Merge remote-tracking branch 'main/main'
This commit is contained in:
@@ -0,0 +1,150 @@
|
||||
---
|
||||
title: fast.ai 课程 — Practical Deep Learning for Coders
|
||||
created: 2026-04-10
|
||||
updated: 2026-04-10
|
||||
type: concept
|
||||
tags: [deep-learning, course, fastai, pytorch, stable-diffusion]
|
||||
sources: [https://course.fast.ai/]
|
||||
---
|
||||
|
||||
# fast.ai 课程 — Practical Deep Learning for Coders
|
||||
|
||||
## 概述
|
||||
|
||||
| 项目 | 内容 |
|
||||
|------|------|
|
||||
| **课程名称** | Practical Deep Learning for Coders |
|
||||
| **发布方** | fast.ai(Jeremy Howard & Rachel Thomas)|
|
||||
| **授课地点** | University of Queensland(Part 1, 2022)|
|
||||
| **课程性质** | 免费,无需大学数学基础 |
|
||||
| **前置要求** | 有一定编程经验 |
|
||||
| **软件栈** | PyTorch、fastai、Hugging Face |
|
||||
| **视频播放量** | 600 万+ 次 |
|
||||
| **配套书籍** | [*Deep Learning for Coders with fastai and PyTorch*](https://www.amazon.com/dp/1492045527)(Amazon 5星,免费在线阅读)|
|
||||
|
||||
fast.ai 课程的核心理念:**"Deep Learning is for everyone"**。课程设计强调实践优先、理论从实践出,不要求事先掌握线性代数或微积分——所需数学会在课程中按需教授。
|
||||
|
||||
---
|
||||
|
||||
## Part 1:Practical Deep Learning(基础篇)
|
||||
|
||||
**9 节课时,每节约 90 分钟**
|
||||
|
||||
| 课时 | 主题 | 核心内容 |
|
||||
|------|------|---------|
|
||||
| L1 | [Getting started](https://course.fast.ai/Lessons/lesson1.html) | 5 分钟跑通第一个模型;计算机视觉 end-to-end 训练流程;图像分类实战(is-it-a-bird)|
|
||||
| L2 | [Deployment](https://course.fast.ai/Lessons/lesson2.html) | 模型部署;Gradio / Hugging Face Spaces;将模型发布为 Web 服务 |
|
||||
| L3 | [Neural net foundations](https://course.fast.ai/Lessons/lesson3.html) | 反向传播直观理解;激活函数;损失函数;fastai 的训练循环 |
|
||||
| L4 | [Natural Language (NLP)](https://course.fast.ai/Lessons/lesson4.html) | 文本分类;语言模型微调;Hugging Face Transformers |
|
||||
| L5 | [From-scratch model](https://course.fast.ai/Lessons/lesson5.html) | 从零实现神经网络;Python 底层代码;理解每一行计算 |
|
||||
| L6 | [Random forests](https://course.fast.ai/Lessons/lesson6.html) | 随机森林;梯度提升;表格数据分析;scikit-learn |
|
||||
| L7 | [Collaborative filtering](https://course.fast.ai/Lessons/lesson7.html) | 推荐系统;嵌入表示;协同过滤实战 |
|
||||
| L8 | [Convolutions (CNNs)](https://course.fast.ai/Lessons/lesson8.html) | 卷积神经网络原理;ResNet 架构;计算机视觉深度模型 |
|
||||
| Bonus | [Data Ethics](https://course.fast.ai/Lessons/lesson8a.html) | 数据伦理;算法偏见;AI 负责任部署 |
|
||||
|
||||
**Part 1 能解决的问题类型:**
|
||||
- 计算机视觉(图像分类、目标检测、分割)
|
||||
- 自然语言处理(文本分类、情感分析)
|
||||
- 表格数据分析(结构化数据建模)
|
||||
- 协同过滤(推荐系统)
|
||||
- 随机森林与回归模型
|
||||
- 模型部署(Web 服务)
|
||||
|
||||
---
|
||||
|
||||
## Part 2:Deep Learning Foundations to Stable Diffusion(进阶篇)
|
||||
|
||||
**17+ 节课时,30+ 小时视频**
|
||||
|
||||
**前置要求:** 完成 Part 1,或具备以下同等能力:
|
||||
- 能从零手写 SGD 训练循环
|
||||
- 熟悉 Kaggle 竞赛级深度学习实战
|
||||
- 熟练使用 PyTorch 和 fastai 进行 NLP/CV 建模
|
||||
|
||||
**Part 2 核心理念:** 从零实现 Stable Diffusion——那个让互联网震惊的 AI 生图 killer app。课程与 Stability.ai 和 Hugging Face Diffusers 团队深度合作,确保覆盖最新技术。
|
||||
|
||||
| 课时 | 主题 | 核心内容 |
|
||||
|------|------|---------|
|
||||
| L9 | [Stable Diffusion](https://course.fast.ai/Lessons/lesson9.html) | Diffusers 库使用;Pipeline API;文生图;扩散模型直观理解 |
|
||||
| L10 | [Diving Deeper](https://course.fast.ai/Lessons/lesson10.html) | 深入 diffusion 机制;噪声调度;条件控制 |
|
||||
| L11 | [Matrix multiplication](https://course.fast.ai/Lessons/lesson11.html) | 矩阵乘法底层;CUDA 优化;GPU 计算原理 |
|
||||
| L12 | [Mean shift clustering](https://course.fast.ai/Lessons/lesson12.html) | 聚类;均值漂移;embedding 空间分析 |
|
||||
| L13 | [Backpropagation & MLP](https://course.fast.ai/Lessons/lesson13.html) | 反向传播推导;多层感知机实现 |
|
||||
| L14 | [Backpropagation](https://course.fast.ai/Lessons/lesson14.html) | 自动求导机制;PyTorch autograd 深入 |
|
||||
| L15 | [Autoencoders](https://course.fast.ai/Lessons/lesson15.html) | 自编码器;变分自编码器(VAE);latent 空间 |
|
||||
| L16 | [The Learner framework](https://course.fast.ai/Lessons/lesson16.html) | fastai Learner 框架设计;训练循环抽象 |
|
||||
| L17 | [Initialization/normalization](https://course.fast.ai/Lessons/lesson17.html) | 权重初始化;BatchNorm/LayerNorm;训练稳定性 |
|
||||
| L18 | [Accelerated SGD & ResNets](https://course.fast.ai/Lessons/lesson18.html) | 动量 SGD;ResNet 跳跃连接;优化器深入 |
|
||||
| L19 | [DDPM and Dropout](https://course.fast.ai/Lessons/lesson19.html) | DDPM(Denoising Diffusion Probabilistic Models);Dropout 理论 |
|
||||
| L20 | [Mixed Precision](https://course.fast.ai/Lessons/lesson20.html) | FP16/BF16 混合精度训练;梯度缩放;节省显存 |
|
||||
| L21 | [DDIM](https://course.fast.ai/Lessons/lesson21.html) | DDIM(Denoising Diffusion Implicit Models);加速采样 |
|
||||
| L22 | [Karras et al (2022)](https://course.fast.ai/Lessons/lesson22.html) | 论文精读;EDM(Elucidating...);改进 diffusion 采样 |
|
||||
| L23 | [Super-resolution](https://course.fast.ai/Lessons/lesson23.html) | 图像超分辨率;diffusion 在图像增强中的应用 |
|
||||
| L24 | [Attention & transformers](https://course.fast.ai/Lessons/lesson24.html) | Attention 机制;Transformer 架构;GPT 原理 |
|
||||
| L25 | [Latent diffusion](https://course.fast.ai/Lessons/lesson25.html) | Latent Diffusion Model(Stable Diffusion 核心);U-Net 在 latent 空间的计算 |
|
||||
|
||||
**Bonus 视频:**
|
||||
- [Lesson 9a](https://youtu.be/0_BBRNYInx8)(Stable Diffusion 进阶技巧)
|
||||
- [Lesson 9b](https://youtu.be/mYpjmM7O-30)(Diffusers 细节)
|
||||
|
||||
**Part 2 覆盖的关键技术栈:**
|
||||
- Stable Diffusion / Latent Diffusion Model
|
||||
- U-Net、CLIP embedding、VAE
|
||||
- DDPM、DDIM、EDM(Karras 2022)
|
||||
- Transformer / Attention 机制
|
||||
- 混合精度训练(FP16/BF16)
|
||||
- 论文阅读方法论(如何读懂顶会论文并复现)
|
||||
|
||||
---
|
||||
|
||||
## 配套资源
|
||||
|
||||
| 资源 | 链接 | 说明 |
|
||||
|------|------|------|
|
||||
| 官方书籍(免费在线)| [fast.ai/book](https://course.fast.ai/Resources/book.html) | 与 Part 1 完全对应 |
|
||||
| 课程论坛 | [forums.fast.ai](https://forums.fast.ai) | 问答、组队学习 |
|
||||
| Kaggle Notebooks | [course.fast.ai/Resources/kaggle](https://course.fast.ai/Resources/kaggle.html) | 所有课时 Notebook,可直接运行 |
|
||||
| Paperspace Gradient | [$10 免费额度](https://gradient.run/notebooks) | GPU 云端运行(推荐,国内可访问)|
|
||||
| 测试imonials | [course.fast.ai/Resources/testimonials](https://course.fast.ai/Resources/testimonials.html) | 学员反馈 |
|
||||
|
||||
---
|
||||
|
||||
## fastai 库简介
|
||||
|
||||
**fastai** 是课程核心框架,建立在 PyTorch 之上,提供:
|
||||
|
||||
- **高层 API**:`vision_learner`、`text_classifier_learner`、`tabular_learner` 等,一行代码构建训练流程
|
||||
- **DataLoaders**:统一的 DataLoader 封装,自动处理 transforms、splits
|
||||
- **Learner**:训练循环抽象,内置 metrics、callbacks、fp16
|
||||
- **导出部署**:`learn.export()` 一键导出,可配合 Gradio/HuggingFace 部署
|
||||
|
||||
fastai 的设计哲学:**让研究员少写代码,让代码研究员能理解底层**。
|
||||
|
||||
---
|
||||
|
||||
## 学习路径建议
|
||||
|
||||
```
|
||||
入门路径:
|
||||
1. Part 1 L1(Getting Started)→ 立刻跑出第一个模型,建立信心
|
||||
2. Part 1 L2(Deployment)→ 学会把模型上线,立即有获得感
|
||||
3. Part 1 L3-L8 → 系统掌握 CV/NLP/表格/推荐四大任务
|
||||
4. Part 1 Bonus(Data Ethics)→ 理解 AI 伦理风险
|
||||
|
||||
进阶路径(Part 2):
|
||||
5. Part 2 L9(Stable Diffusion)→ 生图实战,建立直觉
|
||||
6. Part 2 L13-L14(Backpropagation)→ 彻底理解训练过程
|
||||
7. Part 2 L16(Learner framework)→ 掌握 fastai 底层设计
|
||||
8. Part 2 L24(Attention & transformers)→ 大模型基础
|
||||
9. Part 2 L25(Latent diffusion)→ Stable Diffusion 核心原理
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 相关页面
|
||||
|
||||
> [[ml-intro]](机器学习入门概念)
|
||||
> [[pytorch-basics]](PyTorch 基础)
|
||||
> [[transformer-architecture]](Transformer 架构)
|
||||
> [[diffusion-models]](Diffusion 模型)
|
||||
> [[huggingface-quickstart]](Hugging Face 快速上手)
|
||||
@@ -9,7 +9,7 @@ tags: [meta, index]
|
||||
# Wiki Index
|
||||
|
||||
> 機場智能化工程 wiki — 涵蓋**智算中心技術**與**航班運營管理**兩大領域
|
||||
> Last updated: 2026-04-10 | Total pages: 44
|
||||
> Last updated: 2026-04-10 | Total pages: 35
|
||||
|
||||
---
|
||||
|
||||
|
||||
@@ -198,6 +198,15 @@ tags: [meta, log]
|
||||
- 總計收錄約 200+ 縮寫,完整覆蓋智算中心 + 航班運營兩大領域
|
||||
- 更新:glossary.md、index.md、log.md
|
||||
|
||||
## [2026-04-10] ingest | fast.ai Practical Deep Learning 課程文檔
|
||||
|
||||
- 來源:https://course.fast.ai/ 首頁 + L1/L9/part2 overview 頁面
|
||||
- 新建:concepts/flight-operations/fast-ai-course.md
|
||||
- Part 1(9 課時):Getting Started / Deployment / Neural Net Foundations / NLP / From-scratch / Random Forests / Collaborative Filtering / CNNs / Data Ethics
|
||||
- Part 2(17 課時):Stable Diffusion 從零實現 / Backpropagation / Autoencoders / Attention & Transformers / Latent Diffusion 等
|
||||
- **已移出**:文件移至 vault 根目錄 03_Resources/Development/fast-ai-course.md(不在 airport-wiki 內)
|
||||
- 更新:index.md(Total pages: 36→35)、log.md
|
||||
|
||||
## [2026-04-10] update | 重写 机场智算中心技术方案
|
||||
- 基于 wiki 最新内容全面重写:GPU 集群拓扑、液冷 PUE/WUE/CUE 三级指标、IB NDR 400G 组网、存储架构、国产化方案
|
||||
- 修正:GB200 NVL72 HBM3e 带宽 16 TB/s(原错误值已更正)、HGX H100 NVSwitch 6 芯片拓扑、昇腾 HCCS 392 GB/s(原误写 128 GB/s)
|
||||
|
||||
Reference in New Issue
Block a user