Pytorch_ Geometric (pyg) uses dataloader to report an error runtimeerror: sizes of tenants must match except in dimension 0

Use Pytorch_Geometric（PyG） Time to build DataLoader, from DataLoader Take samples Batch Times wrong ：RuntimeError: Sizes of tensors must match except in dimension 0.

The error is caused by data alignment error ,1 individual batch Is a collection of multiple samples , An error occurred when the samples were spliced into a set , The law is as follows ：

• Use pytorch-geometric Of dataloader when ,batch Merge rules for each sample of
  • attribute edge_index Special rules , Each sample edge_index by $2\times e_i$, Then merge n Samples form a batch After that batch.edge_index The size is $2\times ({\sum }_{i=1}^n{e}_i)$
  • If all other attributes are tensor, Expand according to the first dimension , For example, for attributes $x$, The first sample size is $d_1\times d_2$, The second sample size is $d_3\times d_2$, If there is one batch These two samples are included ,batch.x The size of will be $(d_3+d_1)\times d_2$. Here is a huge pit , It requires that in addition to the first dimension , All other dimensions must be the same size ！！ Otherwise, an error will be reported RuntimeError: Sizes of tensors must match except in dimension 0.
  • Other properties if not tensor, It will return normally according to the list ,batch.x=[ sample 1 Of x, sample 2 Of x, sample 3 Of x]

How to solve ：

• If using torch tensor Caused by the , Think of ways to align other than the first dimension , Width of other dimensions .
• If there is no way to align , Use non tensor Data type substitution , For example, a list of .
• The last choice , Appoint batch_size=1 To avoid .dataloader=DataLoader(MyData,batch_size=1)