### PyTorch 视频分类的数据集加载与训练 #### 使用 `VideoFrameDataset` 类进行视频数据集加载 为了简化视频数据集的处理过程，PyTorch 提供了一个名为 `VideoFrameDataset` 的类来帮助开发者更方便地加载和预处理视频文件。此方法不仅提高了效率还减少了复杂度[^1]。 ```python from video_dataset import VideoFrameDataset, read_video_clips # 设置路径和其他必要参数 root_path = 'path/to/video/folder' annotation_file = 'annotations.txt' dataset = VideoFrameDataset( root_path=root_path, annotationfile_path=annotation_file, num_segments=5, frames_per_segment=16, image_tmpl='img_{:05d}.jpg', transform=torchvision.transforms.Compose([ torchvision.transforms.Resize((224, 224)), torchvision.transforms.ToTensor() ]) ) data_loader = torch.utils.data.DataLoader(dataset, batch_size=4, shuffle=True) ``` 上述代码展示了如何创建一个基于帧采样的自定义视频数据集实例，并通过 DataLoader 来批量读取这些样本以便后续使用于模型训练过程中。 #### 构建简单的 CNN-LSTM 结构来进行视频分类任务 对于视频分类问题来说，通常会采用一种混合架构——即先利用卷积神经网络(CNN)提取空间特征再送入长短时记忆(LSTM)单元捕捉时间序列上的变化规律从而完成最终预测工作[^3]。 ```python import torch.nn as nn class ConvLstm(nn.Module): def __init__(self, num_classes=10): super(ConvLstm, self).__init__() # Define a convolutional network backbone self.backbone = models.resnet18(pretrained=False) feature_extractor_layers = list(self.backbone.children())[:-1] self.feature_extractor = nn.Sequential(*feature_extractor_layers) # Flatten the output of conv net and feed it into LSTM self.lstm_input_dim = 512 * 7 * 7 # Assuming input size is (224x224), adjust accordingly. self.hidden_size = 256 self.rnn = nn.LSTM(input_size=self.lstm_input_dim, hidden_size=self.hidden_size, num_layers=2, bidirectional=True) # Fully connected layer to map from RNN's last time step output to classes self.fc = nn.Linear(2*self.hidden_size, num_classes) def forward(self, x): n_frames = x.shape[1] cnn_features = [] for i in range(n_frames): frame_feature = self.feature_extractor(x[:,i,:,:,:]) cnn_features.append(frame_feature.squeeze(-1).squeeze(-1)) rnn_input = torch.stack(cnn_features,dim=0) lstm_output,(hn,cn)=self.rnn(rnn_input.permute(1,0,2)) out = self.fc(lstm_output[-1]) return F.log_softmax(out, dim=-1) ``` 这段代码实现了带有 ResNet 骨干网作为前端特征抽取器以及双向两层 LSTM 层组成的简单但有效的视频分类模型结构。 #### 开始训练过程 最后一步就是编写脚本来执行实际的学习算法迭代更新权重直至收敛： ```python device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') model = ConvLstm(num_classes=len(classes)).to(device) criterion = nn.CrossEntropyLoss().to(device) optimizer = optim.Adam(model.parameters(), lr=learning_rate) for epoch in range(num_epochs): running_loss = 0.0 for inputs, labels in data_loader: optimizer.zero_grad() outputs = model(inputs.to(device)) loss = criterion(outputs, labels.to(device)) loss.backward() optimizer.step() running_loss += loss.item() print(f'[Epoch {epoch + 1}] Loss: {running_loss / len(data_loader)}') ``` 以上便是完整的从准备数据到构建并训练一个基本版本的视频分类系统的全过程介绍。