<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>재미있는 삽질 기록</title>
    <link>https://woohyun.dev/</link>
    <description>재미있는 삽질 기록</description>
    <language>ko</language>
    <atom:link href="https://woohyun.dev/rss.xml" rel="self" type="application/rss+xml" />
  <item>
    <title>claude code v2.1.199 release note 살펴보기</title>
    <link>https://woohyun.dev/2.1.199/</link>
    <guid isPermaLink="true">https://woohyun.dev/2.1.199/</guid>
    <pubDate>Fri, 03 Jul 2026 00:00:00 GMT</pubDate>
    <description>Claude Code v2.1.199 릴리스 노트 살펴보기 안녕하세요. 이번 글에서는 Claude Code v2.1.199 릴리스 노트의 변경 사항을 정리합니다. 도입 v2.1.199 는 2.1.x 계열의 패치 릴리스입니다. 릴리스 노트에는 24개 항목이 올라와 있고, 이 중 대부...</description>
  </item>
  <item>
    <title>260701 회고</title>
    <link>https://woohyun.dev/260701/</link>
    <guid isPermaLink="true">https://woohyun.dev/260701/</guid>
    <pubDate>Thu, 02 Jul 2026 00:00:00 GMT</pubDate>
    <description>계속 개발과 취업 준비를 병행하면서 든 생각이다. 최근에 취업 시장에서는 AI를 잘 활용할 수 있는 능력을 중요시하게 여기기 시작했다. 이 능력을 검증할 수 있는 방법에 대해서 생각해봤는데, 가장 직관적인 방법은 Claude code나 Codex 등의 Agentic tool을 어떤...</description>
  </item>
  <item>
    <title>Claude Fable 5, Mythos 5 살펴보기</title>
    <link>https://woohyun.dev/claude-fable-5/</link>
    <guid isPermaLink="true">https://woohyun.dev/claude-fable-5/</guid>
    <pubDate>Wed, 10 Jun 2026 00:00:00 GMT</pubDate>
    <description>2026년 7월 1일, Fable 5·Mythos 5 재배포 6월 12일 미국 정부의 수출 통제(export controls)로 중단됐던 두 모델은 6월 30일 통제 해제 후 7월 1일 다시 배포됐다. Fable 5는 전 세계 사용자에게, Mythos 5는 승인된 일부 미국 조직...</description>
  </item>
  <item>
    <title>Claude Opus 4.8 살펴보기</title>
    <link>https://woohyun.dev/claude-opus-4-8/</link>
    <guid isPermaLink="true">https://woohyun.dev/claude-opus-4-8/</guid>
    <pubDate>Fri, 29 May 2026 00:00:00 GMT</pubDate>
    <description>오늘 새벽 두 시에 갑자기 Claude Opus 4.8이 출시됐다. 요새 클로드의 모델 업데이트 주기가 부쩍 빨라졌다. 작년만 해도 프론티어 모델의 업데이트는 최소 세 달을 간격으로 진행했었는데, 올해 들어서는 Opus 4.6에서 4.7로 약 두 달, 4.7에서 4.8로는 그보다...</description>
  </item>
  <item>
    <title>사공이 많아지면 - 260529 회고</title>
    <link>https://woohyun.dev/260529/</link>
    <guid isPermaLink="true">https://woohyun.dev/260529/</guid>
    <pubDate>Thu, 28 May 2026 00:00:00 GMT</pubDate>
    <description>Claude code의 개인적인 harness를 개발하는 중이고, 현재 review gate를 만들고 있는데 방향성을 잘못 잡아 며칠 째 고생하는 중이다. OMC 같은 이미 존재하는 harness를 써도 되겠지만, 개인 프로젝트를 진행하는데에는 harness가 너무 무거워 보였다....</description>
  </item>
  <item>
    <title>Antigravity CLI 로그인이 반복되는 문제 해결하기</title>
    <link>https://woohyun.dev/antigravity-cli-fix-auth-persistancy/</link>
    <guid isPermaLink="true">https://woohyun.dev/antigravity-cli-fix-auth-persistancy/</guid>
    <pubDate>Thu, 21 May 2026 00:00:00 GMT</pubDate>
    <description>안녕하세요. 최근 Google I/O에서 공개된 Antigravity CLI를 WSL에서 써보다가, 종료 후 다시 들어가면 새로 로그인해야 하는 문제가 있어서 이에 대한 문제 해결 과정을 공유하고자 작성했습니다. 문제 상황 agy 에 한 번 로그인한 뒤 종료하고 다시 실행하면, 매...</description>
  </item>
  <item>
    <title>나만의 하네스 구축하기 - CLAUDE.md</title>
    <link>https://woohyun.dev/setup-harness/</link>
    <guid isPermaLink="true">https://woohyun.dev/setup-harness/</guid>
    <pubDate>Mon, 18 May 2026 00:00:00 GMT</pubDate>
    <description>오늘은 CLAUDE.md를 제대로 다루는 법을 정리해보려고 한다. CLAUDE.md는 개인이 구축할 수 있는 가장 영향력 있는 하네스라고 생각한다. 그만큼 잘 다뤄야 하는 요소이므로, 잘 다루는 방법을 정리해봤다. CLAUDE.md가 중요한 이유 본격적으로 시작하기 전에 다음 내용...</description>
  </item>
  <item>
    <title>claude code에서 잘 사용하지 않는 skills 다루기</title>
    <link>https://woohyun.dev/claude-code-unused-skills/</link>
    <guid isPermaLink="true">https://woohyun.dev/claude-code-unused-skills/</guid>
    <pubDate>Fri, 15 May 2026 00:00:00 GMT</pubDate>
    <description>이번 글에서는 Claude Code의 Skills 설정을 다룬다. GitHub에는 Claude Code용 오픈소스 스킬 번들이 넘쳐난다. gstack , superpowers 같은 잘 알려진 컬렉션부터 개인이 묶어 올린 모음집까지, 한 번 설치하면 적게는 한두 개, 많게는 20개가...</description>
  </item>
  <item>
    <title>codex-plugin-cc 설치 및 리뷰 - Claude Code에서 Codex로 리뷰 받기</title>
    <link>https://woohyun.dev/codex-plugin-cc-review/</link>
    <guid isPermaLink="true">https://woohyun.dev/codex-plugin-cc-review/</guid>
    <pubDate>Thu, 14 May 2026 00:00:00 GMT</pubDate>
    <description>이번 글은 OpenAI가 공식적으로 내놓은, Claude Code에서 쓸 수 있는 Codex 리뷰 플러그인을 다룬다. 사실 나온 지는 좀 됐는데, 정리할 겸 설치 과정과 리뷰 사용법을 함께 적어 보았다. 이 플러그인이 갖는 의의 Claude Code로 작성한 코드를 Codex에게...</description>
  </item>
  <item>
    <title>Matt Pocock - skills for real engineers 리뷰</title>
    <link>https://woohyun.dev/matt-pocock-skills-review/</link>
    <guid isPermaLink="true">https://woohyun.dev/matt-pocock-skills-review/</guid>
    <pubDate>Sat, 09 May 2026 00:00:00 GMT</pubDate>
    <description>오늘은 Matt Pocock이 작성한 Skills For Real Engineers 를 리뷰해보려고 한다. 해당 Github 레포지토리는 바이브 코딩이 아닌 실제 엔지니어링을 위해, 작성자가 매일 사용하는 에이전트 스킬을 모아 둔 것이라고 한다. 즉, 바이브 코딩보다는 에이전틱 엔...</description>
  </item>
  <item>
    <title>Dropout 이해하기</title>
    <link>https://woohyun.dev/deeplearning-dropout/</link>
    <guid isPermaLink="true">https://woohyun.dev/deeplearning-dropout/</guid>
    <pubDate>Mon, 05 Jan 2026 00:00:00 GMT</pubDate>
    <description>Dropout을 공부하면서 헷갈렸던 내용들을 정리해보았다. 일반화 성능을 높이는 핵심적인 요소인 마스킹과, 그리고 그 마스킹에 걸리지 않은 살아남은 값들을 키우는 스케일업. 이 구분을 중심으로 글을 작성했다. Dropout이란? Dropout 레이어에 들어온 텐서의 각 원소를 독립...</description>
  </item>
  <item>
    <title>NEFTune</title>
    <link>https://woohyun.dev/llm-practice-neftune/</link>
    <guid isPermaLink="true">https://woohyun.dev/llm-practice-neftune/</guid>
    <pubDate>Tue, 30 Dec 2025 00:00:00 GMT</pubDate>
    <description>이번 대회 학습 데이터는 약 2,000개로 적은데 모델은 27B였다. 거대 모델이 소규모 데이터를 학습하면 과적합 위험이 크다. 이걸 막으려고 NEFTune을 도입했는데, 임베딩에 노이즈를 더하는 단순한 기법이 왜 과적합을 막는지 정리해봤다. 문제: SFT 단계의 소규모 과적합 L...</description>
  </item>
  <item>
    <title>체크포인트 선정하기</title>
    <link>https://woohyun.dev/llm-practice-bestmetric/</link>
    <guid isPermaLink="true">https://woohyun.dev/llm-practice-bestmetric/</guid>
    <pubDate>Fri, 26 Dec 2025 00:00:00 GMT</pubDate>
    <description>학습이 끝나면 보통 eval/loss가 가장 낮은 체크포인트를 골라 추론에 쓴다. 나도 그렇게 했는데, 정작 loss가 더 높은 마지막 epoch 체크포인트가 실제 점수는 더 높게 나와서 한참 갸웃했다. WandB 차트를 보면서 어떤 체크포인트를 써야 하는지 다시 생각했던 걸 정리...</description>
  </item>
  <item>
    <title>Unsloth 적용기</title>
    <link>https://woohyun.dev/llm-practice-unsloth/</link>
    <guid isPermaLink="true">https://woohyun.dev/llm-practice-unsloth/</guid>
    <pubDate>Fri, 26 Dec 2025 00:00:00 GMT</pubDate>
    <description>27B 모델을 V100 32GB에서 굴리려니 속도와 메모리 양쪽이 빠듯했다. unsloth를 적용하면서 왜 쓰는지, 그리고 직접 batch size를 1·4·8로 바꿔보다 batch 1이 제일 빨랐던 이상한 결과를 추적했던 걸 정리해봤다. 왜 unsloth였나 unsloth는 같은...</description>
  </item>
  <item>
    <title>completion-only loss - 질문 말고 정답만 학습시키기</title>
    <link>https://woohyun.dev/llm-practice-completion-only-loss/</link>
    <guid isPermaLink="true">https://woohyun.dev/llm-practice-completion-only-loss/</guid>
    <pubDate>Sun, 21 Dec 2025 00:00:00 GMT</pubDate>
    <description>SFT 데이터를 학습시킬 때 질문과 정답이 한 시퀀스로 들어가는데, 처음엔 모델이 이 전체를 다 학습하는 줄 알았다. 그런데 질문은 학습하지 않고 정답 부분에만 loss를 거는 방법이 따로 있었다. 입력으로 받는 것과 loss를 거는 것의 차이를 정리해봤다. 입력으로 받는 것 ≠...</description>
  </item>
  <item>
    <title>LLM 추론 최적화 이해하기</title>
    <link>https://woohyun.dev/nlp-inference/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-inference/</guid>
    <pubDate>Tue, 16 Dec 2025 00:00:00 GMT</pubDate>
    <description>LLM 추론은 토큰을 auto-regressive하게 하나씩 생성하는 과정이라 바닐라로 돌리면 매우 느리다. 이걸 빠르게 만드는 최적화 기법들을 메모리 이동, 반복 계산, 배치 처리 관점으로 나눠 정리해봤다. 생성 단계의 sampling/search 자체는 3. Sampling 기...</description>
  </item>
  <item>
    <title>1. argmax 분해와 Greedy</title>
    <link>https://woohyun.dev/nlp-decoding/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-decoding/</guid>
    <pubDate>Fri, 28 Nov 2025 00:00:00 GMT</pubDate>
    <description>자연어 생성에서 decoding이 정확히 뭘 하는 과정인지, 그리고 왜 문장 전체를 한 번에 구하지 않고 토큰을 하나씩 뽑아 나가는지 정리해봤다. decoding이 풀려는 문제 입력 문장 x가 주어졌을 때, 출력 문장 y가 나올 확률을 최대화하는 y를 찾는 것이 decoding의...</description>
  </item>
  <item>
    <title>2. Beam Search와 그 한계</title>
    <link>https://woohyun.dev/nlp-decoding-beamsearch/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-decoding-beamsearch/</guid>
    <pubDate>Fri, 28 Nov 2025 00:00:00 GMT</pubDate>
    <description>필기에 내가 적어둔 메모 중에 2³개...? (X) 라는 게 있었다. beam search가 매 스텝 후보를 k개씩 펼치니까 스텝이 지날수록 k 2 , k 3 , … 으로 지수적으로 폭발하는 거 아닌가 싶었던 흔적이다. 결론부터 말하면 그렇지 않다. 그게 왜 그런지 따라가보면서 b...</description>
  </item>
  <item>
    <title>3. Sampling 기반 디코딩 - Temperature, Top-k, Top-p</title>
    <link>https://woohyun.dev/nlp-decoding-sampling/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-decoding-sampling/</guid>
    <pubDate>Fri, 28 Nov 2025 00:00:00 GMT</pubDate>
    <description>beam search까지 정리하고 나니, 확률을 최대화하는 방식만으로는 단조로운 문장밖에 못 만든다는 게 분명해졌다. 그래서 필기에 이런 질문을 적어뒀었다. Question 확률이 낮은 단어를 더 혹은 덜 뽑히게 할 수 있을까? 확률분포에 따라 단어를 random sampling하...</description>
  </item>
  <item>
    <title>Input Embedding과 Positional Encoding</title>
    <link>https://woohyun.dev/nlp-embedding/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-embedding/</guid>
    <pubDate>Wed, 26 Nov 2025 00:00:00 GMT</pubDate>
    <description>Transformer 입력을 정리하다가, 수식으로는 분명 행렬곱인데 PyTorch는 행렬곱을 안 한다는 부분에서 한 번 멈췄었다. 그런데도 학습이 된다는 게 처음엔 어 이게 왜 되지 싶었다. 그 부분과 positional encoding의 scaling factor를 같이 정리해봤...</description>
  </item>
  <item>
    <title>LoRA를 사용하는 이유</title>
    <link>https://woohyun.dev/llm-peft-lora/</link>
    <guid isPermaLink="true">https://woohyun.dev/llm-peft-lora/</guid>
    <pubDate>Wed, 05 Nov 2025 00:00:00 GMT</pubDate>
    <description>LoRA의 메커니즘 자체는 머릿속에 잘 그려져 있었다. 큰 weight 행렬을 두 개의 작은 행렬로 쪼개 덧붙이는 것. 그런데 정작 LoRA가 왜 PEFT 중에서 가장 널리 쓰이는지, 그 결정적 장점은 정리해두지 못했었다. 그 부분까지 채워봤다. 메커니즘 - low-rank 분해...</description>
  </item>
  <item>
    <title>Casual LM을 사용하는 이유</title>
    <link>https://woohyun.dev/llm-causallm/</link>
    <guid isPermaLink="true">https://woohyun.dev/llm-causallm/</guid>
    <pubDate>Tue, 04 Nov 2025 00:00:00 GMT</pubDate>
    <description>최근 LLM이 대부분 Decoder-Only 구조라는 게 처음엔 의아했다. 그런데 학습 효율 관점에서 보면 납득이 됐다. 문장 하나가 주어지면 매 토큰 위치마다 다음 토큰이라는 학습 신호가 생기므로, 사실상 문장 하나로 거의 모든 토큰에서 학습이 일어난다. 이 점을 정리하면서 두...</description>
  </item>
  <item>
    <title>Pretrained LM과 LLM</title>
    <link>https://woohyun.dev/llm-pretrainedlm/</link>
    <guid isPermaLink="true">https://woohyun.dev/llm-pretrainedlm/</guid>
    <pubDate>Tue, 04 Nov 2025 00:00:00 GMT</pubDate>
    <description>둘 다 대량의 코퍼스로 사전학습한 언어모델인데, 왜 Pretrained LM과 LLM을 굳이 나눠 부르는지가 처음엔 잘 와닿지 않았다. 핵심은 모델 크기보다 하나의 모델로 몇 개의 태스크를 푸느냐에 있었다. Pretrained LM BERT, GPT-1/2 같은 모델이 여기에 속한...</description>
  </item>
  <item>
    <title>Pretrain Corpus 정제와 Memorization</title>
    <link>https://woohyun.dev/llm-pretraining/</link>
    <guid isPermaLink="true">https://woohyun.dev/llm-pretraining/</guid>
    <pubDate>Tue, 04 Nov 2025 00:00:00 GMT</pubDate>
    <description>LLM은 온라인에서 긁어모을 수 있는 최대한 많은 텍스트로 사전학습한다. 그런데 그 원시 데이터를 그대로 쓰지 않고 굳이 정제 작업을 거치는 이유가 무엇인지 정리해봤다. Corpus와 원시 데이터 코퍼스는 사전학습을 위한 대량의 텍스트 데이터 집합이다. 원시 데이터는 블로그, 뉴스...</description>
  </item>
  <item>
    <title>Stacking 앙상블 직접 구현하기</title>
    <link>https://woohyun.dev/ml-stacking/</link>
    <guid isPermaLink="true">https://woohyun.dev/ml-stacking/</guid>
    <pubDate>Fri, 24 Oct 2025 00:00:00 GMT</pubDate>
    <description>경진대회에서 Stacking 앙상블을 직접 구현하면서 배운 내용을 정리한다. Stacking은 여러 모델의 장점을 모아 더 강한 모델을 만드는 앙상블 기법이다. 직접 해보며 가장 크게 깨달은 점은, Stacking을 KFold와 반드시 함께 써야 한다는 것이었다. Stacking에...</description>
  </item>
  <item>
    <title>문자열 단위 Language Model</title>
    <link>https://woohyun.dev/nlp-language-model/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-language-model/</guid>
    <pubDate>Thu, 02 Oct 2025 00:00:00 GMT</pubDate>
    <description>RNN으로 문자 하나씩을 예측하는 문자열 단위 Language Model을 hello 예제로 따라가며, 입력을 One-Hot으로 주는 이유와 학습/추론의 차이를 정리한다. hello 예제 Vocabulary를 [h, e, l, o] 로 두고, 학습 문자열 hello 를 학습한다고...</description>
  </item>
  <item>
    <title>ResNet에서 LSTM까지</title>
    <link>https://woohyun.dev/nlp-lstm/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-lstm/</guid>
    <pubDate>Thu, 02 Oct 2025 00:00:00 GMT</pubDate>
    <description>LSTM이 RNN의 기울기 문제를 어떻게 우회하는지를, ResNet의 skip connection에서 Highway network를 거쳐 LSTM에 이르는 흐름으로 따라간다. 마지막에는 nn.LSTM 으로 양방향 RNN을 쓰는 법까지 정리한다. ResNet의 skip connect...</description>
  </item>
  <item>
    <title>BPTT와 TBPTT</title>
    <link>https://woohyun.dev/nlp-rnn-bptt/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-rnn-bptt/</guid>
    <pubDate>Thu, 02 Oct 2025 00:00:00 GMT</pubDate>
    <description>RNN을 시간 축으로 펼쳐 역전파하는 BPTT와, 긴 시퀀스를 잘라 부분적으로 역전파하는 TBPTT를 정리한다. BPTT(Backpropagation Through Time) RNN은 시퀀스를 따라 forward를 진행하면서 각 타임스텝마다 출력과 Loss를 만든다. Loss는 f...</description>
  </item>
  <item>
    <title>Static, Dynamic Embedding과 VectorDB, RAG</title>
    <link>https://woohyun.dev/nlp-kindofembedding/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-kindofembedding/</guid>
    <pubDate>Tue, 30 Sep 2025 00:00:00 GMT</pubDate>
    <description>Word2Vec 같은 Static Embedding의 한계와 문맥에 따라 변하는 Dynamic Embedding, 학습되지 않은 정보를 다루기 위한 VectorDB와 RAG에 대한 내용을 짧게 정리해보았다. Static Embedding의 한계 Word Embedding과 Word...</description>
  </item>
  <item>
    <title>Word, Character, Subword 토큰화와 BPE</title>
    <link>https://woohyun.dev/nlp-tokenize-bpe/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-tokenize-bpe/</guid>
    <pubDate>Tue, 30 Sep 2025 00:00:00 GMT</pubDate>
    <description>텍스트를 어떤 단위로 쪼갤지에 따라 Word / Character / Subword level 토큰화가 나뉜다. 각각의 장단점과, Subword 방법의 대표적인 BPE의 동작 원리를 정리해봤다. Word-level 토큰화 Token을 단어(Word) 단위로 구분하는 방식이다. 일반...</description>
  </item>
  <item>
    <title>Word Embedding과 Word2Vec</title>
    <link>https://woohyun.dev/nlp-word2vec/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-word2vec/</guid>
    <pubDate>Tue, 30 Sep 2025 00:00:00 GMT</pubDate>
    <description>단어를 벡터로 표현하는 두 방식(One-Hot Encoding과 Word Embedding)을 비교하고, Word Embedding을 만드는 대표적 방법인 Word2Vec의 원리와 forward 구현까지 정리한다. One-Hot Encoding과 Word Embedding One-...</description>
  </item>
  <item>
    <title>ReLU의 존재 의미</title>
    <link>https://woohyun.dev/deeplearning-whyrelu/</link>
    <guid isPermaLink="true">https://woohyun.dev/deeplearning-whyrelu/</guid>
    <pubDate>Mon, 29 Sep 2025 00:00:00 GMT</pubDate>
    <description>ReLU가 기울기 문제를 해결한다고 뭉뚱그려 말하기 쉽지만, ReLU가 다루는 것은 vanishing이지 exploding이 아니다. 이 구분을 정리해봤다. ReLU가 해결하는 건 vanishing이다 역전파에서 gradient는 각 층의 activation 미분을 계속 곱하면서...</description>
  </item>
  <item>
    <title>Multi-Head Attention 직접 구현하기</title>
    <link>https://woohyun.dev/nlp-transformer-mha/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-transformer-mha/</guid>
    <pubDate>Sun, 21 Sep 2025 00:00:00 GMT</pubDate>
    <description>멀티헤드 어텐션을 직접 구현하면서 헷갈렸던 지점들을 정리한다. 헤드의 차원은 embedding_dim을 헤드 수로 나눈 것 먼저 헤드를 빼고 보면, 입력은 (B, L, D) = (batch_size, sequence_length, embedding_dim)이고, W q ​ , W...</description>
  </item>
  <item>
    <title>Tokenizer 이해하기</title>
    <link>https://woohyun.dev/nlp-tokenizer/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-tokenizer/</guid>
    <pubDate>Sat, 20 Sep 2025 00:00:00 GMT</pubDate>
    <description>Transformer에 문장을 넣기 전, 텍스트를 어떻게 숫자로 바꾸는지 tokenizer를 정리한다. 왜 필요한가 Transformer에 넣을 입력이 numeric 데이터가 아니라면 그대로 넣을 수 없다. 텍스트를 숫자로 바꾸는 과정이 필요한데, 그게 tokenizer가 하는 일...</description>
  </item>
  <item>
    <title>nn.Embedding에 이해하기</title>
    <link>https://woohyun.dev/pytorch-embedding/</link>
    <guid isPermaLink="true">https://woohyun.dev/pytorch-embedding/</guid>
    <pubDate>Sat, 20 Sep 2025 00:00:00 GMT</pubDate>
    <description>토큰 번호로 바뀐 문장이 어떻게 벡터가 되는지, nn.Embedding 의 동작을 간단하게 정리해봤다. (vocab_size, model_dim) 가중치 행렬 nn.Embedding 은 결국 (vocab_size, model_dimension) 크기의 가중치 행렬 하나다. 이 행렬...</description>
  </item>
  <item>
    <title>Self-Attention 이해하기</title>
    <link>https://woohyun.dev/nlp-attention/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-attention/</guid>
    <pubDate>Thu, 18 Sep 2025 00:00:00 GMT</pubDate>
    <description>attention이 무엇을 계산하는지, Q와 K와 V가 각각 어떤 역할인지, 그리고 한 단어를 어떻게 contextualize하는지 간단히 정리해봤다. attention은 value의 가중 평균 attention 함수는 세 가지를 입력으로 받는다. Attention ( Q , K...</description>
  </item>
  <item>
    <title>Teacher Forcing과 Auto-regressive 생성</title>
    <link>https://woohyun.dev/nlp-autoregressive/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-autoregressive/</guid>
    <pubDate>Thu, 18 Sep 2025 00:00:00 GMT</pubDate>
    <description>seq2seq decoder가 출력을 하나씩 만들어내는 방식과, 학습 때와 추론 때 입력을 다르게 주는 이유를 정리한다. Auto-regressive 생성 decoder는 한 시점에 한 토큰씩 출력을 만든다. 각 단계의 hidden state는 입력 시퀀스 전체에 대한 정보(con...</description>
  </item>
  <item>
    <title>RNN과 기울기 소실, 폭발</title>
    <link>https://woohyun.dev/nlp-rnn-gradient-vanishing/</link>
    <guid isPermaLink="true">https://woohyun.dev/nlp-rnn-gradient-vanishing/</guid>
    <pubDate>Thu, 18 Sep 2025 00:00:00 GMT</pubDate>
    <description>RNN이 시계열을 어떻게 처리하고, 왜 기울기 소실과 폭발에 시달리는지 정리한다. RNN과 recurrence RNN은 시계열 데이터를 처리하기 위한 모델이다. 시점 t의 hidden state는 이전 시점의 hidden state와 현재 입력으로 정해진다. h t ​ = f W...</description>
  </item>
  <item>
    <title>Backpropagation과 Chain Rule</title>
    <link>https://woohyun.dev/deeplearning-backpropagation/</link>
    <guid isPermaLink="true">https://woohyun.dev/deeplearning-backpropagation/</guid>
    <pubDate>Wed, 17 Sep 2025 00:00:00 GMT</pubDate>
    <description>역전파가 어떻게 gradient를 계산하는지, local gradient와 chain rule을 중심으로 정리해봤고, 역전파 계산에 자주 나오는 연산 게이트들이 gradient를 어떻게 흘려보내는지까지 정리해봤다. local gradient와 chain rule 어떤 연산 노드 f...</description>
  </item>
  <item>
    <title>clone과 detach, 그리고 계산 그래프</title>
    <link>https://woohyun.dev/pytorch-whycloneanddetach/</link>
    <guid isPermaLink="true">https://woohyun.dev/pytorch-whycloneanddetach/</guid>
    <pubDate>Sat, 06 Sep 2025 00:00:00 GMT</pubDate>
    <description>Pytorch에서 Tensor를 다룰 때 경고문을 자주 보았을 것이다. UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.detach().clone() ... rather than...</description>
  </item>
  <item>
    <title>Linear Layer와 ReLU를 직접 구현해보기</title>
    <link>https://woohyun.dev/pytorch-implement-linear/</link>
    <guid isPermaLink="true">https://woohyun.dev/pytorch-implement-linear/</guid>
    <pubDate>Fri, 05 Sep 2025 00:00:00 GMT</pubDate>
    <description>nn.Linear 와 nn.ReLU 라는 이미 완성된 클래스가 존재하지만, 직접 구현해보면 더 명확하게 이해할 수 있고 재미있는 사실들도 발견할 수 있다. torch.nn 없이 직접 구현해보았다. Linear Layer 구현하기 선형 계층은 입력에 가중치를 곱하고 편향을 더하는 y...</description>
  </item>
  <item>
    <title>가중치 초기화와 Kaiming(He) 초기화</title>
    <link>https://woohyun.dev/weight-initialization/</link>
    <guid isPermaLink="true">https://woohyun.dev/weight-initialization/</guid>
    <pubDate>Fri, 05 Sep 2025 00:00:00 GMT</pubDate>
    <description>간단한 Linear Layer를 pytorch로 직접 구현해보면서, 가중치를 어떻게 초기화하느냐에 따라 학습 결과가 달라지는 것을 확인해보았다. 해당 내용에 대한 짧은 글이다. 표준정규분포로 초기화 torch.randn 으로 가중치를 초기화하면 표준정규분포(평균 0, 표준편차 1)...</description>
  </item>
  <item>
    <title>model.eval()과 torch.no_grad()</title>
    <link>https://woohyun.dev/pytorch-training-eval/</link>
    <guid isPermaLink="true">https://woohyun.dev/pytorch-training-eval/</guid>
    <pubDate>Thu, 04 Sep 2025 00:00:00 GMT</pubDate>
    <description>Pytorch의 모델 평가 단계에서 거의 항상 같이 등장하는 model.eval() 과 torch.no_grad() 를 정리해봤다. model.eval() model.eval() 은 모델을 평가, 즉 Evaluation에 최적화된 상태로 만드는 과정이다. Training 과정에는...</description>
  </item>
  <item>
    <title>Dataset과 DataLoader, 그리고 데이터 파이프라인</title>
    <link>https://woohyun.dev/pytorch-utils-datapipe/</link>
    <guid isPermaLink="true">https://woohyun.dev/pytorch-utils-datapipe/</guid>
    <pubDate>Thu, 04 Sep 2025 00:00:00 GMT</pubDate>
    <description>PyTorch에서 데이터가 모델에 들어가기까지의 파이프라인을 Dataset , DataLoader , 그리고 이미지 변환인 ToTensor 을 중심으로 정리. Dataset 클래스 PyTorch에서는 데이터의 전처리와 배치 처리를 용이하게 할 수 있도록 Dataset 과 DataL...</description>
  </item>
  <item>
    <title>zero_grad 가 필요한 이유</title>
    <link>https://woohyun.dev/pytorch-zerograd/</link>
    <guid isPermaLink="true">https://woohyun.dev/pytorch-zerograd/</guid>
    <pubDate>Wed, 03 Sep 2025 00:00:00 GMT</pubDate>
    <description>학습 루프에서 자주 빠뜨리는 게 optimizer.zero_grad() 다. 그런데 이게 없으면 학습이 망가진다. 왜 매 스텝마다 기울기를 초기화해야 하는지 짧게 정리해봤다. loss.backward()는 기울기를 누적한다 핵심은 loss.backward() 의 동작이다. loss...</description>
  </item>
  <item>
    <title>1. dim 쉽게 이해하기</title>
    <link>https://woohyun.dev/pytorch-tensor-easydim/</link>
    <guid isPermaLink="true">https://woohyun.dev/pytorch-tensor-easydim/</guid>
    <pubDate>Tue, 02 Sep 2025 00:00:00 GMT</pubDate>
    <description>한참 Pytorch를 공부하고 있을 때, tensor.sum(dim=x) , torch.stack((tensors), dim=x) 등의 연산에서 dim=x 가 어떤 방향으로 연산이 이루어지는 건지 이해하기 어려웠었는데, 쉽게 이해할 수 있었던 방법을 정리해봤다. 단순한 규칙 알아두...</description>
  </item>
  <item>
    <title>2. View의 작동 원리 이해하기</title>
    <link>https://woohyun.dev/pytorch-tensor-view/</link>
    <guid isPermaLink="true">https://woohyun.dev/pytorch-tensor-view/</guid>
    <pubDate>Tue, 02 Sep 2025 00:00:00 GMT</pubDate>
    <description>PyTorch 공부를 하던 도중, 퀴즈를 풀다가 틀렸는데 아래 지문이 옳은 지문이라는 사실을 알고 나서 조금 놀랐었다. 당시 퀴즈 t = Tensor(3, 2) 이고, s = t[:2, :1] 일 때, view 를 사용해 s의 모양을 변경할 수 있다. 나는 s 가 contiguou...</description>
  </item>
  </channel>
</rss>