pytorch 自定义参数不更新方式

nn.Module中定义参数：不需要加cuda，可以求导，反向传播

class BiFPN(nn.Module):
 def __init__(self, fpn_sizes):

self.w1 = nn.Parameter(torch.rand(1))

print("no---------------------------------------------------",self.w1.data, self.w1.grad)

下面这个例子说明中间变量可能没有梯度，但是最终变量有梯度：

cy1 cd都有梯度

import torch

xP=torch.Tensor([[ 3233.8557, 3239.0657, 3243.4355, 3234.4507, 3241.7087,
    3243.7292, 3234.6826, 3237.6609, 3249.7937, 3244.8623,
    3239.5349, 3241.4626, 3251.3457, 3247.4263, 3236.4924,
    3251.5735, 3246.4731, 3242.4692, 3239.4958, 3247.7283,
    3251.7134, 3249.0237, 3247.5637],
   [ 1619.9011, 1619.7140, 1620.4883, 1620.0642, 1620.2191,
    1619.9796, 1617.6597, 1621.1522, 1621.0869, 1620.9725,
    1620.7130, 1620.6071, 1620.7437, 1621.4825, 1620.5107,
    1621.1519, 1620.8462, 1620.5944, 1619.8038, 1621.3364,
    1620.7399, 1621.1178, 1618.7080],
   [ 1619.9330, 1619.8542, 1620.5176, 1620.1167, 1620.1577,
    1620.0579, 1617.7155, 1621.1718, 1621.1338, 1620.9572,
    1620.6288, 1620.6621, 1620.7074, 1621.5305, 1620.5656,
    1621.2281, 1620.8346, 1620.6021, 1619.8228, 1621.3936,
    1620.7616, 1621.1954, 1618.7983],
   [ 1922.6078, 1922.5680, 1923.1331, 1922.6604, 1922.9589,
    1922.8818, 1920.4602, 1923.8107, 1924.0142, 1923.6907,
    1923.4465, 1923.2820, 1923.5728, 1924.4071, 1922.8853,
    1924.1107, 1923.5465, 1923.5121, 1922.4673, 1924.1871,
    1923.6248, 1923.9086, 1921.9496],
   [ 1922.5948, 1922.5311, 1923.2850, 1922.6613, 1922.9734,
    1922.9271, 1920.5950, 1923.8757, 1924.0422, 1923.7318,
    1923.4889, 1923.3296, 1923.5752, 1924.4948, 1922.9866,
    1924.1642, 1923.6427, 1923.6067, 1922.5214, 1924.2761,
    1923.6636, 1923.9481, 1921.9005]])

yP=torch.Tensor([[ 2577.7729, 2590.9868, 2600.9712, 2579.0195, 2596.3684,
    2602.2771, 2584.0305, 2584.7749, 2615.4897, 2603.3164,
    2589.8406, 2595.3486, 2621.9116, 2608.2820, 2582.9534,
    2619.2073, 2607.1233, 2597.7888, 2591.5735, 2608.9060,
    2620.8992, 2613.3511, 2614.2195],
   [ 673.7830,  693.8904,  709.2661,  675.4254,  702.4049,
     711.2085,  683.1571,  684.6160,  731.3878,  712.7546,
     692.3011,  701.0069,  740.6815,  720.4229,  681.8199,
     736.9869,  718.5508,  704.3666,  695.0511,  721.5912,
     739.6672,  728.0584,  729.3143],
   [ 673.8367,  693.9529,  709.3196,  675.5266,  702.3820,
     711.2159,  683.2151,  684.6421,  731.5291,  712.6366,
     692.1913,  701.0057,  740.6229,  720.4082,  681.8656,
     737.0168,  718.4943,  704.2719,  695.0775,  721.5616,
     739.7233,  728.1235,  729.3387],
   [ 872.9419,  891.7061,  905.8004,  874.6565,  899.2053,
     907.5082,  881.5528,  883.0028,  926.3083,  908.9742,
     890.0403,  897.8606,  934.6913,  916.0902,  880.4689,
     931.3562,  914.4233,  901.2154,  892.5759,  916.9590,
     933.9291,  923.0745,  924.4461],
   [ 872.9661,  891.7683,  905.8128,  874.6301,  899.2887,
     907.5155,  881.6916,  883.0234,  926.3242,  908.9561,
     890.0731,  897.9221,  934.7324,  916.0806,  880.4300,
     931.3933,  914.5662,  901.2715,  892.5501,  916.9894,
     933.9813,  923.0823,  924.3654]])

shape=[4000, 6000]
cx,cy1=torch.rand(1,requires_grad=True),torch.rand(1,requires_grad=True)

cd=torch.rand(1,requires_grad=True)
ox,oy=cx,cy1
print('cx:{},cy:{}'.format(id(cx),id(cy1)))
print('ox:{},oy:{}'.format(id(ox),id(oy)))
cx,cy=cx*shape[1],cy1*shape[0]
print('cx:{},cy:{}'.format(id(cx),id(cy)))
print('ox:{},oy:{}'.format(id(ox),id(oy)))
distance=torch.sqrt(torch.pow((xP-cx),2)+torch.pow((yP-cy),2))
mean=torch.mean(distance,1)
starsFC=cd*torch.pow((distance-mean[...,None]),2)
loss=torch.sum(torch.mean(starsFC,1).squeeze(),0)
loss.backward()
print(loss)
print(cx)
print(cy1)
print("cx",cx.grad)
print("cy",cy1.grad)
print("cd",cd.grad)
print(ox.grad)
print(oy.grad)
print('cx:{},cy:{}'.format(id(cx),id(cy)))
print('ox:{},oy:{}'.format(id(ox),id(oy)))

来源：https://blog.csdn.net/jacke121/article/details/103672674