<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Transformer on 南巷</title><link>https://nx.wmlab.top/tags/transformer/</link><description>Recent content in Transformer on 南巷</description><generator>Hugo -- gohugo.io</generator><language>zh-CN</language><managingEditor>liuwenhao1968@163.com (南巷)</managingEditor><webMaster>liuwenhao1968@163.com (南巷)</webMaster><copyright>© 2026 南巷</copyright><lastBuildDate>Mon, 15 Jun 2026 02:14:35 +0800</lastBuildDate><atom:link href="https://nx.wmlab.top/tags/transformer/index.xml" rel="self" type="application/rss+xml"/><item><title>CodeLab-LLaMA2 / 星语 MoE：大模型训练实践</title><link>https://nx.wmlab.top/portfolio/xingyu-moe/</link><pubDate>Wed, 15 Apr 2026 00:00:00 +0000</pubDate><author>liuwenhao1968@163.com (南巷)</author><guid>https://nx.wmlab.top/portfolio/xingyu-moe/</guid><description>围绕 LLaMA2 内部原理与工程实现，系统拆解架构、组件、预训练、SFT、LoRA、推理与 RAG/Agent，形成从理论到应用的大模型学习与实现闭环。</description></item><item><title>SegFormer</title><link>https://nx.wmlab.top/blog/segformer/</link><pubDate>Mon, 15 Jun 2026 02:14:35 +0800</pubDate><author>liuwenhao1968@163.com (南巷)</author><guid>https://nx.wmlab.top/blog/segformer/</guid><description>提出一个层级（多尺度）且不依赖位置编码的 Transformer 编码器（MiT），可以直接输出多尺度特征（1/4、1/8、1/16、1/32），避免了 ViT 在不同测试分辨率下必须插值位置编码带来的性能下降。</description></item><item><title>MaskFormer</title><link>https://nx.wmlab.top/blog/maskformer/</link><pubDate>Mon, 15 Jun 2026 02:14:28 +0800</pubDate><author>liuwenhao1968@163.com (南巷)</author><guid>https://nx.wmlab.top/blog/maskformer/</guid><description>具体表现：</description></item><item><title>Mask2Former</title><link>https://nx.wmlab.top/blog/mask2former/</link><pubDate>Mon, 15 Jun 2026 02:14:20 +0800</pubDate><author>liuwenhao1968@163.com (南巷)</author><guid>https://nx.wmlab.top/blog/mask2former/</guid><description>MaskFormer 让分割任务变成 “预测一组掩码 + 类别”； Mask2Former 让模型 “只看自己负责的掩码区域，更聪明地预测掩码”。</description></item><item><title>DETR</title><link>https://nx.wmlab.top/blog/detr/</link><pubDate>Mon, 15 Jun 2026 02:14:04 +0800</pubDate><author>liuwenhao1968@163.com (南巷)</author><guid>https://nx.wmlab.top/blog/detr/</guid><description>在做DETR之前，目标检测都是下面两种任务方式：</description></item><item><title>Deformable DETR</title><link>https://nx.wmlab.top/blog/deformable-detr/</link><pubDate>Mon, 15 Jun 2026 02:14:02 +0800</pubDate><author>liuwenhao1968@163.com (南巷)</author><guid>https://nx.wmlab.top/blog/deformable-detr/</guid><description>原来大DETR有几个缺点：</description></item><item><title>LLaMA 2 动手实现</title><link>https://nx.wmlab.top/blog/llama2-hands-on-implementation/</link><pubDate>Sun, 14 Jun 2026 17:52:34 +0800</pubDate><author>liuwenhao1968@163.com (南巷)</author><guid>https://nx.wmlab.top/blog/llama2-hands-on-implementation/</guid><description>围绕 LLaMA 2 的数据集、Tokenizer、模型结构与训练代码，记录从原理拆解到工程实现的完整实践过程。</description></item><item><title>LLaMA 2 解析</title><link>https://nx.wmlab.top/blog/llama2-analysis/</link><pubDate>Sun, 14 Jun 2026 17:49:58 +0800</pubDate><author>liuwenhao1968@163.com (南巷)</author><guid>https://nx.wmlab.top/blog/llama2-analysis/</guid><description>从模型架构、Tokenizer、RMSNorm、GQA、RoPE、FFN 到输出层，系统拆解 LLaMA 2 的核心组件与设计动机。</description></item></channel></rss>