Add files via upload

large/arxiv.bat

@@ -0,0 +1,7 @@
+@echo off
+echo Running ogbn-arxiv...
+python main.py --method sgformer --dataset ogbn-arxiv --metric acc --lr 0.001 --hidden_channels 256 --use_graph --graph_weight 0.5 ^
+--gnn_num_layers 3 --gnn_dropout 0.5 --gnn_weight_decay 0.0 --gnn_use_residual --gnn_use_weight --gnn_use_bn --gnn_use_act ^
+--trans_num_layers 1 --trans_dropout 0.5 --trans_weight_decay 0.0 --trans_use_residual --trans_use_weight --trans_use_bn ^
+--seed 123 --runs 5 --epochs 1000 --eval_step 9 --device 0 --data_dir D:\ogb_dataset\ogb\arxiv
+pause

large/build_graph.py

@@ -0,0 +1,126 @@
+import numpy as np
+import torch
+from torch import long
+from sklearn.metrics import pairwise_distances
+from sklearn.neighbors import NearestNeighbors
+from sklearn.decomposition import PCA
+from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
+from sklearn.preprocessing import StandardScaler
+
+
+def build_graph_by_pos(height: long, width: long, node_features=None):
+    """
+    简单的八向建图 8*height*width
+    """
+    edge_index = []
+    for i in range(height):
+        for j in range(width):
+            node_idx = i * width + j
+            if i > 0:  # up
+                edge_index.append([node_idx, node_idx - width])
+            if i < height - 1:  # down
+                edge_index.append([node_idx, node_idx + width])
+            if j > 0:  # left
+                edge_index.append([node_idx, node_idx - 1])
+            if j < width - 1:  # right
+                edge_index.append([node_idx, node_idx + 1])
+            if i > 0 and j > 0:  # left and up
+                edge_index.append([node_idx, node_idx - width - 1])
+            if i > 0 and j < width - 1:  # right and up
+                edge_index.append([node_idx, node_idx - width + 1])
+            if i < height - 1 and j > 0:  # left and down
+                edge_index.append([node_idx, node_idx + width - 1])
+            if i < height - 1 and j < width - 1:  # right and down
+                edge_index.append([node_idx, node_idx + width + 1])
+    edge_index = torch.tensor(edge_index, dtype=torch.long).t()
+    return edge_index
+
+def build_graph_by_Knn(height: long, width: long, node_features=None, k=10):
+    num_nodes = height * width
+    edge_index = []
+    nbrs = NearestNeighbors(n_neighbors=k, algorithm='auto').fit(node_features)
+    distances, indices = nbrs.kneighbors(node_features)
+
+    # 遍历每个节点的近邻
+    for i in range(num_nodes):
+        for j in range(1, k):  # 从1开始，跳过自己
+            neighbor_index = indices[i][j]
+            edge_index.append((i, neighbor_index))
+
+    # 将边列表转换为 NumPy 数组
+    edge_index = torch.tensor(edge_index, dtype=torch.long).T  # 转置为 [2, num_edges]
+
+    return edge_index
+
+
+def get_weight(node_features, edge_index, sigma=10):
+    """根据节点特征和边索引计算边权重，并返回张量形式的权重"""
+    num_edges = edge_index.shape[1]  # 边的数量
+    edge_weights = torch.zeros(num_edges, dtype=torch.float)
+
+    def gaussian_kernel(x, y, sig=1.0):
+        return torch.exp(-torch.norm(x - y) ** 2 / (sig ** 2))
+
+    # 遍历每条边
+    for i in range(num_edges):
+        start_node = edge_index[0, i]  # 从节点
+        end_node = edge_index[1, i]  # 到节点
+
+        # 提取节点特征
+        x_i = node_features[start_node]  # 从节点特征
+        x_j = node_features[end_node]  # 到节点特征
+
+        # 计算高斯核权重
+        edge_weights[i] = gaussian_kernel(x_i, x_j, sigma)
+
+    return edge_weights
+
+
+def build_graph_by_fix(node_fea, ground_truth, train_idx, row, col):
+    """
+    先对数据进行LDA降维之后 对降维光谱特征+位置特征进行knn建图
+    主要是加入pos 但是单纯的加入pos对于欧式距离的衡量改变不大，所以先降维成channel再增加pos信息
+    这样pos信息所占比重就能上升，以此查看与单纯的220个波段的knn之间的区别
+    """
+    edge_index = []
+    k = 15  # 邻居数量
+    weight_factor = 15  # 权重系数
+
+    x = node_fea[train_idx]
+    y = ground_truth[train_idx]
+    lda = LinearDiscriminantAnalysis()
+    # 调用fit方法　对训练集像素点的特征和标签进行拟合
+    lda.fit(x, y - 1)
+    x_new = lda.transform(node_fea)
+
+    all_indices = np.arange(x_new.shape[0])  # 计算 x_new 中所有节点的索引
+    x_coords = all_indices // col  # 计算 x 坐标
+    y_coords = all_indices % col  # 计算 y 坐标
+
+    # 将坐标信息添加到 x_new 中
+    coords = np.column_stack((x_coords, y_coords))  # 合并 x 和 y 坐标
+    x_new_with_coords = np.hstack((x_new, coords))  # 将坐标与 x_new 合并
+
+    # 进行最小-最大标准化
+    x_min = np.min(x_new_with_coords, axis=0)  # 每列的最小值
+    x_max = np.max(x_new_with_coords, axis=0)  # 每列的最大值
+
+    # 标准化
+    x_normalized = (x_new_with_coords - x_min) / (x_max - x_min)
+
+    x_weighted = x_normalized.copy()
+    x_weighted[:, -2:] *= weight_factor  # 增加最后两个维度的权重
+
+    knn = NearestNeighbors(n_neighbors=k, algorithm='auto').fit(x_weighted)
+    distances, indices = knn.kneighbors(x_weighted)
+
+    # 遍历每个节点的近邻
+    for i in range(row * col):
+        for j in range(1, k):  # 从1开始，跳过自己
+            neighbor_index = indices[i][j]
+            edge_index.append((i, neighbor_index))
+
+    # 将边列表转换为 NumPy 数组
+    edge_index = torch.tensor(edge_index, dtype=torch.long).T  # 转置为 [2, num_edges]
+
+    return edge_index