[teamai] Push 87 resource(s) from XingfenD

2026-09-10 16:10:45 +08:00
parent 425c9c078a
commit 65c04def51
1314 changed files with 211681 additions and 0 deletions
@@ -0,0 +1,219 @@
+# 神经网络架构图模板
+
+本文件用于生成"论文中那种神经网络架构图"：
+
+- Transformer / Encoder-Decoder 架构图
+- U-Net / FPN / 多尺度网络架构
+- GAN / Diffusion / VAE 架构
+- Attention 机制示意
+- 自定义模型架构图
+
+特征：
+
+- 多个 layer 块按数据流方向排布（横向或竖向）
+- 每个 layer 块有：层名 + tensor shape 标注（H × W × C）
+- 跳连 / residual / attention 连线清晰
+- 颜色编码不同 layer 类型（Conv / Attention / FC / Norm）
+- 出版物级，白底克制
+
+## 适用范围
+
+- 论文中的 model architecture figure
+- 综述论文 framework
+- 答辩 PPT 模型介绍页
+- 教学 slide 中的网络示意
+
+## 何时使用
+
+- 用户提到 "网络架构 / network architecture / model architecture / Transformer / U-Net / GAN / Diffusion / VAE"
+- 用户希望「层级清晰、tensor shape 标准、跳连一目了然」
+- 用户希望视觉「论文风、白底、彩色编码 layer 类型」
+
+不要使用：
+
+- 用户要的是「方法 pipeline 总览」（多 stage 业务流）→ 用 `academic-figures/method-pipeline-overview.md`
+- 用户要的是「系统架构图」（前端 + 后端 + DB）→ 用 `technical-diagrams/system-architecture.md`
+- 用户要的是「数据流向 / ER 图」 → 用 `technical-diagrams/er-diagram.md`
+- 用户要的是「概念示意 / 注意力可视化」（自由度高）→ 用 `academic-figures/scientific-schematic.md`
+
+## 缺失信息优先提问顺序
+
+1. 模型类型（Encoder-Decoder / U-Net / Transformer / GAN / Diffusion / 自定义）
+2. 主干网络层数 / 每层类型（如「6 层 Transformer encoder + 6 层 decoder + 8 头 attention」）
+3. Tensor shape（输入分辨率 / 通道数 / 序列长度）
+4. 是否有跳连 / residual / cross-attention
+5. 是否有 multi-task / multi-head 输出
+6. 是否要中文标签（论文图通常英文）
+7. 比例（横向 16:9 / 2:1，符合论文双栏）
+
+## 主模板：Transformer / Encoder-Decoder 架构图
+
+📖 描述
+
+整张图横向流动：左输入 embedding → 多层 encoder 块 → cross-attention → 多层 decoder 块 → 右输出 head。每个 layer 块标注层类型与 tensor shape，跳连用弧形虚线。
+
+📝 提示词
+
+```json
+{
+  "type": "神经网络架构图（neural network architecture diagram）",
+  "goal": "生成论文级别的网络架构图：层级清晰、tensor shape 标注、跳连分明、可单色印刷可读",
+  "canvas": {
+    "aspect_ratio": "{argument name=\"aspect_ratio\" default=\"16:9\"}",
+    "background": "white #FFFFFF",
+    "outer_padding": "60px"
+  },
+  "model_meta": {
+    "name": "{argument name=\"model_name\" default=\"Our Transformer\"}",
+    "input_spec": "{argument name=\"input_spec\" default=\"Input Image: 224×224×3\"}",
+    "output_spec": "{argument name=\"output_spec\" default=\"Class Logits: 1000\"}"
+  },
+  "layer_groups": {
+    "rule": "use color-coded blocks per layer type — keep palette ≤ 5 muted academic colors",
+    "color_legend": [
+      { "type": "Embedding / PatchEmbed", "fill": "#E0E7FF", "border": "#6366F1" },
+      { "type": "Self-Attention", "fill": "#FEE2E2", "border": "#DC2626" },
+      { "type": "Cross-Attention", "fill": "#FEF3C7", "border": "#D97706" },
+      { "type": "Feed Forward / MLP", "fill": "#D1FAE5", "border": "#059669" },
+      { "type": "Norm / Residual", "fill": "#F3F4F6", "border": "#6B7280" }
+    ]
+  },
+  "layers": {
+    "count": "{argument name=\"layer_count\" default=\"8\"}",
+    "items": [
+      { "id": "L1", "type": "Embedding / PatchEmbed", "name": "Patch Embed", "shape": "196×768" },
+      { "id": "L2", "type": "Norm / Residual", "name": "LayerNorm", "shape": "196×768" },
+      { "id": "L3", "type": "Self-Attention", "name": "Multi-head Self-Attn (×8)", "shape": "196×768", "annotation": "× N=6 (encoder)" },
+      { "id": "L4", "type": "Feed Forward / MLP", "name": "FFN", "shape": "196×768" },
+      { "id": "L5", "type": "Cross-Attention", "name": "Cross-Attn", "shape": "K×768" },
+      { "id": "L6", "type": "Self-Attention", "name": "Decoder Self-Attn", "shape": "K×768", "annotation": "× N=6 (decoder)" },
+      { "id": "L7", "type": "Feed Forward / MLP", "name": "FFN", "shape": "K×768" },
+      { "id": "L8", "type": "Norm / Residual", "name": "Output Head (Linear)", "shape": "K×C" }
+    ]
+  },
+  "block_style": {
+    "shape": "rounded rectangle (corner radius 4-6px)",
+    "size_rule": "blocks of same layer type share identical width and height; visually grouped",
+    "border": "1.2px solid (use the type's border color)",
+    "fill": "use the type's fill color (very light tint)",
+    "label_text": "layer name on first line (sans-serif bold 10-11pt) + tensor shape on second line (monospace italic 9pt)",
+    "annotation_text": "if 'annotation' present (e.g. '× N=6'), draw it as a curly brace with label on the right side of the repeated block"
+  },
+  "connections": {
+    "main_flow": {
+      "style": "thin black solid arrows (1.2px), horizontal left → right",
+      "arrowhead": "small filled triangle"
+    },
+    "residual": {
+      "enabled": "{argument name=\"residual_enabled\" default=\"true\"}",
+      "style": "curved dashed arrow arcing above the main flow, label '+' near join",
+      "rule": "draw residual from input of attention block to its output"
+    },
+    "cross_attention": {
+      "enabled": "{argument name=\"cross_attention_enabled\" default=\"true\"}",
+      "style": "horizontal arrow from encoder side feeding into decoder cross-attn, label 'K, V'",
+      "rule": "encoder output is shown as K, V input to decoder cross-attn"
+    }
+  },
+  "extras": {
+    "show_param_count": {
+      "enabled": "{argument name=\"show_params\" default=\"false\"}",
+      "rule": "if true, add parameter count below each major group (e.g. '85M params')"
+    },
+    "highlight_novelty": {
+      "enabled": "{argument name=\"highlight_novelty\" default=\"true\"}",
+      "rule": "if true, surround the user's contributed module with a thicker dashed orange border + label 'Ours' / 'Novel'"
+    }
+  },
+  "constraints": {
+    "must_keep": [
+      "tensor shapes are accurate and labeled in monospace font",
+      "color encodes layer type consistently across the figure",
+      "all layers of the same type have identical block size",
+      "white background, no gradient, no decoration",
+      "all labels in English by default (or all Chinese if explicitly requested), no mixing",
+      "must remain readable when printed in grayscale (rely on shape and label, not color alone)",
+      "novel contribution (if any) is clearly marked"
+    ],
+    "avoid": [
+      "3D extruded blocks, drop shadows, glossy fills",
+      "rainbow palette (>5 colors)",
+      "cartoon icons, emoji",
+      "freeform 'art-style' blobs instead of crisp rectangles",
+      "tensor shapes typeset in proportional font",
+      "arrows crossing through blocks",
+      "missing tensor shape labels (the figure is then useless for paper review)",
+      "unlabeled cross-attention (must say K, V)"
+    ]
+  }
+}
+```
+
+### 参数策略
+
+- **必问**：`layer_count`、每层的 `type` 和 `shape`
+- **可默认**：`aspect_ratio`（16:9）、`background`（白）、`color_legend`（默认 5 类配色）、`block_style`
+- **可随机**：blocks 内每行的精确字号 / padding，annotation 摆放位置
+
+### 自动补全策略
+
+- 用户给「我用 Transformer」但没给细节 → 反问关键参数（层数、头数、隐藏维度、序列长度）；不要瞎编模型规模
+- 用户给「U-Net」 → 自动用 contracting + expansive 双臂布局变体（见变体 2）
+- 用户没说有没有 residual → 默认 `residual_enabled: true`（绝大多数现代网络都有）
+- 用户没说有没有 novelty → 默认 `highlight_novelty: true`（论文图一般要标自己的贡献）
+- 用户没说参数量 → 默认 `show_params: false`（除非用户提到模型规模对比）
+
+## 变体 1：U-Net / FPN 双臂架构
+
+```json
+{
+  "type": "U-Net / FPN 双臂架构图",
+  "modify": {
+    "layout": "U 形：左臂下采样（contracting path）+ 中央 bottleneck + 右臂上采样（expansive path），每层之间有水平 skip connection",
+    "annotation": "skip 用横向虚线箭头标注，特征图用渐窄 / 渐宽的矩形示意 spatial 维度变化"
+  }
+}
+```
+
+适用：U-Net、FPN、HRNet、所有 encoder-decoder 分割网络。
+
+## 变体 2：GAN / Diffusion 双网络对抗 / 多步推理
+
+```json
+{
+  "type": "GAN / Diffusion 架构图",
+  "modify": {
+    "layout_gan": "上方 Generator（noise → image）+ 下方 Discriminator（image → real/fake），中间共享生成图像作为 D 的输入",
+    "layout_diffusion": "横向 timestep 序列 t=T → t=0，每个 timestep 是同一个 U-Net 实例，标 't' 嵌入条件"
+  }
+}
+```
+
+适用：GAN 系列、扩散模型、Score-based 模型。
+
+## 变体 3：Multi-task / Multi-head 输出
+
+```json
+{
+  "type": "多任务 / 多头输出架构图",
+  "modify": {
+    "layout": "共享 backbone 在中央 → 右侧分叉成 2-4 个 task head（如 classification head / regression head / segmentation head）",
+    "annotation": "每个 head 旁边标对应 loss 函数和权重 λ"
+  }
+}
+```
+
+适用：多任务学习、检测 + 分割、辅助监督。
+
+## 避免事项
+
+- tensor shape 缺失或随便写 → 论文图核心信息没了
+- 用渐变 / 3D 立方体堆叠 → 像 PPT 不像论文
+- 颜色 ≥ 6 种 → 失去 layer 类型语义
+- 没有 residual / cross-attention 标注（如果架构里有）→ 误导读者
+- 用 Comic Sans / 手写字体
+- 跳连箭头穿过 layer 块
+- 同一类 layer 块大小不一致
+- 中英文标签混用
+- 把"训练 loss"画进结构图（应该单独一张 training figure 或 caption 里说明）
+- 在结构图里塞具体超参数表（应该走 table，不进 figure）