📉 学习率调度可视化（Warmup + Cosine）Learning Rate Schedule Visualizer (Warmup + Cosine)
输入峰值学习率、最小学习率、总步数和 warmup 步数，实时画出 warmup + 余弦衰减、线性衰减、常数、余弦重启、WSD（预热-稳定-衰减）和逆平方根等学习率曲线，可查询任意步的学习率，导出每步数值 CSV、SVG 图和对应的 transformers 代码。Enter peak and minimum learning rate, total and warmup steps to plot warmup + cosine, linear, constant, cosine with restarts, WSD (warmup-stable-decay) and inverse-sqrt schedules; read the LR at any step, and export per-step CSV, an SVG chart and matching transformers code.
公式与 Hugging Face transformers 的 get_cosine_schedule_with_warmup 等函数一致，并加入了最小学习率：lr = min + (peak − min) × f(进度)。Warmup 阶段从 0 线性升到峰值。Formulas match Hugging Face transformers (get_cosine_schedule_with_warmup etc.) with a minimum LR added: lr = min + (peak − min) × f(progress). Warmup rises linearly from 0 to the peak.
使用说明与常见问题How-to & FAQ
「学习率调度可视化（Warmup + Cosine）」是什么？What is the Learning Rate Schedule Visualizer (Warmup + Cosine)?
输入峰值学习率、最小学习率、总步数和 warmup 步数，实时画出 warmup + 余弦衰减、线性衰减、常数、余弦重启、WSD（预热-稳定-衰减）和逆平方根等学习率曲线，可查询任意步的学习率，导出每步数值 CSV、SVG 图和对应的 transformers 代码。Enter peak and minimum learning rate, total and warmup steps to plot warmup + cosine, linear, constant, cosine with restarts, WSD (warmup-stable-decay) and inverse-sqrt schedules; read the LR at any step, and export per-step CSV, an SVG chart and matching transformers code.
如何使用「学习率调度可视化（Warmup + Cosine）」？How do I use the Learning Rate Schedule Visualizer (Warmup + Cosine)?
选择调度类型并填写参数，曲线实时更新；鼠标移到图上可查看每一步的学习率，橙色区域是 warmup 阶段。在“查询步数”中输入步数可直接读出数值。导出的 CSV 每行一步，适合画进论文图表。Pick a schedule and fill in the numbers; the curve updates live. Hover to read the LR at each step (orange = warmup) or type a step to get the exact value. The CSV export has one row per step for paper figures.
和 Hugging Face 的调度器一致吗？Does it match Hugging Face schedulers?
一致。最小学习率为 0 时，曲线与 transformers 的 get_cosine_schedule_with_warmup、get_linear_schedule_with_warmup 等函数逐步相同；设置最小学习率后按 min + (peak − min) × f 计算。Yes. With min LR 0 the curves equal transformers’ get_cosine_schedule_with_warmup, get_linear_schedule_with_warmup and friends step by step; with a min LR the formula is min + (peak − min) × f.
WSD 是什么？What is WSD?
WSD（Warmup-Stable-Decay）先 warmup，再长时间保持峰值，最后一小段（如 10%–20%）快速衰减，近年在 MiniCPM、DeepSeek 等大模型训练中常用，便于中途续训。Warmup-Stable-Decay: warm up, hold the peak for most of training, then decay quickly over the last 10%–20%; popular in recent LLM training (MiniCPM, DeepSeek) because it makes continued training easy.
「学习率调度可视化（Warmup + Cosine）」免费吗？文件会上传吗？Is the Learning Rate Schedule Visualizer (Warmup + Cosine) free? Are my files uploaded?
「学习率调度可视化（Warmup + Cosine）」完全免费，无需注册。你输入的内容只在本机浏览器里处理（并保存在本机方便下次继续），不会上传到任何服务器，适合处理未发表的论文和数据。The Learning Rate Schedule Visualizer (Warmup + Cosine) is free with no sign-up. Everything you enter is processed in this browser only (and kept locally so you can continue later); nothing is uploaded, so it is safe for unpublished papers and data.