使用 ffmpeg 作为 python 子进程将视频帧读取到 numpy 数组时发生内存泄漏

Question

我可以在 python 中将视频逐帧流式传输到 OpenGL Texture2D OK（pi3d 模块，例如 pi3d_demos/VideoWalk.py），但我注意到它逐渐泄漏内存。下面是显示问题的代码的精简版本。

有人能看到我漏水的地方吗？当 python 停止时，内存似乎会恢复。我尝试过将其明确设置为

None

或手动调用垃圾收集器。

#!/usr/bin/python
import os
import numpy as np
import subprocess
import threading
import time
import json

def get_dimensions(video_path):
    probe_cmd = f'ffprobe -v error -show_entries stream=width,height,avg_frame_rate -of json "{video_path}"'
    probe_result = subprocess.check_output(probe_cmd, shell=True, text=True)
    video_info_list = [vinfo for vinfo in json.loads(probe_result)['streams'] if 'width' in vinfo]
    if len(video_info_list) > 0:
        video_info = video_info_list[0] # use first if more than one!
        return(video_info['width'], video_info['height'])
    else:
        return None

class VideoStreamer:
    def __init__(self, video_path):
        self.flag = False # use to signal new texture
        self.kill_thread = False
        self.command = [ 'ffmpeg', '-i', video_path, '-f', 'image2pipe',
                        '-pix_fmt', 'rgb24', '-vcodec', 'rawvideo', '-']
        dimensions = get_dimensions(video_path)
        if dimensions is not None:
            (self.W, self.H) = dimensions
            self.P = 3
            self.image = np.zeros((self.H, self.W, self.P), dtype='uint8')
            self.t = threading.Thread(target=self.pipe_thread)
            self.t.start()
        else: # couldn't get dimensions for some reason - assume not able to read video
            self.W = 240
            self.H = 180
            self.P = 3
            self.image = np.zeros((self.H, self.W, self.P), dtype='uint8')
            self.t = None

    def pipe_thread(self):
        pipe = None
        while not self.kill_thread:
            st_tm = time.time()
            if pipe is None:
                pipe = subprocess.Popen(self.command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, bufsize=-1)
            self.image = np.frombuffer(pipe.stdout.read(self.H * self.W * self.P), dtype='uint8') # overwrite array
            pipe.stdout.flush() # presumably nothing else has arrived since read()
            pipe.stderr.flush() # ffmpeg sends commentary to stderr
            if len(self.image) < self.H * self.W * self.P: # end of video, reload
                pipe.terminate()
                pipe = None
            else:
                self.image.shape = (self.H, self.W, self.P)
                self.flag = True
            step = time.time() - st_tm
            time.sleep(max(0.04 - step, 0.0)) # adding fps info to ffmpeg doesn't seem to have any effect
        if pipe is not None:
            pipe.terminate()
            pipe = None

    def kill(self):
        self.kill_thread = True
        if self.t is not None:
            self.t.join()

vs = None
try:
    while True:
        for (path, _, videos) in os.walk("/home/patrick/Pictures/videos"):
            for video in videos:
                print(video)
                os.system("free") # shows gradually declining memory available
                vs = VideoStreamer(os.path.join(path, video))
                for i in range(500):
                    tries = 0
                    while not vs.flag and tries < 5:
                        time.sleep(0.001)
                        tries += 1
                    # at this point vs.image is a numpy array HxWxP bytes
                    vs.flag = False
                vs.kill()
except KeyboardInterrupt:
    if vs is not None:
        vs.kill()


os.system("free")

Answer 1

好吧，有两个因素让我感到困惑。第一个是 Wayland 在 Raspberry Pi 上（至少）的书虫上使用的

pcmanfm

中的一个错误，该错误挂在图像数据上。

第二个是由于释放流句柄的时间造成的，后者可以通过使用上下文管理器来修复（即

with Popen(...stdout=PIPE) as pipe:

可能应该总是这样做，但通常不会明确显示

subprocess.Popen

。所以如上所述，但是螺纹部分：

    def pipe_thread(self):
        while not self.kill_thread:
            with subprocess.Popen(self.command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, bufsize=-1) as pipe:
                while pipe.poll() is None and not self.kill_thread:
                    st_tm = time.time()
                    self.flag = False
                    self.image = np.frombuffer(pipe.stdout.read(self.H * self.W * self.P), dtype='uint8') # overwrite array
                    self.image.shape = (self.H, self.W, self.P)
                    self.flag = True
                    step = time.time() - st_tm
                    time.sleep(max(0.04 - step, 0.0)) # adding fps info to ffmpeg doesn't seem to have any effect