PyNVVL

PyNVVL 是 NVIDIA 视频加载器 (NVVL) 的一个薄封装。此软件包使您能够将视频直接加载到 GPU 内存中，并以无拷贝方式将其作为 CuPy ndarrays 访问。PyNVVL 的预构建二进制文件包括 NVVL 本身，因此您不需要安装 NVVL。

需求

• CUDA 8.0, 9.0, 9.1 或 9.2
• Python 2.7.6+、3.4.7+、3.5.1+ 或 3.6.0+
• CuPy v4.5.0

测试环境

• Ubuntu 16.04
• Python 2.7.6+、3.4.7+、3.5.1+ 和 3.6.0+
• CUDA 8.0、9.0、9.1 和 9.2

安装预构建二进制文件

请根据您的 CUDA 版本选择合适的软件包。

# [For CUDA 8.0]
pip install pynvvl-cuda80

# [For CUDA 9.0]
pip install pynvvl-cuda90

# [For CUDA 9.1]
pip install pynvvl-cuda91

# [For CUDA 9.2]
pip install pynvvl-cuda92

用法

import pynvvl
import matplotlib.pyplot as plt

# Create NVVLVideoLoader object
loader = pynvvl.NVVLVideoLoader(device_id=0, log_level='error')

# Show the number of frames in the video
n_frames = loader.frame_count('examples/sample.mp4')
print('Number of frames:', n_frames)

# Load a video and return it as a CuPy array
video = loader.read_sequence(
    'examples/sample.mp4',
    horiz_flip=True,
    scale_height=512,
    scale_width=512,
    crop_y=60,
    crop_height=385,
    crop_width=512,
    scale_method='Linear',
    normalized=True
)

print(video.shape)  # => (91, 3, 385, 512): (n_frames, channels, height, width)
print(video.dtype)  # => float32

# Get the first frame as numpy array
frame = video[0].get()
frame = frame.transpose(1, 2, 0)

plt.imshow(frame)
plt.savefig('examples/sample.png')

此视频来自 Moments-In-Time 数据集的 flickr-2-6-3-3-5-2-7-6-5626335276_4.mp4。

请注意，裁剪是在缩放之后进行的。在上面的示例中，NVVL 首先从 256 x 256 缩放到 512 x 512，然后裁剪区域 [60:60 + 385, 0:512]。有关转换选项的更多信息，请参阅以下部分。

VideoLoader 选项

当创建 NVVLVideoLoader 对象时，请指定 GPU 设备 ID。您还可以通过构造函数的 log_level 参数指定日志级别。

Wrapper of NVVL VideoLoader

    Args:
        device_id (int): Specify the device id used to load a video.
        log_level (str): Logging level which should be either 'debug',
            'info', 'warn', 'error', or 'none'.
            Logs with levels >= log_level is shown. The default is 'warn'.

转换选项

pynvvl.NVVLVideoLoader.read_sequence 可以接受一些选项来指定颜色空间、值范围以及您想要对视频执行的转换。

Loads the video from disk and returns it as a CuPy ndarray.

    Args:
        filename (str): The path to the video.
        frame (int): The initial frame number of the returned sequence.
            Default is 0.
        count (int): The number of frames of the returned sequence.
            If it is None, whole frames of the video are loaded.
        channels (int): The number of color channels of the video.
            Default is 3.
        scale_height (int): The height of the scaled video.
            Note that scaling is performed before cropping.
            If it is 0 no scaling is performed. Default is 0.
        scale_width (int): The width of the scaled video.
            Note that scaling is performed before cropping.
            If it is 0, no scaling is performed. Default is 0.
        crop_x (int): Location of the crop within the scaled frame.
            Must be set such that crop_y + height <= original height.
            Default is 0.
        crop_y (int): Location of the crop within the scaled frame.
            Must be set such that crop_x + width <= original height.
            Default is 0.
        crop_height (int): The height of cropped region of the video.
            If it is None, no cropping is performed. Default is None.
        crop_width (int): The width of cropped region of the video.
            If it is None, no cropping is performed. Default is None.
        scale_method (str): Scaling method. It should be either of
            'Nearest' or 'Lienar'. Default is 'Linear'.
        horiz_flip (bool): Whether horizontal flipping is performed or not.
            Default is False.
        normalized (bool): If it is True, the values of returned video is
            normalized into [0, 1], otherwise the value range is [0, 255].
            Default is False.
        color_space (str): The color space of the values of returned video.
            It should be either 'RGB' or 'YCbCr'. Default is 'RGB'.
        chroma_up_method (str): How the chroma channels are upscaled from
            yuv 4:2:0 to 4:4:4. It should be 'Linear' currently.
        out (cupy.ndarray): Alternate output array where place the result.
            It must have the same shape and the dtype as the expected
            output, and its order must be C-contiguous.

如何构建

使用 Docker 构建 wheels

需求

• Docker
• nvidia-docker (v1/v2)

bash docker/build_wheels.sh

无 Docker 设置开发环境

脚本 setup.py 会查找必要的库。

需求：以下库在 LIBRARY_PATH 中可用。

• libnvvl.so
• libavformat.so.57
• libavfilter.so.6
• libavcodec.so.57
• libavutil.so.55

您可以在 nvvl 仓库中构建 libnvvl.so。遵循 nvvl 库的说明。 build 目录必须在 LIBRARY_PATH 中。

其他三个库在 Ubuntu 16.04 中作为软件包提供。它们安装在 /usr/lib/x86_64-linux-gnu 下，因此它们也必须在 LIBRARY_PATH 中。

python setup.py develop
python setup.py bdist_wheel