[torch]t.sparse.mm&t.spmm的使用与区别

代码：

a = torch.randn(2, 3)
a1 =a.to_sparse().requires_grad_(True)
a2 = torch.randn(5,2,3).to_sparse()
a3 = a.to_sparse().requires_grad_(False)
b = torch.randn(3, 2)
b1 =b.to_sparse().requires_grad_(True)
b2 = torch.randn(5,3,32).to_sparse()
b3 = b.to_sparse().requires_grad_(False)

y1 = torch.sparse.mm(a,b)   # 两个dense矩阵
y2 = torch.sparse.mm(a1,b1) # 两个sparse矩阵
y3 = torch.sparse.mm(a1, b) #sparse,dense
#y4 = torch.sparse.mm(a, b1) # dense, sparse --不成立

# z0 = torch.spmm(a1, b) #sparse-有梯度.dense --不成立
z1 = torch.spmm(a3, b) #sparse-无梯度,dense  ok
z2 = torch.spmm(a, b)  #dense,dense ok
z3 =torch.spmm(a1, b1) #sparse-有梯度,sparse-有梯度        ---不成立
#z4 =torch.spmm(a, b1)  #dense,sparse-有梯度  ---不成立
#z5 = torch.spmm(a, b3) # dense,sparse -无梯度  ----不成立

结论：
torch.sparse.mm

• dense $\times$ dense
• sparse $times$ sparse
• sparse $times$ dense

torch.spmm

• dense $times$ dense
• sparse $times$ dense