视频API¶

注意

尝试在Colab上运行或转到末尾下载完整的示例代码。

这个例子展示了torchvision提供的一些用于视频的API，以及如何构建数据集等的示例。

1. 介绍：构建一个新的视频对象并检查其属性¶

首先我们选择一个视频来测试对象。为了讨论方便，我们使用的是kinetics400数据集中的一个视频。要创建它，我们需要定义路径和我们想要使用的流。

选定的视频统计信息：

WUzgd7C1pWA.mp4
- source:
  
  动力学-400
- video:
  
  H-264
  
  MPEG-4 AVC（第10部分）（avc1）
  
  帧率: 29.97
- audio:
  
  MPEG AAC 音频 (mp4a)
  
  采样率：48K Hz

import torch
import torchvision
from torchvision.datasets.utils import download_url
torchvision.set_video_backend("video_reader")

# Download the sample video
download_url(
    "https://github.com/pytorch/vision/blob/main/test/assets/videos/WUzgd7C1pWA.mp4?raw=true",
    ".",
    "WUzgd7C1pWA.mp4"
)
video_path = "./WUzgd7C1pWA.mp4"

Downloading https://raw.githubusercontent.com/pytorch/vision/refs/heads/main/test/assets/videos/WUzgd7C1pWA.mp4 to ./WUzgd7C1pWA.mp4

7%
4%
1%
7%
4%
1%
8%
5%
2%
8%
5%
2%
9%
6%
3%
9%
6%
3%
0%
7%
4%
0%
7%
4%
1%
8%
5%
0%

流的定义方式与torch设备类似。我们将它们编码为stream_type:stream_id形式的字符串，其中stream_type是字符串，stream_id是长整型。构造函数接受仅传递stream_type，在这种情况下，流会自动发现。首先，让我们获取特定视频的元数据：

stream = "video"
video = torchvision.io.VideoReader(video_path, stream)
video.get_metadata()

{'video': {'duration': [10.9109], 'fps': [29.97002997002997]}, 'audio': {'duration': [10.9], 'framerate': [48000.0]}, 'subtitles': {'duration': []}, 'cc': {'duration': []}}

在这里我们可以看到视频有两个流 - 一个视频流和一个音频流。当前可用的流类型包括 [‘video’, ‘audio’]。每个描述符由两部分组成：流类型（例如‘video’）和一个唯一的流ID （由视频编码决定）。通过这种方式，如果视频容器包含多个相同类型的流，用户可以访问他们想要的那个。如果只传递流类型，解码器会自动检测该类型的第一个流并返回它。

让我们从视频流中读取所有帧。默认情况下，next(video_reader) 的返回值是一个包含以下字段的字典。

返回的字段是：

data: 包含一个 torch.tensor
pts: 包含此特定帧的浮点时间戳

metadata = video.get_metadata()
video.set_current_stream("audio")

frames = []  # we are going to save the frames here.
ptss = []  # pts is a presentation timestamp in seconds (float) of each frame
for frame in video:
    frames.append(frame['data'])
    ptss.append(frame['pts'])

print("PTS for first five frames ", ptss[:5])
print("Total number of frames: ", len(frames))
approx_nf = metadata['audio']['duration'][0] * metadata['audio']['framerate'][0]
print("Approx total number of datapoints we can expect: ", approx_nf)
print("Read data size: ", frames[0].size(0) * len(frames))

PTS for first five frames  [0.0, 0.021332999999999998, 0.042667, 0.064, 0.08533299999999999]
Total number of frames:  511
Approx total number of datapoints we can expect:  523200.0
Read data size:  523264

但是如果我们只想读取视频的某个时间段呢？这可以通过结合我们的seek函数轻松实现，并且每次调用next都会返回返回帧的呈现时间戳（以秒为单位）。

鉴于我们的实现依赖于python迭代器，我们可以利用itertools来简化过程并使其更加符合python风格。

例如，如果我们想从第二秒开始读取十帧：

import itertools
video.set_current_stream("video")

frames = []  # we are going to save the frames here.

# We seek into a second second of the video and use islice to get 10 frames since
for frame, pts in itertools.islice(video.seek(2), 10):
    frames.append(frame)

print("Total number of frames: ", len(frames))

Total number of frames:  10

或者如果我们想从第2秒读取到第5秒，我们跳转到视频的第2秒，然后我们使用itertools的takewhile来获取正确数量的帧：

video.set_current_stream("video")
frames = []  # we are going to save the frames here.
video = video.seek(2)

for frame in itertools.takewhile(lambda x: x['pts'] <= 5, video):
    frames.append(frame['data'])

print("Total number of frames: ", len(frames))
approx_nf = (5 - 2) * video.get_metadata()['video']['fps'][0]
print("We can expect approx: ", approx_nf)
print("Tensor size: ", frames[0].size())

Total number of frames:  90
We can expect approx:  89.91008991008991
Tensor size:  torch.Size([3, 256, 340])

2. 构建一个示例 read_video 函数¶

我们可以利用上述方法来构建读取视频功能，该功能与现有的read_video函数遵循相同的API。

def example_read_video(video_object, start=0, end=None, read_video=True, read_audio=True):
    if end is None:
        end = float("inf")
    if end < start:
        raise ValueError(
            "end time should be larger than start time, got "
            f"start time={start} and end time={end}"
        )

    video_frames = torch.empty(0)
    video_pts = []
    if read_video:
        video_object.set_current_stream("video")
        frames = []
        for frame in itertools.takewhile(lambda x: x['pts'] <= end, video_object.seek(start)):
            frames.append(frame['data'])
            video_pts.append(frame['pts'])
        if len(frames) > 0:
            video_frames = torch.stack(frames, 0)

    audio_frames = torch.empty(0)
    audio_pts = []
    if read_audio:
        video_object.set_current_stream("audio")
        frames = []
        for frame in itertools.takewhile(lambda x: x['pts'] <= end, video_object.seek(start)):
            frames.append(frame['data'])
            audio_pts.append(frame['pts'])
        if len(frames) > 0:
            audio_frames = torch.cat(frames, 0)

    return video_frames, audio_frames, (video_pts, audio_pts), video_object.get_metadata()


# Total number of frames should be 327 for video and 523264 datapoints for audio
vf, af, info, meta = example_read_video(video)
print(vf.size(), af.size())

torch.Size([327, 3, 256, 340]) torch.Size([523264, 1])

3. 构建一个随机抽样的示例数据集（可应用于kinetics400的训练数据集）¶

很好，所以现在我们可以使用相同的原则来制作样本数据集。我们建议为此目的尝试可迭代数据集。在这里，我们将构建一个示例数据集，该数据集读取随机选择的10帧视频。

制作样本数据集

import os
os.makedirs("./dataset", exist_ok=True)
os.makedirs("./dataset/1", exist_ok=True)
os.makedirs("./dataset/2", exist_ok=True)

下载视频

from torchvision.datasets.utils import download_url
download_url(
    "https://github.com/pytorch/vision/blob/main/test/assets/videos/WUzgd7C1pWA.mp4?raw=true",
    "./dataset/1", "WUzgd7C1pWA.mp4"
)
download_url(
    "https://github.com/pytorch/vision/blob/main/test/assets/videos/RATRACE_wave_f_nm_np1_fr_goo_37.avi?raw=true",
    "./dataset/1",
    "RATRACE_wave_f_nm_np1_fr_goo_37.avi"
)
download_url(
    "https://github.com/pytorch/vision/blob/main/test/assets/videos/SOX5yA1l24A.mp4?raw=true",
    "./dataset/2",
    "SOX5yA1l24A.mp4"
)
download_url(
    "https://github.com/pytorch/vision/blob/main/test/assets/videos/v_SoccerJuggling_g23_c01.avi?raw=true",
    "./dataset/2",
    "v_SoccerJuggling_g23_c01.avi"
)
download_url(
    "https://github.com/pytorch/vision/blob/main/test/assets/videos/v_SoccerJuggling_g24_c01.avi?raw=true",
    "./dataset/2",
    "v_SoccerJuggling_g24_c01.avi"
)

Downloading https://raw.githubusercontent.com/pytorch/vision/refs/heads/main/test/assets/videos/WUzgd7C1pWA.mp4 to ./dataset/1/WUzgd7C1pWA.mp4

7%
4%
1%
7%
4%
1%
8%
5%
2%
8%
5%
2%
9%
6%
3%
9%
6%
3%
0%
7%
4%
0%
7%
4%
1%
8%
5%
0%
Downloading https://raw.githubusercontent.com/pytorch/vision/refs/heads/main/test/assets/videos/RATRACE_wave_f_nm_np1_fr_goo_37.avi to ./dataset/1/RATRACE_wave_f_nm_np1_fr_goo_37.avi

4%
9%
3%
7%
1%
6%
0%
4%
0%
Downloading https://raw.githubusercontent.com/pytorch/vision/refs/heads/main/test/assets/videos/SOX5yA1l24A.mp4 to ./dataset/2/SOX5yA1l24A.mp4

8%
7%
5%
4%
2%
1%
9%
8%
6%
5%
3%
2%
0%
9%
7%
6%
4%
0%
Downloading https://raw.githubusercontent.com/pytorch/vision/refs/heads/main/test/assets/videos/v_SoccerJuggling_g23_c01.avi to ./dataset/2/v_SoccerJuggling_g23_c01.avi

4%
9%
3%
8%
2%
7%
1%
6%
0%
5%
9%
3%
8%
2%
7%
0%
Downloading https://raw.githubusercontent.com/pytorch/vision/refs/heads/main/test/assets/videos/v_SoccerJuggling_g24_c01.avi to ./dataset/2/v_SoccerJuggling_g24_c01.avi

3%
5%
8%
0%
3%
6%
8%
1%
3%
6%
9%
1%
4%
6%
9%
2%
4%
7%
9%
0%

家政和公用事业

import os
import random

from torchvision.datasets.folder import make_dataset
from torchvision import transforms as t


def _find_classes(dir):
    classes = [d.name for d in os.scandir(dir) if d.is_dir()]
    classes.sort()
    class_to_idx = {cls_name: i for i, cls_name in enumerate(classes)}
    return classes, class_to_idx


def get_samples(root, extensions=(".mp4", ".avi")):
    _, class_to_idx = _find_classes(root)
    return make_dataset(root, class_to_idx, extensions=extensions)

我们将定义数据集和一些基本参数。我们假设FolderDataset的结构，并添加以下参数：

clip_len: 剪辑的长度，以帧为单位
frame_transform: 对每一帧单独进行变换
video_transform: 对视频序列进行变换

注意

我们实际上添加了epoch大小，因为使用IterableDataset()类允许我们在需要时自然地过采样每个视频的剪辑或图像。

class RandomDataset(torch.utils.data.IterableDataset):
    def __init__(self, root, epoch_size=None, frame_transform=None, video_transform=None, clip_len=16):
        super(RandomDataset).__init__()

        self.samples = get_samples(root)

        # Allow for temporal jittering
        if epoch_size is None:
            epoch_size = len(self.samples)
        self.epoch_size = epoch_size

        self.clip_len = clip_len
        self.frame_transform = frame_transform
        self.video_transform = video_transform

    def __iter__(self):
        for i in range(self.epoch_size):
            # Get random sample
            path, target = random.choice(self.samples)
            # Get video object
            vid = torchvision.io.VideoReader(path, "video")
            metadata = vid.get_metadata()
            video_frames = []  # video frame buffer

            # Seek and return frames
            max_seek = metadata["video"]['duration'][0] - (self.clip_len / metadata["video"]['fps'][0])
            start = random.uniform(0., max_seek)
            for frame in itertools.islice(vid.seek(start), self.clip_len):
                video_frames.append(self.frame_transform(frame['data']))
                current_pts = frame['pts']
            # Stack it into a tensor
            video = torch.stack(video_frames, 0)
            if self.video_transform:
                video = self.video_transform(video)
            output = {
                'path': path,
                'video': video,
                'target': target,
                'start': start,
                'end': current_pts}
            yield output

给定文件夹结构中的视频路径，例如：

dataset
- class 1
  
  文件 0
  
  文件 1
  
  …
- class 2
  
  文件 0
  
  文件 1
  
  …
- …

我们可以生成一个数据加载器并测试数据集。

transforms = [t.Resize((112, 112))]
frame_transform = t.Compose(transforms)

dataset = RandomDataset("./dataset", epoch_size=None, frame_transform=frame_transform)

from torch.utils.data import DataLoader
loader = DataLoader(dataset, batch_size=12)
data = {"video": [], 'start': [], 'end': [], 'tensorsize': []}
for batch in loader:
    for i in range(len(batch['path'])):
        data['video'].append(batch['path'][i])
        data['start'].append(batch['start'][i].item())
        data['end'].append(batch['end'][i].item())
        data['tensorsize'].append(batch['video'][i].size())
print(data)

{'video': ['./dataset/1/RATRACE_wave_f_nm_np1_fr_goo_37.avi', './dataset/2/SOX5yA1l24A.mp4', './dataset/1/WUzgd7C1pWA.mp4', './dataset/2/v_SoccerJuggling_g24_c01.avi', './dataset/2/v_SoccerJuggling_g23_c01.avi'], 'start': [0.5340898311348212, 6.514303497110079, 4.142249898746436, 7.604998914489408, 2.865074628013495], 'end': [1.066667, 7.040367, 4.671333, 8.1081, 3.370033], 'tensorsize': [torch.Size([16, 3, 112, 112]), torch.Size([16, 3, 112, 112]), torch.Size([16, 3, 112, 112]), torch.Size([16, 3, 112, 112]), torch.Size([16, 3, 112, 112])]}

4. 数据可视化¶

可视化视频示例

import matplotlib.pyplot as plt

plt.figure(figsize=(12, 12))
for i in range(16):
    plt.subplot(4, 4, i + 1)
    plt.imshow(batch["video"][0, i, ...].permute(1, 2, 0))
    plt.axis("off")

清理视频和数据集：

import os
import shutil
os.remove("./WUzgd7C1pWA.mp4")
shutil.rmtree("./dataset")

脚本总运行时间： (0 分钟 4.896 秒)

Gallery generated by Sphinx-Gallery