Chaoke Du
Graduate Student
Exploring the intersection of AI and science.
Featured Projects
Tools and applications I've built
AI Image Studio
A browser-based image generation and editing workspace for prompt-driven visual creation.
Music Visualizer
An interactive audio visualization experience for exploring sound through motion and color.
SchemaForge
An AI-assisted database schema design tool that turns product scenarios into structured database models.
Solution Architect
A visual architecture design tool for drafting systems, flows, and implementation plans.
PaperReader - AI Paper Reading Assistant
An AI-powered tool that reads academic papers with you, paragraph by paragraph.
This Website
My personal portfolio built with Next.js, featuring i18n, Markdown blog, and more.
Latest Posts
Thoughts on tech and research
【Vision Transformer 实战】从 CNN 到 ViT:图像分类 SOTA 之路与完整实现(2026 最新版)
2020 年 Google 提出的 Vision Transformer(ViT)彻底改变了计算机视觉领域。它证明了 Transformer 架构在视觉任务上可以超越传统的 CNN,在 ImageNet 等基准测试上达到 SOTA。本文将从原理到代码,深入讲解 ViT 的架构,并对比 ResNet、EfficientNet、Swin Transformer 等模型。
【深度学习进阶】Transformer 架构深度解析:从零手写 Self-Attention 与完整实现(2026 最新版)
Transformer 是现代深度学习的基石,从 GPT-4、LLaMA 3 到 Vision Transformer,几乎所有 SOTA 模型都基于它。本文将从数学原理出发,**从零手写** Self-Attention、Multi-Head Attention、位置编码,并实现完整的 Transformer,最后对比 2026 年最新的 Flash Attention、RoPE、GQA 等优化技术。适合有一定 PyTorch 基础、想深入理解 Transformer 的读者。
02 | 大模型的"大脑":Transformer与预训练
**核心问题**:预训练到底在学什么?为什么下一个token预测能产生如此强大的能力?