16 - nn.Conv2d的原理以及三路分支残差block的算子融合实现
文章目录1. R-drop1.1 R-drop 原理2. 常见函数2.1 torch常见API1. R-drop论文链接:R-Drop: Regularized Dropout for Neural Networks1.1 R-drop 原理2. 常见函数2.1 torch常见APItorch.functional.padimport torchfrom torch.nn import funct
1. R-drop
1.1 R-drop 原理
- 算法:
- 损失函数:
- 贡献
2. 常见函数
2.1 torch常见API
- torch.functional.pad
import torch
from torch.nn import functional as F
a = torch.ones(2, 3, 4, 5)
paddings_1 = (1, 2, 3, 4, 5, 6, 7, 8)
paddings_2 = (10, 9, 8, 7, 6, 5, 4, 3)
b = F.pad(a, paddings_1)
c = F.pad(a,paddings_2)
print(f"a.shape={a.shape}")
print(f"b.shape={b.shape}")
print(f"c.shape={c.shape}")
a.shape=torch.Size([2, 3, 4, 5])
b.shape=torch.Size([17, 14, 11, 8])
c.shape=torch.Size([9, 14, 19, 24])
- torch.nn.Conv2d (padding=“same”)
以前我们设置二维卷积的时候需要手动设置padding 的值,特别的麻烦,现在方便了,直接设置padding=“same”,那么可以保证,输入张量经过卷积运算后,输出矩阵的高宽前后保持一致;
import torch
from torch import nn
my_conv2d = nn.Conv2d(in_channels=3,out_channels=5,kernel_size=3,padding="same")
my_input = torch.rand(2,3,10,20)
my_output = my_conv2d(my_input)
print(f"my_conv2d={my_conv2d}")
print(f"my_input.shape={my_input.shape}")
print(f"my_output.shape={my_output.shape}")
my_conv2d=Conv2d(3, 5, kernel_size=(3, 3), stride=(1, 1), padding=same)
my_input.shape=torch.Size([2, 3, 10, 20])
my_output.shape=torch.Size([2, 5, 10, 20])
- torch.isclose()
torch.isclose(input, other, rtol=1e-05, atol=1e-08, equal_nan=False) → Tensor
返回一个新的张量,其布尔元素表示输入的每个元素是否“接近”other的相应元素。亲密度定义为:
如果张量a和张量b中的数据的差值在一定的范围内,那么就返回True,否则返回False
print(torch.isclose(torch.Tensor([1.,2.,3.]),torch.Tensor([1.+1e-10,3.,3.])))
# tensor([ True, False, True])
- torch.all(input)
torch.all(input):
测试输入中的所有元素的值是否为True。如果所有的值为True,则返回True,否则返回False
# import library
import torch
# define the tensor a , b
a = torch.Tensor([1,0,2,0])
b = torch.Tensor([1,2,3,4,5])
# convert the values into boolean type
b_bool = b.bool()
a_bool = a.bool()
# if all the values in tensor a_all is true ,it return the final answer for True
# if just including the one False value in tensor a_all,it return the final answer for False
a_all = torch.all(a_bool)
b_all = torch.all(b_bool)
print(f"a={a}")
print(f"a_bool={a_bool}")
print(f"a_all={a_all}")
print("*"*50)
print(f"b={b}")
print(f"b_bool={b_bool}")
print(f"b_all={b_all}")
a=tensor([1., 0., 2., 0.])
a_bool=tensor([ True, False, True, False])
a_all=False
**************************************************
b=tensor([1., 2., 3., 4., 5.])
b_bool=tensor([True, True, True, True, True])
b_all=True
- torch.equal()&torch.eq(a,b)
torch.eq(a,b):
计算张量a和b中每个对应位置的值是否相等,并返回每个位置的True 和 False
torch.equal(a,b):
比较张量a,b在形状和值是否相等,返回整体的True 和False
# import the library
import torch
# define the variable tensor a and b
a = torch.Tensor([[1,2,3],[4,5,6]])
b = torch.Tensor([[1,4,2],[2,3,6]])
# compare the values of the each elements in tensor a and b is whether equal
c_eq = torch.eq(a, b)
# just compare the tensor a and b is whether equal
d_equal = torch.equal(a,b)
print(f"a={a}")
print(f"b={b}")
print(f"c_eq={c_eq}")
print(f"d_equal={d_equal}")
a=tensor([[1., 2., 3.],
[4., 5., 6.]])
b=tensor([[1., 4., 2.],
[2., 3., 6.]])
c_eq=tensor([[ True, False, False],
[False, False, True]])
d_equal=False
3. Resnet 算子融合
3.1 目标
将Resnet模块中
r
e
s
u
l
t
1
=
c
o
n
v
3
×
3
+
c
o
n
v
1
×
1
+
x
result_1=conv3\times3+conv1\times1+x
result1=conv3×3+conv1×1+x
转换成:
r
e
s
u
l
t
2
=
c
o
n
v
3
×
3
+
c
o
n
v
3
×
3
+
c
o
n
v
3
×
3
result_2=conv3\times3+conv3\times3+conv3\times3
result2=conv3×3+conv3×3+conv3×3
最后我们在使用的时候用一个3×3卷积表示即可得到result1的同等效果,达到算子融合的目的
3.2 步骤
- 将1×1卷积改造成3×3卷积
conv_2d_for_pointwise
- 将x本身改造成3×3卷积
conv_2d_for_identity
- 创建一个新的融合3×3卷积
conv_2d_for_fusion
- 将
conv_2d_for_pointwise
,conv_2d_for_identity
和本来就有的3×3卷积conv2d
权重和偏置相加后赋值给conv_2d_for_fusion
conv_2d_for_fusion.weight = conv2d.weight + conv_2d_for_pointwise.weight + conv_2d_for_identity.weight
conv_2d_for_fusion.bias = conv2d.bias + conv_2d_for_pointwise.bias + conv_2d_for_identity.bias
- 这样我们就得到一个新的卷积
conv_2d_for_fusion
;它满足了resnet的所有需求,而且更快
3.3 代码
import torch
from torch import nn
import torch.nn.functional as F
import time
in_channels = 2
ou_channels = 2
kernel_size = 3
w = 9
h = 9
t1 = time.time()
# 方法1:原生写法
x = torch.ones(1, in_channels, w, h)
conv_2d = nn.Conv2d(in_channels, ou_channels, kernel_size, padding="same")
conv_2d_pointwise = nn.Conv2d(in_channels, ou_channels, 1)
result1 = conv_2d(x) + conv_2d_pointwise(x) + x
t2 = time.time()
# 方法2:算子融合
# 把 point_wise 卷积核 x 本身都携程 3*3 的卷积
# 最终把三个卷积写成一个卷积实现算子融合
# step 1: 将 1x1的卷积转换成 3x3 的卷积
# conv_2d_pointwise.weight.shape = (2,2,1,1) 通过填充变成 (2,2,3,3)
pointwise_to_conv_weight = F.pad(conv_2d_pointwise.weight, [1, 1, 1, 1, 0, 0, 0, 0])
conv_2d_for_pointwise = nn.Conv2d(in_channels, ou_channels, kernel_size, padding="same")
conv_2d_for_pointwise.weight = nn.Parameter(pointwise_to_conv_weight)
conv_2d_for_pointwise.bias = nn.Parameter(conv_2d_pointwise.bias)
# step 2: 将 x 本身转换成 3x3 的卷积
# x 经过 3x3 的卷积conv_2d_for_identity还是能够保证值不变
# 为了保证 x 在转换过程中的值不变,需要满足如下两个条件
# 1. 3x3的卷积不能在像素与像素之间具有关联性
# 需要一个 3x3 矩阵,中心为1,其他为0
# 2. 3x3的卷积不能在通道与通道之间具有关联性
# 需要通道上面[中间1矩阵,全0矩阵,全0矩阵,中间1矩阵];这样通道间就没关联了
# 卷积的形状为:(2,2,3,3);我们需要定义卷积的weights和bias来实现自定义卷积
#
conv_2d_for_identity = nn.Conv2d(in_channels, ou_channels, kernel_size, padding="same")
zeros = torch.unsqueeze(torch.zeros(kernel_size, kernel_size), 0)
stars = torch.unsqueeze(F.pad(torch.ones((1, 1)), [1, 1, 1, 1]), 0)
zeros_stars = torch.unsqueeze(torch.cat([zeros, stars], 0), 0)
stars_zeros = torch.unsqueeze(torch.cat([stars, zeros], 0), 0)
conv_2d_for_identity_weight = torch.cat([zeros_stars, stars_zeros], dim=0)
conv_2d_for_identity_bias = torch.zeros([ou_channels])
conv_2d_for_identity.weight = nn.Parameter(conv_2d_for_identity_weight)
conv_2d_for_identity.bias = nn.Parameter(conv_2d_for_identity_bias)
result2 = conv_2d(x) + conv_2d_for_pointwise(x) + conv_2d_for_identity(x)
# result2 = conv_2d_for_pointwise(x)
# result2 = conv_2d_for_identity(x)
# print(f"torch.all(torch.isclose(result1,result2))\t=\t{torch.all(torch.isclose(result1,result2))}")
# print(f"zeros={zeros}")
# print(f"zeros.shape={zeros.shape}")
# print(f"stars.shape={stars.shape}")
# print(f"stars={stars}")
# print(f"zeros_stars={zeros_stars}")
# print(f"zeros_stars.shape={zeros_stars.shape}")
# step 3: 将改造后的矩阵conv_2d(x);conv_2d_for_pointwise(x);conv_2d_for_identity(x)进行融合
conv_2d_for_fusion = nn.Conv2d(in_channels, ou_channels, kernel_size, padding="same")
conv_2d_for_fusion.weight = nn.Parameter(conv_2d.weight.data + conv_2d_for_pointwise.weight.data + conv_2d_for_identity.weight.data)
conv_2d_for_fusion.bias = nn.Parameter(conv_2d.bias.data + conv_2d_for_pointwise.bias.data + conv_2d_for_identity.bias.data)
result3 = conv_2d_for_fusion(x)
t3 = time.time()
# print(f"torch.all(torch.isclose(result1,result3))\t=\t{torch.all(torch.isclose(result1,result3))}")
print(f"原生卷积 time = {1000*(t2-t1)}ms")
print(f"融合卷积 time = {1000*(t3-t2)}ms")
结果:
原生卷积 time = 1.9557476043701172ms
融合卷积 time = 0.9968280792236328ms
3.4 小结
多次实验,我们在不改变值的情况下,我们发现,经过算子融合的网络即使在python这样的语法下,也能在运行速度上提高2倍;其主要思想还是像地铁思想样,我们从A点到B点,如果有三个人甲,乙,丙 ;他们如果单独进行运动,那么时间就会长,如果我们将甲乙丙三个人打包到地铁上,让他们搭载地铁,那么他们一起到达B点的时间就会缩短。
更多推荐
所有评论(0)