diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..be0cf26
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,2 @@
+reference/
+.venv/
diff --git a/.python-version b/.python-version
new file mode 100644
index 0000000..e4fba21
--- /dev/null
+++ b/.python-version
@@ -0,0 +1 @@
+3.12
diff --git a/README.md b/README.md
index 4600e58..205d82a 100644
--- a/README.md
+++ b/README.md
@@ -1,3 +1,24 @@
+
+
# Geo-Layout Transformer 🚀
**A Unified, Self-Supervised Foundation Model for Physical Design Analysis**
@@ -124,6 +145,22 @@ The first step is to convert your GDSII/OASIS files into a graph dataset that th
```
This script will parse the GDS file, divide it into patches, construct a graph for each patch, and save the processed data as `.pt` files for efficient loading.
+#### Polygon handling and per-patch graphs 🧩
+
+When building a graph for each patch, we now preserve both global and per-patch (clipped) polygon information to robustly handle polygons spanning multiple patches:
+
+- Each geometry retains:
+ - **Global polygon**: vertices, bbox, area.
+ - **Clipped polygon(s)** in the patch: vertices (may be multiple fragments), area, and the **area ratio** (clipped/global).
+ - **is_partial** flag indicating cross-patch polygons.
+ - **Layer index** and the **patch bbox**.
+- Node features include centroid, width/height from clipped shape (or global if no clip), clipped area, area ratio, layer id, and partial flag.
+- Extra metadata is attached on the PyG `Data` object:
+ - `data.layer: LongTensor [num_nodes]`
+ - `data.node_meta: List[Dict]` with per-node global/clipped details (for visualization/debugging)
+
+This follows the spirit of LayoutGMN’s structural encoding while staying compatible with our GNN encoder.
+
### 4.2. Stage 2: Model Training
Once the dataset is ready, you can train the Geo-Layout Transformer.
@@ -159,6 +196,17 @@ This project is ambitious and we welcome contributions. Our future roadmap inclu
Please feel free to open an issue or submit a pull request.
+## Acknowledgments
+
+We stand on the shoulders of open-source communities. This project draws inspiration and/or utilities from:
+
+- PyTorch and PyTorch Geometric for model building and graph learning
+- gdstk/klayout for GDSII/OASIS parsing and geometry operations
+- Scientific Python stack (NumPy, SciPy) for numerical robustness
+- Research works such as LayoutGMN (graph matching for structural similarity) that informed our polygon/graph handling design
+
+If your work is used and not listed here, please open an issue or PR so we can properly credit you.
+
---
Made with ❤️ for EDA research and open-source collaboration.
diff --git a/README_zh.md b/README_zh.md
index af59e63..5b399b6 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -1,3 +1,24 @@
+
+
# Geo-Layout Transformer 🚀
**一个用于物理设计分析的统一、自监督基础模型**
@@ -124,6 +145,22 @@ Geo-Layout-Transformer/
```
该脚本将解析 GDS 文件，将其划分为多个区块，为每个区块构建一个图，并将处理后的数据保存为 `.pt` 文件以便高效加载。
+#### 多边形处理与按区块建图 🧩
+
+在为每个区块（patch）构建图时，我们同时保留多边形的全局信息和区块内（裁剪后）的信息，以稳健处理跨越多个区块的多边形：
+
+- 每个几何对象包含：
+ - **全局多边形**：顶点、外接框、面积。
+ - **区块内裁剪多边形（可能多个片段）**：顶点、面积，以及 **面积占比**（裁剪/全局）。
+ - **is_partial 标记**：指示是否跨区块。
+ - **层索引** 与 **区块边界框**。
+- 节点特征包含：基于裁剪形状（若无则基于全局）的质心、宽/高、裁剪面积、面积占比、层 id、是否跨区块标志。
+- 额外元数据保存在 PyG `Data` 对象中：
+ - `data.layer: LongTensor [num_nodes]`
+ - `data.node_meta: List[Dict]`，含每个节点的全局/裁剪细节（用于可视化/调试）
+
+该设计借鉴了 LayoutGMN 的结构编码思想，同时与我们现有的 GNN 编码器保持兼容。
+
### 4.2. 阶段二：模型训练
数据集准备就绪后，您就可以开始训练 Geo-Layout Transformer。
@@ -159,6 +196,17 @@ python main.py --config-file configs/default.yaml --mode pretrain --data-dir dat
欢迎随时提出 Issue 或提交 Pull Request。
+## 致谢
+
+本项目离不开开源社区的贡献与启发，特别感谢：
+
+- PyTorch 与 PyTorch Geometric，为模型构建与图学习提供可靠基石
+- gdstk/klayout，为 GDSII/OASIS 的解析与几何操作提供高效能力
+- 科学计算生态（NumPy、SciPy），保障数值计算的稳定性
+- 研究工作 LayoutGMN（面向结构相似性的图匹配），启发了我们对多边形/图构建的设计
+
+若您的工作被本项目使用但尚未列出，欢迎提交 Issue 或 PR 以便完善致谢。
+
---
Made with ❤️ 面向 EDA 研究与开源协作。
diff --git a/pyproject.toml b/pyproject.toml
index d97c6b9..ac5cede 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,7 +1,7 @@
[project]
-name = "layouttrans"
+name = "geo-layout-transformer"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
-requires-python = ">=3.13"
+requires-python = ">=3.12"
dependencies = []
diff --git a/reference/.DS_Store b/reference/.DS_Store
new file mode 100644
index 0000000..b320be5
Binary files /dev/null and b/reference/.DS_Store differ
diff --git a/src/data/gds_parser.py b/src/data/gds_parser.py
index d6350c2..a61993b 100644
--- a/src/data/gds_parser.py
+++ b/src/data/gds_parser.py
@@ -39,30 +39,76 @@ class GDSParser:
return patches
def extract_geometries_from_patch(self, patch_bbox: Tuple[float, float, float, float]) -> List[Dict]:
- """从给定的区块中提取所有几何对象。
+ """从给定的区块中提取所有几何对象，并记录全局与区块内（裁剪后）的信息。
+
+ 说明：
+ - 为了处理跨越多个区块的多边形，本函数会计算多边形与区块边界框的布尔相交，
+ 得到位于该区块内的裁剪多边形，并同时记录原始（全局）多边形信息。
Args:
patch_bbox: 区块的边界框 (x_min, y_min, x_max, y_max)。
Returns:
- 一个字典列表，每个字典代表一个几何对象及其属性（多边形、层、边界框）。
+ 一个字典列表，每个字典代表一个几何对象及其属性：
+ - global_points: 原始多边形顶点（Nx2 ndarray）
+ - global_bbox: 原始多边形边界框
+ - global_area: 原始多边形面积
+ - clipped_points: 与区块相交后的裁剪多边形顶点（Mx2 ndarray，可能为空）
+ - clipped_area: 裁剪后面积（可能为 0）
+ - area_ratio: 裁剪面积 / 原始面积（用于衡量跨区块比例）
+ - is_partial: 是否为跨区块（裁剪面积 < 原始面积）
+ - layer: 层映射到的整数索引
+ - patch_bbox: 当前区块边界框
"""
x_min, y_min, x_max, y_max = patch_bbox
- # 获取单元内的所有多边形
+ rect = gdstk.rectangle(x_min, y_min, x_max, y_max)
polygons = self.top_cell.get_polygons(by_spec=True)
- geometries = []
- # 遍历所有多边形
+ geometries: List[Dict] = []
+
for (layer, datatype), poly_list in polygons.items():
layer_str = f"{layer}/{datatype}"
- # 只处理在 layer_mapping 中定义的层
- if layer_str in self.layer_mapping:
- for poly in poly_list:
- # 简单的边界框相交检查
- p_xmin, p_ymin, p_xmax, p_ymax = poly.bb()
- if not (p_xmax < x_min or p_xmin > x_max or p_ymax < y_min or p_ymin > y_max):
- geometries.append({
- "polygon": poly,
- "layer": self.layer_mapping[layer_str],
- "bbox": (p_xmin, p_ymin, p_xmax, p_ymax)
- })
+ if layer_str not in self.layer_mapping:
+ continue
+ layer_idx = self.layer_mapping[layer_str]
+
+ for poly in poly_list:
+ p_xmin, p_ymin, p_xmax, p_ymax = poly.bb()
+ # 快速边界框测试（若无相交则跳过）
+ if p_xmax < x_min or p_xmin > x_max or p_ymax < y_min or p_ymin > y_max:
+ continue
+
+ # 全局多边形点与面积
+ global_points = np.array(poly.points, dtype=float)
+ global_area = abs(gdstk.Polygon(global_points).area())
+
+ # 与区块矩形做相交，可能返回多个多边形
+ clipped = gdstk.boolean([poly], [rect], "and", precision=1e-3, layer=layer, datatype=datatype)
+ clipped_points_list: List[np.ndarray] = []
+ clipped_area = 0.0
+ if clipped:
+ for cpoly in clipped:
+ pts = np.array(cpoly.points, dtype=float)
+ if pts.size == 0:
+ continue
+ area = abs(gdstk.Polygon(pts).area())
+ if area <= 0:
+ continue
+ clipped_points_list.append(pts)
+ clipped_area += area
+
+ area_ratio = (clipped_area / global_area) if global_area > 0 else 0.0
+ is_partial = area_ratio < 0.999 # 允许微小数值误差
+
+ geometries.append({
+ "global_points": global_points,
+ "global_bbox": (p_xmin, p_ymin, p_xmax, p_ymax),
+ "global_area": float(global_area),
+ "clipped_points_list": clipped_points_list, # 可能包含多个裁剪片段
+ "clipped_area": float(clipped_area),
+ "area_ratio": float(area_ratio),
+ "is_partial": bool(is_partial),
+ "layer": layer_idx,
+ "patch_bbox": (x_min, y_min, x_max, y_max),
+ })
+
return geometries
diff --git a/src/data/graph_constructor.py b/src/data/graph_constructor.py
index 147229a..9a2efe5 100644
--- a/src/data/graph_constructor.py
+++ b/src/data/graph_constructor.py
@@ -1,4 +1,4 @@
-from typing import List, Dict
+from typing import List, Dict, Tuple
import torch
from torch_geometric.data import Data
from scipy.spatial import cKDTree
@@ -32,21 +32,64 @@ class GraphConstructor:
if not geometries:
return None
- node_features = []
- node_positions = []
- # 提取每个几何图形的特征
- for geo in geometries:
- x_min, y_min, x_max, y_max = geo["bbox"]
- width = x_max - x_min
- height = y_max - y_min
- area = width * height
- centroid_x = x_min + width / 2
- centroid_y = y_min + height / 2
+ node_features: List[List[float]] = []
+ node_positions: List[List[float]] = []
+ node_layers: List[int] = []
+ node_meta: List[Dict] = []
+
+ # 提取每个几何图形的特征（优先使用裁剪后片段的质心；若无裁剪片段，则使用全局质心）
+ for geo in geometries:
+ layer_idx: int = int(geo["layer"]) if "layer" in geo else 0
+ global_bbox = geo.get("global_bbox", None)
+ global_points = geo.get("global_points", None)
+ clipped_points_list = geo.get("clipped_points_list", []) or []
+ clipped_area = float(geo.get("clipped_area", 0.0))
+ global_area = float(geo.get("global_area", 0.0))
+ area_ratio = float(geo.get("area_ratio", 0.0))
+ is_partial = bool(geo.get("is_partial", False))
+
+ # 选择用于节点位置与宽高的几何：若存在裁剪片段，聚合其外接框，否则用全局框
+ if clipped_points_list:
+ # 合并所有裁剪片段点，计算整体外接框与质心
+ all_pts = np.vstack(clipped_points_list)
+ elif global_points is not None:
+ all_pts = np.array(global_points, dtype=float)
+ else:
+ # 回退到 bbox 信息（兼容旧格式）
+ x_min, y_min, x_max, y_max = geo["bbox"]
+ all_pts = np.array([[x_min, y_min], [x_max, y_max]], dtype=float)
+
+ x_min, y_min = np.min(all_pts, axis=0)
+ x_max, y_max = np.max(all_pts, axis=0)
+ width = float(x_max - x_min)
+ height = float(y_max - y_min)
+ centroid_x = float(x_min + width / 2.0)
+ centroid_y = float(y_min + height / 2.0)
+
+ # 节点特征：质心、宽、高、裁剪面积、全局面积占比、层索引（数值化）
+ features = [
+ centroid_x,
+ centroid_y,
+ width,
+ height,
+ clipped_area,
+ (clipped_area / global_area) if global_area > 0 else 0.0,
+ float(layer_idx),
+ 1.0 if is_partial else 0.0,
+ ]
- # 特征包括：中心点坐标、宽度、高度、面积
- features = [centroid_x, centroid_y, width, height, area]
node_features.append(features)
node_positions.append([centroid_x, centroid_y])
+ node_layers.append(layer_idx)
+ # 将原始与裁剪的必要元信息保存在 Data 中（以便后续可视化与调试）
+ node_meta.append({
+ "layer": layer_idx,
+ "global_bbox": tuple(global_bbox) if global_bbox is not None else None,
+ "global_area": global_area,
+ "clipped_area": clipped_area,
+ "area_ratio": area_ratio,
+ "is_partial": is_partial,
+ })
# 将特征和位置转换为 PyTorch 张量
x = torch.tensor(node_features, dtype=torch.float)
@@ -57,6 +100,9 @@ class GraphConstructor:
# 创建图数据对象
data = Data(x=x, edge_index=edge_index, pos=pos, y=torch.tensor([label], dtype=torch.float))
+ # 附加层索引与元信息（元信息以对象列表形式保存，供上层使用；不会参与张量运算）
+ data.layer = torch.tensor(node_layers, dtype=torch.long)
+ data.node_meta = node_meta
return data
def _create_edges(self, node_positions: torch.Tensor) -> torch.Tensor:
+
+ 
+ 
