参考论文
3D Convolutional Neural Networks for Human Action Recognition
3D卷积针对的数据集是视频,所以3D是在2D卷积的基础上再加一个时间维度。它提取的是视频中某一帧和前后几帧的关联特征。
如上,左图是3D卷积,右图是2D卷积。
它的Pytorch编写也很简单…
nn.Conv3d(in_channels, out channels, kernel_size = (7,7,3), padding = (1,1,1)) #假设我用的是7x7x3的卷积核
行为分析说白了就是之前的图片分类任务中的图片,变成了视频而已,多了一个时间维度。所以我们需要运用3D卷积网络,去提取视频的特征。
注:在看此网络之前,建议先去了解普通的残差网络。
#残差网络的基础模块:实线连接的那个基础模块
import torch.nn as nn
import torch.nn.functional as F
class BasicBlock(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, downsample=None):
super().__init__()
self.conv1 = nn.Conv3d(in_channels, out_channels, kernel_size, stride, padding=1)
self.bn1 = nn.BatchNorm3d(out_channels)
self.conv2 = nn.Conv3d(in_channels, out_channels,kernel_size, stride, padding=1)
self.bn2 = nn.BatchNorm3d(out_channels)
self.downsample = downsample
self.stride = stride
def forward(self, x):
residual = x #残差边
feat1 = F.relu(self.bn1(self.conv1(x))) #第一层卷积
feat2 = self.bn2(self.conv2(feat1)) #第二层卷积
if self.downsample is not None:
residual = self.downsample(x)
out = F.relu(feat2+residual)
return out
最后的就是主干网络了
#使用的是resnet34残差主干网络
import torch.nn as nn
import torch
import torch.nn.functional as F
#layers = [3, 4, 6, 3],block_inplanes[64, 128, 256, 512]
class ResNet(nn.Module):
#block指的是上文提到的BasicBlock模块
def __init__(self, block, layers, block_inplanes, n_input_channels=3, conv1_t_size=7,conv1_t_stride=1,no_max_pool=False, n_classes=10):
super().__init__()
self.in_planes = block_inplanes[0]
self.no_max_pool = no_max_pool
self.conv1 = nn.Conv3d(n_input_channels,self.in_planes,kernel_size=(conv1_t_size, 7, 7), stride=(conv1_t_stride, 2, 2),padding=(conv1_t_size // 2, 3, 3),bias=False)
self.bn1 = nn.BatchNorm3d(self.in_planes)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool3d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, block_inplanes[0], layers[0])
self.layer2 = self._make_layer(block,block_inplanes[1],layers[1], stride=2)
self.layer3 = self._make_layer(block, block_inplanes[2], layers[2], stride=2)
self.layer4 = self._make_layer(block,block_inplanes[3],layers[3], stride=2)
self.avgpool = nn.AdaptiveAvgPool3d((1, 1, 1))
self.fc = nn.Linear(block_inplanes[3], n_classes)
for m in self.modules():
if isinstance(m, nn.Conv3d):
nn.init.kaiming_normal_(m.weight,
mode='fan_out',
nonlinearity='relu')
elif isinstance(m, nn.BatchNorm3d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def _downsample_basic_block(self, x, planes, stride):
out = F.avg_pool3d(x, kernel_size=1, stride=stride)
zero_pads = torch.zeros(out.size(0), planes - out.size(1), out.size(2),
out.size(3), out.size(4))
if isinstance(out.data, torch.cuda.FloatTensor):
zero_pads = zero_pads.cuda()
out = torch.cat([out.data, zero_pads], dim=1)
return out
def _make_layer(self, block, planes, blocks, stride=1):
downsample = None
if stride != 1 or self.in_planes != planes:
downsample = nn.Sequential(
nn.Conv2d(self.in_planes, planes, stride),
nn.BatchNorm3d(planes))
layers = []
layers.append(
block(in_planes=self.in_planes, planes=planes, stride=stride, downsample=downsample))
self.in_planes = planes
for i in range(1, blocks):
layers.append(block(self.in_planes, planes))
return nn.Sequential(*layers)
def forward(self, x):
x = F.relu(self.bn1(self.conv1(x)))
if not self.no_max_pool:
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avgpool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)
return x