| 1 | <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1280 720" width="1280" height="720"> |
| 2 | <defs> |
| 3 | <filter id="cardShadow" x="-10%" y="-10%" width="120%" height="120%"> |
| 4 | <feGaussianBlur in="SourceAlpha" stdDeviation="4"/> |
| 5 | <feOffset dx="0" dy="3"/> |
| 6 | <feFlood flood-color="#1A365D" flood-opacity="0.07"/> |
| 7 | <feComposite in2="SourceAlpha" operator="in"/> |
| 8 | <feMerge><feMergeNode/><feMergeNode in="SourceGraphic"/></feMerge> |
| 9 | </filter> |
| 10 | </defs> |
| 11 | |
| 12 | <g id="bg"> |
| 13 | <rect width="1280" height="720" fill="#FFFFFF"/> |
| 14 | </g> |
| 15 | |
| 16 | <g id="header"> |
| 17 | <rect x="60" y="56" width="6" height="40" fill="#3182CE"/> |
| 18 | <text x="84" y="88" font-family="Georgia, "Microsoft YaHei", serif" font-size="36" font-weight="700" fill="#1A365D">How they trained it</text> |
| 19 | <text x="84" y="118" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1.5">TRAINING SETUP · §5 WMT'14 ENGLISH-GERMAN & ENGLISH-FRENCH</text> |
| 20 | </g> |
| 21 | |
| 22 | <g id="card-hardware"> |
| 23 | <rect x="60" y="160" width="580" height="240" rx="14" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5" filter="url(#cardShadow)"/> |
| 24 | <rect x="60" y="160" width="6" height="240" rx="3" fill="#1A365D"/> |
| 25 | <use data-icon="tabler-outline/cpu" x="84" y="180" width="32" height="32" fill="#1A365D"/> |
| 26 | <text x="84" y="240" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1.5">HARDWARE & SCALE</text> |
| 27 | <text x="84" y="296" font-family="Georgia, "Microsoft YaHei", serif" font-size="40" font-weight="700" fill="#1A365D">8 × NVIDIA P100</text> |
| 28 | <text x="84" y="332" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" fill="#4A5568">One machine, eight GPUs — a single-node setup.</text> |
| 29 | <text x="84" y="354" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" fill="#4A5568">Batches of ~25,000 source + 25,000 target tokens,</text> |
| 30 | <text x="84" y="376" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" fill="#4A5568">grouped by similar sequence length for efficiency.</text> |
| 31 | </g> |
| 32 | |
| 33 | <g id="card-time"> |
| 34 | <rect x="660" y="160" width="560" height="240" rx="14" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5" filter="url(#cardShadow)"/> |
| 35 | <rect x="660" y="160" width="6" height="240" rx="3" fill="#3182CE"/> |
| 36 | <use data-icon="tabler-outline/bolt" x="684" y="180" width="32" height="32" fill="#3182CE"/> |
| 37 | <text x="684" y="240" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1.5">TRAINING TIME</text> |
| 38 | <g id="time-base"> |
| 39 | <text x="684" y="296" font-family="Georgia, "Microsoft YaHei", serif" font-size="32" font-weight="700" fill="#1A365D">base</text> |
| 40 | <text x="820" y="296" font-family="Georgia, "Microsoft YaHei", serif" font-size="32" font-weight="700" fill="#3182CE">12 h</text> |
| 41 | <text x="940" y="296" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" fill="#4A5568">· 100k steps</text> |
| 42 | </g> |
| 43 | <g id="time-big"> |
| 44 | <text x="684" y="346" font-family="Georgia, "Microsoft YaHei", serif" font-size="32" font-weight="700" fill="#1A365D">big</text> |
| 45 | <text x="820" y="346" font-family="Georgia, "Microsoft YaHei", serif" font-size="32" font-weight="700" fill="#3182CE">3.5 d</text> |
| 46 | <text x="940" y="346" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="15" fill="#4A5568">· 300k steps</text> |
| 47 | </g> |
| 48 | <text x="684" y="386" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="14" font-style="italic" fill="#4A5568">A small fraction of the training cost of prior best models.</text> |
| 49 | </g> |
| 50 | |
| 51 | <g id="card-optimizer"> |
| 52 | <rect x="60" y="420" width="580" height="240" rx="14" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5" filter="url(#cardShadow)"/> |
| 53 | <rect x="60" y="420" width="6" height="240" rx="3" fill="#1A365D"/> |
| 54 | <use data-icon="tabler-outline/settings" x="84" y="440" width="32" height="32" fill="#1A365D"/> |
| 55 | <text x="84" y="500" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1.5">OPTIMIZER</text> |
| 56 | <text x="84" y="546" font-family="Georgia, "Microsoft YaHei", serif" font-size="32" font-weight="700" fill="#1A365D">Adam</text> |
| 57 | <text x="84" y="578" font-family="Consolas, "Courier New", monospace" font-size="15" fill="#1A202C">β₁ = 0.9 · β₂ = 0.98 · ε = 1e-9</text> |
| 58 | <text x="84" y="610" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="14" fill="#4A5568">Custom learning-rate schedule:</text> |
| 59 | <text x="84" y="632" font-family="Consolas, "Courier New", monospace" font-size="14" fill="#3182CE">lr ∝ min(step⁻⁰·⁵, step · warmup⁻¹·⁵)</text> |
| 60 | <text x="84" y="652" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-style="italic" fill="#4A5568">warm up for 4,000 steps, then decay as 1/√step.</text> |
| 61 | </g> |
| 62 | |
| 63 | <g id="card-regularization"> |
| 64 | <rect x="660" y="420" width="560" height="240" rx="14" fill="#FFFFFF" stroke="#E2E8F0" stroke-width="1.5" filter="url(#cardShadow)"/> |
| 65 | <rect x="660" y="420" width="6" height="240" rx="3" fill="#3182CE"/> |
| 66 | <use data-icon="tabler-outline/shield" x="684" y="440" width="32" height="32" fill="#3182CE"/> |
| 67 | <text x="684" y="500" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" font-weight="700" fill="#4A5568" letter-spacing="1.5">REGULARIZATION</text> |
| 68 | <g id="reg-dropout"> |
| 69 | <text x="684" y="544" font-family="Georgia, "Microsoft YaHei", serif" font-size="22" font-weight="700" fill="#1A365D">Dropout</text> |
| 70 | <text x="850" y="544" font-family="Consolas, "Courier New", monospace" font-size="20" font-weight="700" fill="#3182CE">p = 0.1</text> |
| 71 | <text x="684" y="566" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" fill="#4A5568">on every sub-layer output and on the sum of embedding + PE.</text> |
| 72 | </g> |
| 73 | <g id="reg-label-smoothing"> |
| 74 | <text x="684" y="610" font-family="Georgia, "Microsoft YaHei", serif" font-size="22" font-weight="700" fill="#1A365D">Label smoothing</text> |
| 75 | <text x="950" y="610" font-family="Consolas, "Courier New", monospace" font-size="20" font-weight="700" fill="#3182CE">ε_ls = 0.1</text> |
| 76 | <text x="684" y="632" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="13" fill="#4A5568">Slightly hurts perplexity, but improves BLEU and accuracy.</text> |
| 77 | </g> |
| 78 | </g> |
| 79 | |
| 80 | <g id="footer"> |
| 81 | <text x="60" y="694" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="11" fill="#A0AEC0">Source: Vaswani et al. 2017, §5.1 Training Data and Batching · §5.2 Hardware and Schedule · §5.3 Optimizer · §5.4 Regularization</text> |
| 82 | <text x="1220" y="694" text-anchor="end" font-family="Arial, "Microsoft YaHei", sans-serif" font-size="11" fill="#A0AEC0">13</text> |
| 83 | </g> |
| 84 | </svg> |
| 85 |