记录_broadcast

记录_broadcast

code

seed = 0
n_approx = 5
num_cells = 4
dimension = 3
X = np.random.rand(n_approx, dimension)
print(X, X.shape)
kernel_points = np.random.rand(num_cells, dimension)
print(kernel_points, kernel_points.shape)

>>>
[[0.26214758 0.05170461 0.93400984]
 [0.62591246 0.65007225 0.8046075 ]
 [0.86740066 0.01002227 0.63433347]
 [0.6582109  0.76768031 0.22212046]
 [0.12551867 0.37109102 0.19793781]] (5, 3)
>>>
[[0.51271662 0.17935163 0.50461549]
 [0.52004753 0.02080544 0.33881502]
 [0.23372128 0.87948412 0.67752571]
 [0.95115178 0.11724297 0.1471212 ]] (4, 3)

Y = np.expand_dims(X, axis=1)
print(Y,Y.shape)

>>>
[[[0.26214758 0.05170461 0.93400984]]

 [[0.62591246 0.65007225 0.8046075 ]]

 [[0.86740066 0.01002227 0.63433347]]

 [[0.6582109  0.76768031 0.22212046]]

 [[0.12551867 0.37109102 0.19793781]]] (5, 1, 3)

differences = Y - kernel_points
print(differences, differences.shape)

>>>
[[[-0.25056904 -0.12764702  0.42939435]
  [-0.25789995  0.03089917  0.59519481]
  [ 0.02842629 -0.82777951  0.25648413]
  [-0.6890042  -0.06553836  0.78688864]]

 [[ 0.11319584  0.47072062  0.29999201]
  [ 0.10586493  0.62926681  0.46579248]
  [ 0.39219118 -0.22941187  0.12708179]
  [-0.32523932  0.53282928  0.65748631]]

 [[ 0.35468404 -0.16932936  0.12971798]
  [ 0.34735313 -0.01078317  0.29551844]
  [ 0.63367938 -0.86946185 -0.04319224]
  [-0.08375112 -0.1072207   0.48721227]]

 [[ 0.14549428  0.58832868 -0.28249502]
  [ 0.13816337  0.74687487 -0.11669456]
  [ 0.42448961 -0.11180381 -0.45540525]
  [-0.29294088  0.65043734  0.07499927]]

 [[-0.38719794  0.19173939 -0.30667768]
  [-0.39452886  0.35028558 -0.14087721]
  [-0.10820261 -0.50839309 -0.4795879 ]
  [-0.8256331   0.25384805  0.05081661]]] (5, 4, 3)

conclusion

• shape(5,1,3) - shape(4,3) --> shape(5,4,3) - shape(4,3)
• 直接通过broadcasting：原本 shape(5,3) 的
  • 第一行依次 减去shape(4,3)每一行
  • 第二行依次 减去shape(4,3)每一行
  • …
  • 第五行依次 减去shape(4,3)每一行
• 达到各点遍历减去中心点的效果