<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[zzZ_CMing的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/zzZ_CMing</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; zzZ_CMing]]></copyright><item><title><![CDATA[【LLM】13：大模型算法面试题库]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/145609650</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/145609650</guid><author>zzZ_CMing</author><pubDate>Thu, 13 Feb 2025 13:07:29 +0800</pubDate><description><![CDATA[大模型算法工程师面试问题]]></description><category></category></item><item><title><![CDATA[【LLM】12：MoE混合专家模型实现]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/145510630</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/145510630</guid><author>zzZ_CMing</author><pubDate>Sat, 08 Feb 2025 10:51:31 +0800</pubDate><description><![CDATA[MoE代码实现]]></description><category></category></item><item><title><![CDATA[【LLM】11：LangChain实战应用分享]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143606018</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143606018</guid><author>zzZ_CMing</author><pubDate>Thu, 07 Nov 2024 19:53:11 +0800</pubDate><description><![CDATA[langchain实战应用]]></description><category></category></item><item><title><![CDATA[【LLM】10：LangChain主要节点说明]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143605460</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143605460</guid><author>zzZ_CMing</author><pubDate>Thu, 07 Nov 2024 19:39:56 +0800</pubDate><description><![CDATA[langchain主要节点说明]]></description><category></category></item><item><title><![CDATA[【LLM】3：从零开始训练大语言模型（预训练PT、微调SFT、RLHF）]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143591945</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143591945</guid><author>zzZ_CMing</author><pubDate>Thu, 07 Nov 2024 15:51:41 +0800</pubDate><description><![CDATA[奖励函数用于评估生成模型的输出质量。将输入文本和生成的响应传递给奖励模型，得到分数并作为奖励信号返回。]]></description><category></category></item><item><title><![CDATA[【LLM】7：混合精度量化的实现]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143221583</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143221583</guid><author>zzZ_CMing</author><pubDate>Thu, 24 Oct 2024 22:47:05 +0800</pubDate><description><![CDATA[以上示例展示了如何在 PyTorch 和 TensorFlow 中实现混合精度训练。]]></description><category></category></item><item><title><![CDATA[【LLM】9：RAG流程详解]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143099958</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/143099958</guid><author>zzZ_CMing</author><pubDate>Sun, 20 Oct 2024 22:07:24 +0800</pubDate><description><![CDATA[是为了弥补大模型在特定领域下知识的不足，整个过程相对稳定，大模型可发挥的空间较少；]]></description><category></category></item><item><title><![CDATA[【LLM】6：大语言模型中的 KV 缓存机制]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/142891497</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/142891497</guid><author>zzZ_CMing</author><pubDate>Sat, 12 Oct 2024 23:50:42 +0800</pubDate><description><![CDATA[KV 缓存在自注意力机制中的作用是在推理阶段缓存之前计算过的 Key 和 Value 矩阵，避免在生成新 token 时重复计算，显著提升推理效率，特别是在长序列生成任务中效果明显。这种缓存机制是大模型推理阶段提高性能的关键优化之一。]]></description><category></category></item><item><title><![CDATA[【LLM】5：Self-Attention中注意力分数的含义]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/142891155</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/142891155</guid><author>zzZ_CMing</author><pubDate>Sat, 12 Oct 2024 23:35:24 +0800</pubDate><description><![CDATA[自注意力机制（Self-Attention）是Transformer模型的核心组件，其中Q、K、V矩阵分别代表查询（Query）、键（Key）、值（Value）。它们的作用和含义可以通过信息匹配过程来理解。在自注意力机制中，输入的序列数据会经过线性变换，生成这三个矩阵，然后用于计算注意力权重和最终输出。]]></description><category></category></item><item><title><![CDATA[【LLM】8：大语言模型的量化（GPTQ、GGUF、AWQ原理）]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/142464904</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/142464904</guid><author>zzZ_CMing</author><pubDate>Mon, 23 Sep 2024 18:20:03 +0800</pubDate><description><![CDATA[GPTQ 通过梯度优化对量化误差进行最小化，适用于后训练阶段的精细量化，精度较高。GGUF 采用全局统一的量化策略，具有简单高效的优点，适用于资源受限的部署场景，但可能导致某些模型层的精度损失。AWQ 关注激活值的量化，通过分析激活值的分布对量化策略进行自适应调整，精度更高但计算复杂度较大。]]></description><category></category></item><item><title><![CDATA[【LLM】4：LoRA原理]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/135141992</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/135141992</guid><author>zzZ_CMing</author><pubDate>Thu, 21 Dec 2023 23:29:10 +0800</pubDate><description><![CDATA[有关LoRA模型的原理]]></description><category></category></item><item><title><![CDATA[【LLM】1：如何更好的使用Transformers库]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/134497369</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/134497369</guid><author>zzZ_CMing</author><pubDate>Sun, 19 Nov 2023 23:02:17 +0800</pubDate><description><![CDATA[有关transformers库的学习资料。]]></description><category></category></item><item><title><![CDATA[【LLM】2：Transformer底层算法原理]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/134112362</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/134112362</guid><author>zzZ_CMing</author><pubDate>Mon, 30 Oct 2023 16:59:18 +0800</pubDate><description><![CDATA[transformer架构的原理]]></description><category></category></item><item><title><![CDATA[CLIP模型原理]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133908101</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133908101</guid><author>zzZ_CMing</author><pubDate>Wed, 18 Oct 2023 16:02:40 +0800</pubDate><description><![CDATA[CLIP 模型原理]]></description><category></category></item><item><title><![CDATA[Stable Diffusion算法原理]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133885399</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133885399</guid><author>zzZ_CMing</author><pubDate>Tue, 17 Oct 2023 17:07:43 +0800</pubDate><description><![CDATA[diffusion扩散理论，stable diffusion潜在扩散理论基础知识]]></description><category></category></item><item><title><![CDATA[python图片上写中文，添加字幕]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133884481</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133884481</guid><author>zzZ_CMing</author><pubDate>Tue, 17 Oct 2023 15:08:00 +0800</pubDate><description><![CDATA[在图片上添加中文，添加字幕]]></description><category></category></item><item><title><![CDATA[Flask接口]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133884296</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133884296</guid><author>zzZ_CMing</author><pubDate>Tue, 17 Oct 2023 15:01:19 +0800</pubDate><description><![CDATA[Flask写后台]]></description><category></category></item><item><title><![CDATA[Stable Diffusion的模型选择，采样器选择，关键词]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133883834</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/133883834</guid><author>zzZ_CMing</author><pubDate>Tue, 17 Oct 2023 14:57:28 +0800</pubDate><description><![CDATA[SD的一些知识]]></description><category></category></item><item><title><![CDATA[【SHM】Semantic Human Matting图像分割]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/109490676</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/109490676</guid><author>zzZ_CMing</author><pubDate>Wed, 04 Nov 2020 15:38:23 +0800</pubDate><description><![CDATA[2018年阿里的论文《Semantatic Human Matting》给出了抠图方法复现]]></description><category></category></item><item><title><![CDATA[mask图生成trimap图的方式]]></title><link>https://blog.youkuaiyun.com/zzZ_CMing/article/details/103939055</link><guid>https://blog.youkuaiyun.com/zzZ_CMing/article/details/103939055</guid><author>zzZ_CMing</author><pubDate>Fri, 30 Oct 2020 15:08:14 +0800</pubDate><description><![CDATA[trimap图在AI抠像中的用途是为了得到精准的alpha图，以便后续的合成。trimap确定需要的前景区域位置——下右图的白色区域；确定不需要的背景区域位置——下右图的黑色区域；介于需要与不需要的待分割区域位置——下右图的灰色区域；trimap图大多都是由人工处理得到的，而标记的过程耗时耗力。这里介绍一种基于mask图生成trimap图的方法，时间效率要比手动处理快，但是效果表现有待提高。]]></description><category></category></item></channel></rss>