神经网络必面之手撕 —— IOU、NMS、正向卷积、池化、BN【持续更新】

皮皮鲁与鲁西西�

已于 2022-10-12 21:47:47 修改

阅读量8.1k

点赞数 14

分类专栏：八股文章标签： python 开发语言

于 2022-03-10 18:02:41 首次发布

本文链接：https://blog.csdn.net/weixin_44398263/article/details/123407466

版权

八股专栏收录该内容

4 篇文章 6 订阅

订阅专栏

一、IOU：

在这里插入图片描述

注意：求交区域的时候，一定要和0比较大小，如果是负数就说明压根不相交

import numpy as np

def ComputeIOU(boxA, boxB):
    ## 计算相交框的坐标
    x1 = np.max([boxA[0], boxB[0]])
    x2 = np.min([boxA[2], boxB[2]])
    y1 = np.max([boxA[1], boxB[1]])
    y2 = np.min([boxA[3], boxB[3]])
    
    ## 计算交区域，并区域，及IOU
    interArea = np.max([x2-x1+1, 0])*np.max([y2-y1+1,0])	##一定要和0比较大小，如果是负数就说明压根不相交
    unionArea = (boxA[2]-boxA[0]+1)*(boxA[3]-boxA[1]+1) + (boxB[2]-boxB[0]+1)*(boxB[3]-boxB[1]+1)-interArea
    iou = interArea/unionArea
    return iou

boxA = [1,1,3,3]
boxB = [2,2,4,4]
IOU = ComputeIOU(boxA, boxB)

二、NMS：

1. 对整个bboxes排序的写法

import numpy as np

def nms(dets, iou_thred, cfd_thred):
    if len(dets)==0: return []
    bboxes = np.array(dets)
    ## 对整个bboxes排序
    bboxes = bboxes[np.argsort(bboxes[:,4])]
    pick_bboxes = []
#     print(bboxes)
    while bboxes.shape[0] and bboxes[-1,-1] >= cfd_thred:
        bbox = bboxes[-1]
        x1 = np.maximum(bbox[0], bboxes[:-1,0])
        y1 = np.maximum(bbox[1], bboxes[:-1,1])
        x2 = np.minimum(bbox[2], bboxes[:-1,2])
        y2 = np.minimum(bbox[3], bboxes[:-1,3])
        inters = np.maximum(x2-x1+1, 0) * np.maximum(y2-y1+1, 0)
        unions = (bbox[2]-bbox[0]+1)*(bbox[3]-bbox[1]+1) + (bboxes[:-1,2]-bboxes[:-1,0]+1)*(bboxes[:-1,3]-bboxes[:-1,1]+1) - inters
        ious = inters/unions
        keep_indices = np.where(ious<iou_thred)
        bboxes = bboxes[keep_indices]	## indices一定不包括自己
        pick_bboxes.append(bbox)
    return np.asarray(pick_bboxes)
    


###   肌肉记忆了
import numpy as np

def nms(preds, iou_thred=0.5, score_thred=0.5):
    ## preds: N * 5, [x1, y1, x2, y2, score]
    orders = np.argsort(preds[:,4])
    det = []
    arears = (preds[:, 2] - preds[:, 0]) * (preds[:, 3] - preds[:, 1])
    while orders.shape[0] and preds[orders[-1], 4] >= score_thred:
        pick = preds[orders[-1]]
        xx1 = np.maximum(pick[0], preds[orders[:-1], 0])
        yy1 = np.maximum(pick[1], preds[orders[:-1], 1])
        xx2 = np.minimum(pick[2], preds[orders[:-1], 2])
        yy2 = np.minimum(pick[3], preds[orders[:-1], 3])
        inters = np.maximum((xx2-xx1), 0) * np.maximum((yy2-yy1), 0)
        unions = arears[orders[-1]] + arears[orders[:-1]] - inters
        iou = inters / unions
        keep = iou < iou_thred
        orders = orders[:-1][keep]
        det.append(pick)
    return np.asarray(det)
    
dets = np.asarray([[187, 82, 337, 317, 0.9], [150, 67, 305, 282, 0.75], [246, 121, 368, 304, 0.8]])
nms(dets)
dets = [[187, 82, 337
, 317, 0.9], [150, 67, 305, 282, 0.75], [246, 121, 368, 304, 0.8]]
dets_nms =  nms(dets, 0.5, 0.3)
print(dets_nms)

2. 不改变bboxes，维护orders的写法：

始终维护orders，代表到原bboxes的映射（map）
优化1：仅维护orders，不改变原bboxes
优化2：提前计算好bboxes的面积，以免在循环中多次重复计算

import numpy as np

def nms(dets, iou_thred, cfd_thred):
    if len(dets)==0: return []
    bboxes = np.array(dets)
    ## 维护orders
    orders = np.argsort(bboxes[:,4])
    pick_bboxes = []
    x1 = bboxes[:,0]
    y1 = bboxes[:,1]
    x2 = bboxes[:,2]
    y2 = bboxes[:,3]
    areas = (x2-x1+1)*(y2-y1+1)		## 提前计算好bboxes面积，防止在循环中重复计算
    
    while orders.shape[0] and bboxes[orders[-1],-1] >= cfd_thred:
        bbox = bboxes[orders[-1]]
        xx1 = np.maximum(bbox[0], x1[orders[:-1]])
        yy1 = np.maximum(bbox[1], y1[orders[:-1]])
        xx2 = np.minimum(bbox[2], x2[orders[:-1]])
        yy2 = np.minimum(bbox[3], y2[orders[:-1]])
        inters = np.maximum(xx2-xx1+1, 0) * np.maximum(yy2-yy1+1, 0)
        unions = areas[orders[-1]] + areas[orders[:-1]] - inters
        ious = inters/unions
        keep_indices = np.where(ious<iou_thred)
        pick_bboxes.append(bbox)
        orders = orders[keep_indices]
    return np.asarray(pick_bboxes)
    

dets = [[187, 82, 337, 317, 0.9], [150, 67, 305, 282, 0.75], [246, 121, 368, 304, 0.8]]
dets_nms = nms(dets, 0.5, 0.3)
print(dets_nms)

三、正向卷积：

torch官方的Conv2d需要传入的参数

torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True)

输入和输出的特征图尺寸大小关系：
在这里插入图片描述
1.对于padding的处理是重开一个输出特征图尺寸的矩阵，然后给非padding区域赋值。或者直接用np.pad函数
2.卷积通过逐元素相乘并求和实现。使用numpy的np.multiply和np.sum函数。在inputs上逐行和逐列操作并赋值到outputs中。np.multiply可以广播，所以可以同时对多个卷积核操作，例如卷积核251633与特征图区域1633经过multiply和sum后得到251，就是输出特征图该像素点上的多通道特征。这样可以省去对各卷积核的一层循环。
3.直接利用range中的间隔模拟stride，由于总共有n-k+2p-1个有效位置，因此range的边界是n-k+2p-1。

import numpy as np

def conv2d(inputs, kernels, padding, bias, stride):
    c, w, h = inputs.shape
#     inputs_pad = np.zeros((c,w+2*padding,h+2*padding))
#     inputs_pad[:, padding:w+padding, padding:h+padding] = inputs
#     print(inputs_pad.shape, '\n', inputs_pad)
#     inputs = inputs_pad
    inputs = np.pad(inputs, ((0,0),(1,1),(1,1)))	## 可以直接用np.pad函数实现pad
    kernels_num, kernel_size = kernels.shape[0], kernels.shape[2]
    
    outputs = np.ones((kernels_num, (w-kernel_size+2*padding)//stride+1, (h-kernel_size+2*padding)//stride+1))
    for i in range(0, w-kernel_size+2*padding+1, stride):
        for j in range(0, h-kernel_size+2*padding+1, stride):
            outputs[:, i//stride, j//stride] = np.sum(np.multiply(kernels, inputs[:, i:i+kernel_size, j:j+kernel_size]), axis=(1,2,3))+bias
    return outputs

inputs = np.ones((16,9,9))
kernels = np.ones((25,16,3,3))
bias = np.arange(1,kernels.shape[0]+1)
stride = 2
padding = 1
outputs = conv2d(inputs, kernels, padding, bias, stride)
print("input{}".format(inputs.shape))
print("kenerls{}, stride{}".format(kernels.shape, stride))
print("output{}".format(outputs.shape))
print(outputs)

四、池化：

torch官方的Pool2d需要传入的参数

nn.MaxPool2d(kernel_size=2, stride=(2, 1), padding=(0, 1))

没写padding了，stride在w和h方向也没区分。。。

## 池化操作

def pooling(inputs, pool_size, stride, mode='max'):
    c, w, h = inputs.shape
    k = pool_size
    outputs = np.zeros((c,(w-k)//stride+1, (h-k)//stride+1))
    if mode == 'max':
        for i in range(0, w-k+1, stride):
            for j in range(0, h-k+1, stride):
                outputs[:, i//stride, j//stride] = np.max(inputs[:,i:i+k,j:j+k], axis=(1,2))
        return outputs
    elif mode == 'avg':
        for i in range(0, w-k+1, stride):
            for j in range(0, h-k+1, stride):
                outputs[:, i//stride, j//stride] = np.mean(inputs[:,i:i+k,j:j+k], axis=(1,2))        
        return outputs
    else:
        raise ValueError('not support this mode, choose "max" or "avg" ')
    
    
pool_size = 2
stride = 2
mode = 'max'
inputs = np.arange(1,76).reshape((3,5,5))
print("inputs:{}".format(inputs.shape), '\n',inputs)
outputs = pooling(inputs, pool_size, stride, mode)
print("outputs:{}".format(outputs.shape), '\n',outputs)

五、BN：

正向传播

在这里插入图片描述

反向传播（求导）

在这里插入图片描述

#### 手撕 BN
import torch
from torch import nn

def batch_norm(X, parameters, moving_mean, moving_var, eps, momentum):
    #### 预测模式下
    if not torch.is_grad_enable():
        X_hat = (X-moving_mean) / torch.sqrt(moving_var + eps)
        return x_hat
    ### 训练模式下
    else:
        assert len(X.shape) in (2, 4)
        #### 全连接层
        if len(X.shape) == 2:
            mean = X.mean(dim=0)
            var = ((X-mean)**2).mean(dim=0)
        ### 卷积层
        elif len(X.shape) == 4:
            mean = X.mean(dim=(0,2,3))
            var = ((X-mean)**2).mean(dim=(0,2,3))
        X_hat = (X-mean) / torch.sqrt(var + eps)
        moving_mean = momentum*moving_mean + (1-momentum)*mean
        moving_var = momentum*moving_var + (1-momentum)*moving_var
    Y = parameters['gamma'] * X_hat + parameters['beta']
    return Y, moving_mean, moving_var


class BatchNorm(nn.Module):
    def __ init__(self, num_features, num_dims):
        super.__init__()
        if num_dims == 2:
            shape = (1, num_features)
        else:
            shape = (1, num_features, 1, 1)
        self.parameters = {}
        self.parameters['gamma'] = nn.parameters(torch.ones(shape))
        self.parameters['beta'] = nn.parametersa(torch.zeros(shape))
        self.moving_mean, self.moving_var = torch.ones(shape), torch.zeros(shape)
        
    def forward(self, X):
        Y, self.moving_mean, self.moving_var =  batch_norm(X, self.parameters, self.moving_mean, self.moving_var, eps=1e-5, momentum=0.9)
        return Y

六、resnet：

在这里插入图片描述

import torch
import torch.nn as nn
import torch.nn.functional as F

class ResNetBlock(nn.Module):
    def __init__(self):
        super(ResNetBlock, self).__init__()
        self.bottleneck = nn.Sequential(
            nn.Conv2d(256, 64, 1, padding='same'),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.Conv2d(64, 64, 3, padding='same'),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.Conv2d(64, 256, 1, padding='same')  
        )
        
    def forward(self, x):
        residual = self.bottleneck(x)
        outputs = x + residual
        return outputs

resnet =  ResNetBlock()
inputs = torch.rand(4, 256, 16, 16)
outputs = resnet(inputs)
print(outputs.shape, outputs)

torch搭建简单的全连接神经网络

#### 手撕 torch神经网络
import torch
from torch import nn
import torch.nn.functional as F
from torch.autograd import Variable
import matplotlib.pyplot as plt

### 定义模型
### N,1 -> N,10 -> N,10 -> N,1
class Net(nn.Module):
    def __init__(self, n_input, n_hidden, n_output):
        super(Net, self).__init__()
        self.dense1 = nn.Linear(n_input, n_hidden)
        self.dense2 = nn.Linear(n_hidden, n_hidden)
        self.out = nn.Linear(n_hidden, n_output)
    
    def forward(self, x):
        x = self.dense1(x)
        x = F.relu(x)
        x = self.dense2(x)
        x = F.relu(x)
        x = self.out(x)
        return x
    
model = Net(1, 20, 1)
print(model)

### 准备数据
x = torch.unsqueeze(torch.linspace(-1,1,100),dim=1)
y = x.pow(3)+0.1*torch.randn(x.size())

x , y =(Variable(x),Variable(y))

plt.scatter(x.data,y.data)
# 或者采用如下的方式也可以输出x,y
# plt.scatter(x.data.numpy(),y.data.numpy())
plt.show()


####  pipeline
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss_func = torch.nn.MSELoss()

for t in range(500):
    predict = model(x)
    loss = loss_func(predict, y)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    
    if t%5 ==0:
        plt.cla()
        plt.scatter(x.data.numpy(), y.data.numpy())
        plt.plot(x.data.numpy(), predict.data.numpy(), 'r-', lw=5)
        plt.text(0.5, 0, 'Loss = %.4f' % loss.data, fontdict={'size': 20, 'color': 'red'})
        plt.pause(0.05)

定义模型使用nn.sequential()这种更简单的方式


class Net(torch.nn.Module):
    def __init__(self, n_input, n_hidden, n_output):
        super(Net4, self).__init__()
        self.dense1 = torch.nn.Sequential(
            OrderedDict(
                [
                    ("dense1", torch.nn.Linear(n_input, n_hidden),
                    ("relu1", torch.nn.ReLU()),
                ]
            ))
 
        self.dense2 = torch.nn.Sequential(
            OrderedDict([
                ("dense1", torch.nn.Linear(n_hidden, n_hidden),
                ("relu2", torch.nn.ReLU()),
            ])
        )
    	self.out = nn.Linear(n_hidden, n_output)
 
    def forward(self, x):
        x= self.dense1(x)
        x= self.dense2(x)
        x= self.out(x)
        return x

在这里插入图片描述

参考：
python实现NMS
面试问题总结——编程题关于IOU、NMS
Python手撕实现正向卷积操作
 BatchNorm的最详细解释和手撕实现与应用代码
 BatchNorm反向传播推导
 Pytorch搭建简单神经网络（一）——回归
 Pytorch之搭建神经网络的简化写法

皮皮鲁与鲁西西�

关注

14
点赞
踩
80

收藏

觉得还不错? 一键收藏
4
评论
神经网络必面之手撕 —— IOU、NMS、正向卷积、池化、BN【持续更新】

一、IOU：import numpy as npdef ComputeIOU(boxA, boxB): ## 计算相交框的坐标 x1 = np.max([boxA[0], boxB[0]]) x2 = np.min([boxA[2], boxB[2]]) y1 = np.max([boxA[1], boxB[1]]) y2 = np.min([boxA[3], boxB[3]]) ## 计算交区域，并区域，及IOU interAre
复制链接

扫一扫