flykhan/maths-cs-ai-compendium-zh
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Deployed 2536c93 with MkDocs version: 1.6.1

Browse Source
This commit is contained in:
flykhan
2026-05-03 11:47:30 +08:00
commit 1ff86b66fc
418 changed files with 617336 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
images/transformer_block.svg
+68
View File
@@ -0,0 +1,68 @@
<svg width="700" height="320" xmlns="http://www.w3.org/2000/svg">
<defs>
<marker id="tf-arrow" markerWidth="8" markerHeight="6" refX="8" refY="3" orient="auto">
<path d="M0,0 L8,3 L0,6" fill="none" stroke="#555" stroke-width="1"/>
</marker>
</defs>
<text x="350" y="22" fill="#333" font-size="14" font-weight="bold" text-anchor="middle">Transformer Encoder Block</text>
<!-- Input at bottom -->
<text x="250" y="305" fill="#666" font-size="11" text-anchor="middle">Input embeddings + positional encoding</text>
<line x1="250" y1="293" x2="250" y2="275" stroke="#555" stroke-width="1.5" marker-end="url(#tf-arrow)"/>
<!-- Multi-Head Attention -->
<rect x="170" y="240" width="160" height="35" rx="6" fill="#3498db" opacity="0.15" stroke="#3498db" stroke-width="1.5"/>
<text x="250" y="262" fill="#3498db" font-size="11" font-weight="bold" text-anchor="middle">Multi-Head Attention</text>
<!-- Q K V arrows into attention -->
<text x="190" y="238" fill="#3498db" font-size="9" text-anchor="middle">Q</text>
<text x="250" y="238" fill="#3498db" font-size="9" text-anchor="middle">K</text>
<text x="310" y="238" fill="#3498db" font-size="9" text-anchor="middle">V</text>
<!-- Add & Norm 1 -->
<line x1="250" y1="240" x2="250" y2="218" stroke="#555" stroke-width="1.5" marker-end="url(#tf-arrow)"/>
<rect x="185" y="193" width="130" height="25" rx="6" fill="#9b59b6" opacity="0.15" stroke="#9b59b6" stroke-width="1.5"/>
<text x="250" y="210" fill="#9b59b6" font-size="10" font-weight="bold" text-anchor="middle">Add &amp; LayerNorm</text>
<!-- Residual connection 1 -->
<line x1="350" y1="258" x2="350" y2="206" stroke="#999" stroke-width="1" stroke-dasharray="4,3"/>
<line x1="350" y1="206" x2="315" y2="206" stroke="#999" stroke-width="1" marker-end="url(#tf-arrow)"/>
<text x="365" y="235" fill="#999" font-size="9">residual</text>
<!-- Feed-Forward Network -->
<line x1="250" y1="193" x2="250" y2="170" stroke="#555" stroke-width="1.5" marker-end="url(#tf-arrow)"/>
<rect x="175" y="137" width="150" height="33" rx="6" fill="#e74c3c" opacity="0.15" stroke="#e74c3c" stroke-width="1.5"/>
<text x="250" y="153" fill="#e74c3c" font-size="10" font-weight="bold" text-anchor="middle">Feed-Forward Network</text>
<text x="250" y="165" fill="#e74c3c" font-size="9" text-anchor="middle">(Linear, ReLU, Linear)</text>
<!-- Add & Norm 2 -->
<line x1="250" y1="137" x2="250" y2="118" stroke="#555" stroke-width="1.5" marker-end="url(#tf-arrow)"/>
<rect x="185" y="93" width="130" height="25" rx="6" fill="#9b59b6" opacity="0.15" stroke="#9b59b6" stroke-width="1.5"/>
<text x="250" y="110" fill="#9b59b6" font-size="10" font-weight="bold" text-anchor="middle">Add &amp; LayerNorm</text>
<!-- Residual connection 2 -->
<line x1="350" y1="153" x2="350" y2="106" stroke="#999" stroke-width="1" stroke-dasharray="4,3"/>
<line x1="350" y1="106" x2="315" y2="106" stroke="#999" stroke-width="1" marker-end="url(#tf-arrow)"/>
<text x="365" y="132" fill="#999" font-size="9">residual</text>
<!-- Output -->
<line x1="250" y1="93" x2="250" y2="68" stroke="#555" stroke-width="1.5" marker-end="url(#tf-arrow)"/>
<text x="250" y="58" fill="#666" font-size="11" text-anchor="middle">Output (to next block or head)</text>
<!-- Attention detail (right side) -->
<rect x="430" y="55" width="240" height="130" rx="6" fill="#f5f5f5" stroke="#ddd" stroke-width="1"/>
<text x="550" y="75" fill="#333" font-size="11" font-weight="bold" text-anchor="middle">Scaled Dot-Product Attention</text>
<text x="445" y="98" fill="#333" font-size="10">1. Score = Q · K^T / sqrt(d_k)</text>
<text x="445" y="118" fill="#333" font-size="10">2. Weights = softmax(Score)</text>
<text x="445" y="138" fill="#333" font-size="10">3. Output = Weights · V</text>
<text x="445" y="162" fill="#666" font-size="9">Multi-head: h parallel attentions,</text>
<text x="445" y="175" fill="#666" font-size="9">concatenated and projected</text>
<!-- Key insight box -->
<rect x="430" y="200" width="240" height="65" rx="6" fill="#f5f5f5" stroke="#ddd" stroke-width="1"/>
<text x="550" y="220" fill="#333" font-size="10" font-weight="bold" text-anchor="middle">Why Transformers Work</text>
<text x="445" y="238" fill="#666" font-size="9">Every token attends to every other</text>
<text x="445" y="252" fill="#666" font-size="9">token in parallel (no sequential</text>
<text x="445" y="266" fill="#666" font-size="9">bottleneck like RNNs)</text>
</svg>
Side by Side

After

Width:  |  Height:  |  Size: 4.4 KiB