Created by: OleNet

import paddle.fluid as fluid
import paddle.fluid.layers as L
import numpy 
import numpy as np


mp, sp = fluid.Program(), fluid.Program()

with fluid.program_guard(mp, sp):
    
    a = fluid.layers.data("a", shape=(1,1,2,1,1,2), dtype='float32', append_batch_size=False)
    b = fluid.layers.data("b", shape=(1,1,2,1,3,2), dtype='float32', append_batch_size=False)

    c = a * b 
    fluid.layers.Print(c)
    d = L.reduce_sum(c, dim=-1, keep_dim=True)

a = np.array([0.426218,0.385291,0.299569,0.456716]).reshape([1,1,2,1,1,2,])
b = np.array([0.417022,0.720325,0.146756,0.0923386,0.000114375,0.302333,0.417022,0.720325,0.000114375,0.302333,0.146756,0.0923386,]).reshape([1,1,2,1,3,2,])

print(np.sum(a * b, axis=-1).squeeze())

place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(sp)
d = exe.run(mp,
                    feed={
                        "a": a,
                        "b": b,
                         },
                    fetch_list=[d])
print(d[0].squeeze())

输出是 numpy 计算的广播之后的reduce是：

[[0.45527702 0.09812728 0.11653493]
 [0.45391082 0.13811458 0.08613606]]

paddle 计算的广播并且reduce_sum 之后是：

[[1.77742283e-01 2.77534740e-01 6.25500488e-02]
 [3.55772315e-02 4.87486838e-05 1.16486184e-01]]

简单测试了下， paddle 的高维reduce_sum 和 reduce_mean 都有问题

1.8.0 和 1.8.2 都能复现