批次矩阵乘法 等价于一种特殊的 Conv1D

conv1d-bmm.py

"""
表明：批次矩阵乘法 等价于一种特殊的 Conv1D
"""
import torch

N = 2
Cin = 3
Lin = 4
Cout = 5

# Use conv1d
x = torch.rand(size=(N, Cin, Lin))
w = torch.rand(size=(Cout, Cin, 1))
y0 = torch.nn.functional.conv1d(x, w)
print(y0)

# Use batched matrix multiplication
w_batch = w.reshape(1, Cout, Cin)
y1 = torch.matmul(w_batch, x)
print(y1)

# Print mean error
print(torch.mean(torch.abs(y0 - y1)))

一个 kernel size 为 1 的 Conv1D，实际上就是一个批次矩阵乘法。反之亦然。