行为识别（三）：创建C3D网络模型用于行为识别

1 参考资源

1.1Pytorch-video-recognition

1、链接：https://github.com/jfzhang95/pytorch-video-recognition

2、推荐理由：包含对UCF101、HMDB51的处理，包含经典的C3D、R3D、R2+1D网络模型、包含完整的预处理函数，包含训练好的权重

1.2 Tensorflow小练习（三）：C3D完成视频动作识别

1、链接：https://zhuanlan.zhihu.com/p/32934943

2、推荐理由：C3D网络Tensorflow实现，包含代码

1.3 行为识别/视频分类|知识汇总

1、链接：https://zhuanlan.zhihu.com/p/374671906

1.4 Paper with code

1、链接：https://paperswithcode.com/task/action-recognition-in-videos

2 本文采用的C3D模型代码

model.py

from torch import nn


class C3D(nn.Module):
    def __init__(self):
        super(C3D, self).__init__()
        """
            input (n_batch, n_channel, n_frame, h, w) = (n_batch, 3, 16, 240, 320)
        """
        self.conv1 = nn.Sequential(
            nn.Conv3d(3, 64, kernel_size=(3, 3, 3), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(1, 2, 2), stride=(1, 2, 2)),

            nn.Conv3d(64, 128, kernel_size=(3, 3, 3), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2)),

            nn.Conv3d(128, 256, kernel_size=(3, 3, 3), padding=(1, 1, 1)),
            nn.Conv3d(256, 256, kernel_size=(3, 3, 3), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2)),

            nn.Conv3d(256, 512, kernel_size=(3, 3, 3), padding=(1, 1, 1)),
            nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2)),

            nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1, 1, 1)),
            nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2), padding=(0, 1, 1))
        )
        self.linear = nn.Sequential(
            nn.Flatten(),
            nn.Linear(512*1*8*11,4096),
            nn.Linear(4096,4096),
            nn.Linear(4096,101),
            nn.ReLU()
            # nn.CrossEntropyLoss() realize the softmax layer
            # nn.Softmax(dim=1)
        )

    def forward(self, x):
        x = self.conv1(x)
        x = self.linear(x)
        return x

3 训练所用代码

CPU版本，修改代码以在GPU上训练

train.py

import torch
from torch.utils.tensorboard import SummaryWriter

from model import *
from UCF101_dataset import *

train_set_root = "T:/Dataset/UCF101/UCF101_frames/train/"
test_set_root = "T:/Dataset/UCF101/UCF101_frames/test/"
train_set = UCF101(train_set_root)
test_set = UCF101(test_set_root)
batch_size = 4
train_data_loader = DataLoader(train_set, shuffle=True, drop_last=True, batch_size=batch_size)
test_data_loader = DataLoader(test_set, shuffle=True, drop_last=True, batch_size=batch_size)

model_c3d = C3D()
loss_function = nn.CrossEntropyLoss()
learning_rate = 1e-2
optimizer = torch.optim.Adam(model_c3d.parameters(),lr=learning_rate)

total_train_step = 0
total_test_clip = 0
correct = 0
accuracy = 0.0
writer = SummaryWriter("logs")
for epoch in range(0,20):
    print("epoch: {} ".format(epoch))
    model_c3d.train()
    for clips_fold,labels in train_data_loader:
        clips = train_set.unfold_batch(clips_fold)
        onehot = train_set.onehot(labels)
        out = model_c3d(clips)
        loss = loss_function(onehot,out)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        total_train_step += 1
        if (total_train_step % 100 == 0):
            print("total_train_step: {}, loss: {}".format(total_train_step,loss))
            writer.add_scalar("loss",loss,total_train_step)

    torch.save(model_c3d,"C3D_epoch_{:5d}.pkl".format(epoch))
    print("model saved as C3D_epoch_{:5d}.pkl".format(epoch))

    model_c3d.eval()
    with torch.no_grad():
        correct = 0
        total_test_clip = 0
        for test_clips_fold,test_labels in test_data_loader:
            test_clips = test_set.unfold_batch(test_clips_fold)
            test_onehot = test_set.onehot(test_labels)
            test_out = model_c3d(test_clips)

            total_test_clip += batch_size
            correct += (torch.argmax(test_out,dim=1) == torch.argmax(test_onehot,dim=1)).sum().item()

        accuracy = correct / total_test_clip
        writer.add_scalar("accuracy",accuracy,epoch)
        print("epoch: {}, Accuracy: {}".format(epoch,accuracy))