在自己的机器上通过 `pip install fastai` 来安装 fastai。\n\n为了确保您的机器上拥有最新版本的 PyTorch，建议先按照 [官方指南](https:\u002F\u002Fpytorch.org\u002Fget-started\u002Flocally\u002F) 进行安装。\n\n如果您计划自行开发 fastai，或希望始终处于技术前沿，可以采用可编辑安装方式（如果选择这种方式，还应同时对 [fastcore](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastcore) 进行可编辑安装）。首先安装 PyTorch，然后执行以下命令：\n\n    git clone https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\n    pip install -e \"fastai[dev]\"\n\n## 学习 fastai\n\n开始学习 fastai（以及深度学习）的最佳方式是阅读 [这本书](https:\u002F\u002Fwww.amazon.com\u002FDeep-Learning-Coders-fastai-PyTorch\u002Fdp\u002F1492045527)，并完成 [免费课程](https:\u002F\u002Fcourse.fast.ai)。\n\n要了解 fastai 的强大功能，请查看 [快速入门](https:\u002F\u002Fdocs.fast.ai\u002Fquick_start.html)，其中展示了如何仅用约 5 行代码构建图像分类器、图像分割模型、文本情感分析模型、推荐系统和表格数据模型。对于每种应用场景，代码都非常相似。\n\n浏览 [教程](https:\u002F\u002Fdocs.fast.ai\u002Ftutorial.html)，学习如何在自己的数据集上训练模型。使用导航侧边栏查阅 fastai 的完整文档。这里的每个类、函数和方法都有详细说明。\n\n如需了解该库的设计理念和动机，请阅读 [同行评审论文](https:\u002F\u002Fwww.mdpi.com\u002F2078-2489\u002F11\u002F2\u002F108\u002Fhtm)。\n\n## 关于 fastai\n\nfastai 是一个深度学习库，它为从业者提供了高层次组件，能够快速轻松地在标准深度学习领域取得最先进的成果；同时也为研究人员提供了低层次组件，可供自由组合以构建新的方法。其目标是在易用性、灵活性和性能之间不作重大妥协的情况下实现这两点。这得益于精心设计的分层架构，该架构将许多深度学习和数据处理技术中常见的底层模式抽象为解耦的模块。这些抽象可以通过利用 Python 语言的动态特性和 PyTorch 库的灵活性，以简洁明了的方式表达出来。fastai 包括：\n\n- 一种面向 Python 的新型类型调度系统，以及张量的语义类型层次结构\n- 一个可在纯 Python 中扩展的 GPU 优化计算机视觉库\n- 一种优化器，它将现代优化器的通用功能提炼为两个基本部分，使优化算法仅需 4–5 行代码即可实现\n- 一种新颖的双向回调系统，可在训练过程中的任何时刻访问和修改数据、模型或优化器的任何部分\n- 一个新的数据块 API\n- 以及更多……\n\nfastai 围绕两大设计目标组织：既易于上手且高效，又具有高度可扩展性和可配置性。它建立在一系列低层次 API 之上，这些 API 提供了可组合的构建模块。这样一来，用户若想重写高层 API 的一部分，或添加特定行为以满足自身需求，就不必深入学习最底层的细节。\n\n\u003Cimg alt=\"分层 API\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Ffastai_fastai_readme_9ca41d064328.png\" width=\"345\">\n\n## 从其他库迁移\n\n从原生 PyTorch、Ignite 或任何其他基于 PyTorch 的库迁移非常容易，甚至可以在使用 fastai 的同时与其他库结合使用。通常，您可以继续使用现有的数据处理代码，但用于训练的代码量会大幅减少，并且更容易采用现代最佳实践。以下是一些常用库的迁移指南，可帮助您顺利过渡：\n\n- [原生 PyTorch](https:\u002F\u002Fdocs.fast.ai\u002Fexamples\u002Fmigrating_pytorch.html)\n- [Ignite](https:\u002F\u002Fdocs.fast.ai\u002Fexamples\u002Fmigrating_ignite.html)\n- [Lightning](https:\u002F\u002Fdocs.fast.ai\u002Fexamples\u002Fmigrating_lightning.html)\n- [Catalyst](https:\u002F\u002Fdocs.fast.ai\u002Fexamples\u002Fmigrating_catalyst.html)\n\n## Windows 支持\n\n由于 Jupyter 和 Windows 上的 Python 多进程问题，`Dataloader` 的 `num_workers` 参数会自动重置为 0，以避免 Jupyter 崩溃。这使得在 Windows 上使用 Jupyter 进行计算机视觉等任务的速度比在 Linux 上慢得多。不过，如果直接从脚本中使用 fastai，则不存在这一限制。\n\n请参阅 [此示例](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fblob\u002Fmaster\u002Fnbs\u002Fexamples\u002Fdataloader_spawn.py)，以充分利用 fastai 在 Windows 上的功能。\n\n我们建议改用 Windows Subsystem for Linux (WSL)——这样您就可以采用常规的 Linux 安装方式，并且不会遇到 `num_workers` 相关的问题。\n\n## 测试\n\n要并行运行测试，请执行：\n\n`nbdev_test`\n\n为确保所有测试通过，您需要安装 `settings.ini` 中 `dev_requirements` 所列的依赖项：\n\n`pip install -e .[dev]`\n\n测试使用 `nbdev` 编写，例如 `test_eq` 的文档。\n\n## 贡献\n\n克隆本仓库后，请确保在终端中运行过 `nbdev_install_hooks`。此命令会安装 Jupyter 和 Git 钩子，以便自动清理、信任并解决笔记本中的合并冲突。\n\n在仓库中做出更改后，应运行 `nbdev_prepare`，并进行必要的调整，以通过所有测试。\n\n## Docker 容器\n\n对于有兴趣使用本项目官方 Docker 容器的用户，可以在此处找到：[fastai Docker 容器](https:\u002F\u002Fgithub.com\u002Ffastai\u002Fdocker-containers#fastai)。","# fastai 快速上手指南\n\nfastai 是一个基于 PyTorch 的深度学习库，旨在让从业者通过少量代码快速获得业界领先的结果，同时为研究者提供灵活的低层组件以构建新方法。\n\n## 环境准备\n\n*   **系统要求**：支持 Linux、macOS 和 Windows（Windows 用户建议在 Jupyter 中使用时配置 `num_workers=0`，或推荐使用 WSL 以获得最佳性能）。\n*   **前置依赖**：\n    *   Python 3.6+\n    *   **PyTorch**：建议优先安装适合你机器环境的最新 PyTorch 版本，以确保获得最佳性能。\n    *   硬件：推荐使用带有 GPU 的环境进行加速训练。\n\n> **提示**：无需本地安装即可体验，可直接使用 [Google Colab](https:\u002F\u002Fcolab.research.google.com\u002F)（记得将运行时更改为\"GPU\"）。\n\n## 安装步骤\n\n### 1. 安装 PyTorch\n请访问 [PyTorch 官网](https:\u002F\u002Fpytorch.org\u002Fget-started\u002Flocally\u002F) 获取适合你环境的安装命令并先行安装。\n\n*(注：国内用户若访问官网缓慢，可尝试使用清华源或阿里源安装 PyTorch，例如：`pip install torch torchvision torchaudio --index-url https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple`)*\n\n### 2. 安装 fastai\n使用 pip 进行标准安装：\n\n```bash\npip install fastai\n```\n\n### 3. 开发者安装（可选）\n如果你计划参与 fastai 开发或使用最新边缘功能，建议使用可编辑模式安装（需先安装 PyTorch）：\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\npip install -e \"fastai[dev]\"\n```\n*注意：开发者模式下，建议同时也对 [fastcore](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastcore) 进行可编辑安装。*\n\n## 基本使用\n\nfastai 的核心优势在于其高层 API 的一致性。仅需约 5 行代码，即可完成图像分类、分割、文本情感分析、推荐系统或表格模型的任务。\n\n以下是一个构建**图像分类器**的最简示例：\n\n```python\nfrom fastai.vision.all import *\n\n# 1. 准备数据 (使用内置数据集或指定路径)\npath = untar_data(URLs.PETS)\ndls = ImageDataLoaders.from_name_re(\n    path, get_image_files(path\u002F\"images\"), pat=r'(.+)_\\d+.jpg$', item_tfms=Resize(224)\n)\n\n# 2. 创建学习器 (使用预训练模型 resnet34)\nlearn = vision_learner(dls, resnet34, metrics=error_rate)\n\n# 3. 微调模型\nlearn.fine_tune(1)\n\n# 4. 预测\nimg = PILImage.create(path\u002F\"images\u002FBritish_Shorthair_12.jpg\")\npred, pred_idx, probs = learn.predict(img)\nprint(f\"Prediction: {pred}, Probability: {probs[pred_idx]:.4f}\")\n```\n\n**下一步学习建议：**\n*   阅读官方文档中的 [Quick Start](https:\u002F\u002Fdocs.fast.ai\u002Fquick_start.html) 查看更多应用场景。\n*   参考 [Tutorials](https:\u002F\u002Fdocs.fast.ai\u002Ftutorial.html) 学习如何在自定义数据集上训练模型。\n*   深入学习可参阅《Deep Learning for Coders with fastai and PyTorch》书籍或免费课程。","一家初创医疗影像公司急需在两周内构建一个高精度的肺部 CT 结节分类原型，以向投资人演示技术可行性。\n\n### 没有 fastai 时\n- 工程师需花费数天手动编写繁琐的数据加载、增强和归一化代码，且容易在处理 3D 医学影像维度时出错。\n- 调试模型训练过程极其痛苦，缺乏内置的监控工具，难以快速判断是学习率设置不当还是模型过拟合。\n- 复现业界最先进（SOTA）结果门槛极高，需要深入阅读大量论文并从头实现复杂的优化器和回调逻辑。\n- 尝试不同架构或微调策略时，代码耦合度高，每次修改都牵一发而动全身，迭代周期长达数小时甚至数天。\n\n### 使用 fastai 后\n- 利用 `DataBlock` API，仅用几行声明式代码即可完成复杂的医学影像数据管道搭建，自动处理增强与标准化。\n- 调用 `fit_one_cycle` 方法即可自动应用最佳实践的学习率调度策略，并实时展示训练指标，快速定位问题。\n- 直接加载预训练的 ResNet 等 SOTA 模型，通过简单的迁移学习流程，在少量数据上也能迅速达到高精度。\n- 借助灵活的回调系统，无需修改核心训练循环即可轻松添加自定义逻辑（如动态早停或特定保存策略），将实验迭代时间缩短至分钟级。\n\nfastai 通过高度封装的最佳实践与灵活的底层设计，让团队将原本需要数月的研发工作压缩至几天，成功按时交付了高质量原型。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Ffastai_fastai_7b8d4484.png","fast.ai","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Ffastai_50edd66d.png","",null,"fastdotai","https:\u002F\u002Ffast.ai","https:\u002F\u002Fgithub.com\u002Ffastai",[80,84,88],{"name":81,"color":82,"percentage":83},"Jupyter Notebook","#DA5B0B",98.5,{"name":85,"color":86,"percentage":87},"Python","#3572A5",1.5,{"name":89,"color":90,"percentage":91},"CSS","#663399",0,27958,7675,"2026-04-06T15:29:37","Apache-2.0","Linux, macOS, Windows","非必需（可在 CPU 运行），但推荐使用 GPU 加速；具体型号、显存大小及 CUDA 版本未在文中明确说明，需参考 PyTorch 官方安装指南","未说明",{"notes":100,"python":98,"dependencies":101},"建议在安装 fastai 前先单独安装最新版本的 PyTorch。在 Windows 系统的 Jupyter 环境中，数据加载器的 num_workers 会自动重置为 0 以避免卡死，这会导致计算机视觉等任务变慢；建议改用脚本运行或安装 Windows Subsystem for Linux (WSL) 以获得完整性能。开发模式下需同时以可编辑模式安装 fastcore。官方提供 Docker 容器镜像。",[102,103],"torch","fastcore",[14],[106,107,108,109,110,64,111,112],"deep-learning","machine-learning","pytorch","python","gpu","notebooks","colab","2026-03-27T02:49:30.150509","2026-04-07T08:13:25.954069",[116,121,126,131,136,140],{"id":117,"question_zh":118,"answer_zh":119,"source_url":120},21356,"在 Linux 系统上加载在 Windows 上训练的 fastai 模型时，出现 'cannot instantiate WindowsPath' 错误怎么办？","这是因为模型序列化时保存了 Windows 路径对象，而在 Linux 上无法实例化。解决方法是在加载模型前添加以下代码进行路径类映射：\n\nimport pathlib\nimport platform\nplt = platform.system()\nif plt == 'Linux':\n    pathlib.WindowsPath = pathlib.PosixPath\n\n然后即可正常加载模型。如果此方法无效，建议重新在 Linux 环境（如 Google Colab 或 Ubuntu）中训练模型以避免跨平台路径问题。","https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F1482",{"id":122,"question_zh":123,"answer_zh":124,"source_url":125},21357,"使用 ImageDataLoaders 且 num_workers > 0 时，训练报错 'Cannot pickle CUDA storage' 如何解决？","该错误通常发生在多进程数据加载尝试序列化 CUDA 张量时。解决方案是将 num_workers 设置为 0，即禁用多进程数据加载：\n\ndata = ImageDataLoaders.from_df(..., num_workers=0)\n\n虽然这会略微降低数据加载速度，但能避免 CUDA 存储无法序列化的问题。若必须使用多进程，需确保数据预处理中不包含已移动到 GPU 的张量。","https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F2899",{"id":127,"question_zh":128,"answer_zh":129,"source_url":130},21358,"运行 predict 或 TTA 时遇到 'RuntimeError: received 0 items of ancdata' 错误如何处理？","该错误与 PyTorch 的多进程数据加载有关。维护者已将数据加载器从 ProcessPool 改为 ThreadPool 以解决此问题。请确保你使用的是最新版本的 fastai 库。如果问题仍存在，可尝试手动设置 num_workers=0 来绕过该问题：\n\ndata = ImageClassifierData.from_paths(..., num_workers=0)\n\n更新库后通常无需额外配置即可解决。","https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F23",{"id":132,"question_zh":133,"answer_zh":134,"source_url":135},21359,"调用 model.fit 时系统直接死机并重启，可能是什么原因？","这通常是由于 CUDA\u002FcuDNN 版本冲突或显存溢出导致的系统级崩溃。检查点包括：\n1. 确认安装的 cuDNN 版本与 CUDA 版本兼容（例如 CUDA 9.0 应搭配对应的 cuDNN 7.x）；\n2. 避免通过 conda 和手动安装混合不同版本的 cuDNN；\n3. 尝试减小 batch size 或使用 precompute=True 减少显存占用；\n4. 在 Windows 上，某些驱动版本可能导致内核级崩溃，建议更新显卡驱动至最新稳定版。\n\n若日志无详细错误信息，可尝试在命令行运行 Python 脚本而非 Jupyter Notebook 以获取更清晰的报错。","https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F751",{"id":137,"question_zh":138,"answer_zh":139,"source_url":130},21360,"如何在 fastai 中正确设置数据加载的 worker 数量以避免多进程错误？","在创建 DataBunch 或 ImageDataLoaders 时，通过 num_workers 参数控制工作进程数。对于大多数用户，尤其是 Windows 用户或遇到序列化错误的用户，推荐设置为 0：\n\ndata = ImageDataLoaders.from_df(..., num_workers=0)\n\n在 Linux 系统且数据管道简单时可尝试设为 2 或 4 以提升性能，但若出现 'ancdata' 或 'pickle CUDA' 相关错误，应立即改回 0。",{"id":141,"question_zh":142,"answer_zh":143,"source_url":120},21361,"跨平台部署 fastai 模型（Windows 训练，Linux 部署）时的最佳实践是什么？","为避免路径序列化问题（如 WindowsPath 错误），最佳实践是：\n1. 在与部署环境相同的操作系统上进行模型训练；\n2. 若必须在 Windows 训练，部署前在 Linux 环境中重新加载模型并再次 export，以生成兼容的路径对象；\n3. 或在部署代码开头加入路径兼容性补丁：\n\nimport pathlib, platform\nif platform.system() == 'Linux':\n    pathlib.WindowsPath = pathlib.PosixPath\n\n但最可靠的方式仍是统一训练与部署环境，推荐使用 Docker 或云笔记本（如 Colab）保持一致性。",[145,150,155,160,165,170,175,180,185,190,195,200,205,210,215,220,225,230,235,240],{"id":146,"version":147,"summary_zh":148,"released_at":149},127359,"2.8.7","- 允许任何 pytorch\u003C3","2026-02-14T01:58:56",{"id":151,"version":152,"summary_zh":153,"released_at":154},127360,"2.8.6","- 新的 fastcore 依赖","2025-12-15T18:05:24",{"id":156,"version":157,"summary_zh":158,"released_at":159},127361,"2.8.5","### 新特性\n\n- 支持 PyTorch 2.9 ([#4116](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F4116))\n\n### 已修复的 bug\n\n- 解决 fp16 回调中的弃用警告 ([#4124](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4124))，感谢 [@FacuRoffet99](https:\u002F\u002Fgithub.com\u002FFacuRoffet99)\n- 使 SaveModelCallback 的修复与 TextLearner 兼容 ([#4121](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4121))，感谢 [@austinvhuang](https:\u002F\u002Fgithub.com\u002Faustinvhuang)","2025-10-26T23:05:26",{"id":161,"version":162,"summary_zh":163,"released_at":164},127362,"2.8.4","### 修复的 bug\n\n- 为 `load_model_text` 设置 `weights_only=False`，修复了 LRFinder \u002F `lr_find` 问题 ([#4120](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4120))，感谢 [@austinvhuang](https:\u002F\u002Fgithub.com\u002Faustinvhuang)\n- 修复：在 TextLearner 构造函数中将 alpha 和 beta 参数传递给 `rnn_cbs()` ([#4119](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4119))，感谢 [@austinvhuang](https:\u002F\u002Fgithub.com\u002Faustinvhuang)\n- 修复了 SaveModelCallback ([#4118](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4118))，感谢 [@FacuRoffet99](https:\u002F\u002Fgithub.com\u002FFacuRoffet99)","2025-08-19T22:00:39",{"id":166,"version":167,"summary_zh":168,"released_at":169},127363,"2.8.3","### 新功能\n\n- 支持 PyTorch 2.8 ([#4116](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F4116))\n\n### 已修复的 bug\n\n- 更新学习率查找图 ([#4098](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4098))，感谢 [@Timmecom](https:\u002F\u002Fgithub.com\u002FTimmecom)","2025-08-07T02:45:58",{"id":171,"version":172,"summary_zh":173,"released_at":174},127364,"2.8.2","### 新功能\n\n- 将 PyTorch 版本更新至 2.7（[#4095](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4095)），感谢 [@tonyhoo](https:\u002F\u002Fgithub.com\u002Ftonyhoo)","2025-05-24T03:52:50",{"id":176,"version":177,"summary_zh":178,"released_at":179},127365,"2.8.1","### 新特性\n\n- 使用 fasttransform ([#4074](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4074))，感谢 [@RensDimmendaal](https:\u002F\u002Fgithub.com\u002FRensDimmendaal)","2025-04-18T21:14:58",{"id":181,"version":182,"summary_zh":183,"released_at":184},127366,"2.7.19","### 新特性\n\n- 添加对 PyTorch 2.6 的支持（[#4078](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4078)），感谢 [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)","2025-03-12T19:09:15",{"id":186,"version":187,"summary_zh":188,"released_at":189},127367,"2.7.18","### 新特性\n\n- 支持 PyTorch 2.5","2024-10-19T03:20:13",{"id":191,"version":192,"summary_zh":193,"released_at":194},127368,"2.7.17","### 新特性\n\n- 在文档输出中添加 Markdown 支持 ([#4044](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F4044))\n- 支持 PyTorch 2.4 ([#4040](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4040))，感谢 [@tonyhoo](https:\u002F\u002Fgithub.com\u002Ftonyhoo)\n- 更新 Fastcore 的最大版本号","2024-08-27T06:50:47",{"id":196,"version":197,"summary_zh":198,"released_at":199},127369,"2.7.16","### New Features\n\n- Support PyTorch 2.4 ([#4040](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4040)), thanks to [@tonyhoo](https:\u002F\u002Fgithub.com\u002Ftonyhoo)\n- Support for loss function pickling ([#4034](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4034)), thanks to [@kevin-vitro](https:\u002F\u002Fgithub.com\u002Fkevin-vitro)","2024-07-30T22:26:23",{"id":201,"version":202,"summary_zh":203,"released_at":204},127370,"2.7.15","### New Features\n\n- Support PyTorch 2.3 ([#4026](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F4026)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Add `log` and `show_epochs` to `log_ploss` ([#3964](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3964)), thanks to [@turbotimon](https:\u002F\u002Fgithub.com\u002Fturbotimon)","2024-04-27T22:14:04",{"id":206,"version":207,"summary_zh":208,"released_at":209},127371,"2.7.14","### New Features\n\n- PyTorch 2.2 support, thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)","2024-02-01T00:26:04",{"id":211,"version":212,"summary_zh":213,"released_at":214},127372,"2.7.13","### New Features\n\n- PyTorch 2.1 compatibility ([#3970](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3970)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Add `MutableMapping` to `torch_core.apply` to Support Moving Transformers Dicts ([#3969](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3969)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Added Jaccard coefficient metric for multiclass target in segmentation ([#3951](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3951)), thanks to [@Hazem-Ahmed-Abdelraouf](https:\u002F\u002Fgithub.com\u002FHazem-Ahmed-Abdelraouf)\n- Support TorchVision's Multi-Weight API ([#3944](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3944)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Fix the Deploy to GitHub Pages Action ([#3942](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3942)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n\n### Bugs Squashed\n\n- Fix Pandas Categorical FutureWarning ([#3973](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3973)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Fix torch.jit.script on TimmBody ([#3948](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3948)), thanks to [@johan12345](https:\u002F\u002Fgithub.com\u002Fjohan12345)\n- Resolve CutMix Deprecation Warning ([#3937](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3937)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Fixed format string ([#3934](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3934)), thanks to [@bkowshik](https:\u002F\u002Fgithub.com\u002Fbkowshik)\n- Fix casting types for mps ([#3912](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3912)), thanks to [@MSciesiek](https:\u002F\u002Fgithub.com\u002FMSciesiek)\n- Fix AccumMetric name.setter ([#3621](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3621)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)","2023-10-15T04:03:45",{"id":216,"version":217,"summary_zh":218,"released_at":219},127373,"2.7.12","### New Features\n\n- PyTorch 2.0 compatibility ([#3890](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3890)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Pytorch 2.0 compiler compatibility ([#3899](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3899)), thanks to [@ggosline](https:\u002F\u002Fgithub.com\u002Fggosline)\n- Better version support for `TensorBase.new_empty` ([#3887](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3887)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- TensorBase deepcopy Compatibility ([#3882](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3882)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- ChannelsLast Callback Improvements & Bug Fix ([#3876](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3876)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Add support for a batch transforms `to` method ([#3875](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3875)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Allow Pillow Image to be passed to PILBase.create ([#3872](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3872)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n\n### Bugs Squashed\n\n- Fix `Learn.predict` Errors Out if Passed a PILImage ([#3884](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3884)), thanks to [@nglillywhite](https:\u002F\u002Fgithub.com\u002Fnglillywhite)\n- Set DataLoaders device if not None and to exists ([#3873](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3873)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Fix `default_device` to correctly detect + use mps (Apple Silicon) ([#3858](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3858)), thanks to [@wolever](https:\u002F\u002Fgithub.com\u002Fwolever)","2023-03-28T20:28:42",{"id":221,"version":222,"summary_zh":223,"released_at":224},127374,"2.7.11","### New Features\n\n- ChannelsLast Callback Improvements, Additional Documentation, & Bug Fix ([#3876](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3876)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Add support for a batch transforms `to` method ([#3875](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3875)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Allow Pillow Image to be passed to PILBase.create ([#3872](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3872)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Compat with latest numpy ([#3871](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3871)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Move training-only step to separate function in `Learner` ([#3857](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3857)), thanks to [@kunaltyagi](https:\u002F\u002Fgithub.com\u002Fkunaltyagi)\n- TabularPandas data transform reproducibility ([#2826](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F2826))\n\n### Bugs Squashed\n\n- Set DataLoaders device if not None and to exists ([#3873](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3873)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Fix `default_device` to correctly detect + use mps (Apple Silicon) ([#3858](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3858)), thanks to [@wolever](https:\u002F\u002Fgithub.com\u002Fwolever)\n- Fix load hanging in distributed processes ([#3839](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3839)), thanks to [@muellerzr](https:\u002F\u002Fgithub.com\u002Fmuellerzr)\n- `default_device` logic is repeated twice, related to `mps` \u002F OSX support. ([#3785](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3785))\n- revert auto-enable of mac mps due to pytorch limitations ([#3769](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3769))\n- Fix Classification Interpretation ([#3563](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3563)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- vision tutorial failed at `learner.fine_tune(1)` ([#3283](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3283))","2023-02-15T02:47:21",{"id":226,"version":227,"summary_zh":228,"released_at":229},127375,"2.7.10","### New Features\n\n- Add torch save and load kwargs ([#3831](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3831)), thanks to [@JonathanGrant](https:\u002F\u002Fgithub.com\u002FJonathanGrant)\n  - This lets us do nice things like set pickle_module to cloudpickle\n- PyTorch 1.13 Compatibility ([#3828](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3828)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n- Recursive copying of attribute dictionaries for TensorImage subclass ([#3822](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3822)), thanks to [@restlessronin](https:\u002F\u002Fgithub.com\u002Frestlessronin)\n- `OptimWrapper` sets same param groups as `Optimizer` ([#3821](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3821)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)\n  - This PR harmonizes the default parameter group setting between `OptimWrapper` and `Optimizer` by modifying `OptimWrapper` to match `Optimizer`'s logic.\n- Support normalization of 1-channel images in unet ([#3820](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3820)), thanks to [@marib00](https:\u002F\u002Fgithub.com\u002Fmarib00)\n- Add `img_cls` param to `ImageDataLoaders` ([#3808](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3808)), thanks to [@tcapelle](https:\u002F\u002Fgithub.com\u002Ftcapelle)\n  - This is particularly useful for passing `PILImageBW` for MNIST.\n- Add support for `kwargs` to `tensor()` when arg is an `ndarray` ([#3797](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3797)), thanks to [@SaadAhmedGit](https:\u002F\u002Fgithub.com\u002FSaadAhmedGit)\n- Add latest TorchVision models on fastai ([#3791](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3791)), thanks to [@datumbox](https:\u002F\u002Fgithub.com\u002Fdatumbox)\n- Option to preserve filenames in `download_images` ([#2983](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F2983)), thanks to [@mess-lelouch](https:\u002F\u002Fgithub.com\u002Fmess-lelouch)\n\n### Bugs Squashed\n\n- `get_text_classifier` fails with custom `AWS_LSTM` ([#3817](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3817))\n- revert auto-enable of mac mps due to pytorch limitations ([#3769](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3769))\n- Workaround for performance bug in PyTorch with subclassed tensors ([#3683](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3683)), thanks to [@warner-benjamin](https:\u002F\u002Fgithub.com\u002Fwarner-benjamin)","2022-11-02T03:04:39",{"id":231,"version":232,"summary_zh":233,"released_at":234},127376,"2.7.8","### New Features\n\n- add split value argument to ColSplitter ([#3737](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3737)), thanks to [@DanteOz](https:\u002F\u002Fgithub.com\u002FDanteOz)\n- deterministic repr for PIL images ([#3762](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3762))\n- option to skip default callbacks in `Learner` ([#3739](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3739))\n- update for nbdev2 ([#3747](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3747))\n\n### Bugs Squashed\n\n- IntToFloatTensor failing on Mac mps due to missing op ([#3761](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3761))\n- fix for pretrained in vision.learner ([#3746](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3746)), thanks to [@peterdudfield](https:\u002F\u002Fgithub.com\u002Fpeterdudfield)\n- fix same file error message when resizing image ([#3743](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fpull\u002F3743)), thanks to [@cvergnes](https:\u002F\u002Fgithub.com\u002Fcvergnes)","2022-08-02T19:19:53",{"id":236,"version":237,"summary_zh":238,"released_at":239},127377,"2.7.6","### New Features\n\n- Initial Mac GPU (mps) support ([#3719](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3719))","2022-07-07T18:47:13",{"id":241,"version":242,"summary_zh":243,"released_at":244},127378,"2.7.5","### New Features\n\n- auto-normalize timm models ([#3716](https:\u002F\u002Fgithub.com\u002Ffastai\u002Ffastai\u002Fissues\u002F3716))\n- PyTorch 1.12 support","2022-07-04T04:02:49"]