[PointNet代码详解]PointNet各模块代码实现超详细注释

pointnet.py pointnet模型各个模块的实现

import torch
import torch.nn as nn
import torch.nn.parallel
import torch.utils.data
from torch.autograd import Variable
import numpy as np
import torch.nn.functional as F
​
# STN3d: T-Net 3*3 transform
# 类似一个mini-PointNet
class STN3d(nn.Module):
    def __init__(self, channel):
        super(STN3d, self).__init__()
        # torch.nn.Conv1d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True)
        self.conv1 = torch.nn.Conv1d(channel, 64, 1) 
        self.conv2 = torch.nn.Conv1d(64, 128, 1)#channel为输入他通道数， 为3指的是输入点云特征的三个通道（X,Y,Z）,channel=6表示6个通道（X.Y.Z.和X,Y,Z方向的法向量）
        self.conv3 = torch.nn.Conv1d(128, 1024, 1)
        self.fc1 = nn.Linear(1024, 512)
        self.fc2 = nn.Linear(512, 256)
        self.fc3 = nn.Linear(256, 9) # 9=3*3
        self.relu = nn.ReLU()
​
        self.bn1 = nn.BatchNorm1d(64)#输入特征的尺度会影响梯度下降算法的迭代步数以及梯度更新的难度，从而影响训练的收敛性。因此，我们需要对特征进行归一化，即使得各个特征有相似的尺度。
        self.bn2 = nn.BatchNorm1d(128)
        self.bn3 = nn.BatchNorm1d(1024)
        self.bn4 = nn.BatchNorm1d(512)
        self.bn5 = nn.BatchNorm1d(256)
​
    def forward(self, x):#前向传播函数
        batchsize = x.size()[0]
        x = F.relu(self.bn1(self.conv1(x)))
        x = F.relu(self.bn2(self.conv2(x)))
        x = F.relu(self.bn3(self.conv3(x)))
        # Symmetric function: max poo