torch_geometric 使用 batch

__momo__

已于 2024-01-02 10:59:41 修改

阅读量1.6k

点赞数 9

分类专栏： PyTorch # Python-Error Handling # ML-GNN 文章标签： python 深度学习图神经网络

于 2023-12-11 21:30:23 首次发布

本文链接：https://blog.csdn.net/yaoyao_chen/article/details/134936884

版权

PyTorch 同时被 3 个专栏收录

10 篇文章 0 订阅

订阅专栏

Python-Error Handling

9 篇文章 0 订阅

订阅专栏

ML-GNN

2 篇文章 0 订阅

订阅专栏

基于GCNConv, SAGEConv

问题描述

基于torch.geometric的data具有node和edge的特征，其中node的维度表示为[num_nodes, node features]，edge的维度表示为[2, edge_features]。

torch.geometric的batch不支持CNN中广泛使用的batch维度。CNN中使用batch的数据具有[batch_size, channel, h, w]的维度。如果使用[batch_size, num_nodes, node_features]和[batch_size, 2, edge_features]，代码会有如下报错：

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size XX but got size XXX for tensor number 1 in the list.

解决方案：

解决方案

完整问题

更多细节

在对角线上重复edge_index特征

batch_edge_index = Batch.from_data_list([Data(edge_index=edge_index)]*batch_size)

使用node_dim

conv = GCNConv(in_channels, out_channels, node_dim=1)
conv(x,edge_index)
# 在这种情况下，x（node）的维度是[batch_size, num_nodes, num_features]
# edge_index的维度是(2, edge_features)
# edge_index holds indices < num_nodes

edge_index always need to be two-dimensional, so the mini-batching of [batch_size, num_nodes, num_features] node features is designed to operate on the same underlying graph of shape [2, num_edges]. If your graph is not static across examples, you will have to use PyG’s approach of diagonal adjacency matrix stacking.

⚠️值得注意的是 edge_index中起始点和终点的index都必须小于nodes的个数（因为默认从0开始）

基于 GATConv

问题描述

与GCNConv, SAGEConv不同的是，当使用

x（node）的维度是[batch_size, num_nodes, num_features]
edge_index的维度是(2, edge_features)

会遇到如下报错：Static graphs not supported in GATConv。

解决方案

解决方案，可以使用Batch.from_data_list(data_list)将数据转换成Batch格式，或者手动将node concate成[batchsize*num_nodes, num_features]，edge_index concate成[2, batch_size*edge_features]。
具体维度解析：

edge_index_s = torch.tensor([
    [0, 0, 0, 0],
    [1, 2, 3, 4],
])
x_s = torch.randn(5, 16)  # 5 nodes.
edge_index_t = torch.tensor([
    [0, 0, 0],
    [1, 2, 3],
])
x_t = torch.randn(4, 16)  # 4 nodes.

edge_index_3 = torch.tensor([[0, 1, 1, 2],
                           [1, 0, 2, 1]], dtype=torch.long)
x_3 = torch.randn(4, 16)

data1= Data(x=x_s,edge_index=edge_index_s)
data2= Data(x=x_t,edge_index=edge_index_t)
data3= Data(x=x_3,edge_index=edge_index_3)
#上面是构建3张Data图对象
# * `Batch(Data)` in case `Data` objects are batched together
#* `Batch(HeteroData)` in case `HeteroData` objects are batched together

data_list = [data1, data2,data3]


loader = Batch.from_data_list(data_list)#调用该函数data_list里的data1、data2、data3 三张图形成一张大图，也就是batch
————————————————
版权声明：本文为CSDN博主「知行合一，至于至善」的原创文章，遵循CC 4.0 BY-SA版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/qq_41800917/article/details/120444534