python 如何对两个数组做差处理

Posted 2023-05-08

参考技术A 转成集合，通过集合的求差方法，然后将结果再转成list 参考技术B Python中的列表中的元素不能直接相加减。
最佳的方式是将列表转换成Python中的科学计算包numpy包的array类型，再进行加减。

import numpy as npa = np.array([1,2,3,4])b = np.array([7,8,9,10])s = a + b

如何对大型 3d 图像堆栈使用多处理？ Python

【中文标题】如何对大型 3d 图像堆栈使用多处理？ Python

【英文标题】：How to use multiprocessing for a big 3d image stack? python

【发布时间】：2021-09-02 08:54:04

【问题描述】：

我有一个 3d 图像堆栈（4000×2048×2048），我想在每个 2d 数组（2048×2048）中做一些操作，例如。高斯滤波、图像增强、resize img ...

import numpy as np
from tifffile import imread,imwrite
import multiprocessing as mp
import cv2

def gaussian_blur_2d(img):
    blur = cv2.GaussianBlur(img,(5,5),0) 
    return blur

file_path = "F:\\Ctest\\123.tif"
img = imread(file_path)
for i in range(0,img.shape[0]):
    img[i,:,:] = gaussian_blur_2d(img[i,:,:])

如何通过使用多处理来加速 for 循环？ 我的想法是将原始图像堆栈分成四个或八个部分，并使用pool.map 到拆分堆栈。但是我如何使用拆分处理结果来获得最终的完整堆栈。我不想写拆分堆栈.这会增加额外的 IO 时间。根据我的经验，当拆分堆栈太大时，会导致pool.map 中的返回错误。

另一方面，我尝试将多维数组粘贴到 mp.Array 中，这给了我 TypeError: only size-1 数组可以转换为 Python 标量。

【问题讨论】：

真的都适合内存吗？ （约 16 gigs，8 位灰度）如果是这样，我会在每个进程中从 multiprocessing.shared_memory 对象构造 numpy 数组。你的大问题几乎肯定是有效地将数据传入和传出子进程。跨度>

【参考方案1】：

正如我在 cmets 中提到的，让多个工作进程之间可访问所有数据是这里最大的挑战，因为使用多处理的关键租户之一是进程之间通常不共享内存。因此，我们必须明确告诉操作系统我们想要访问进程之间“共享”的一块内存，并使用该内存块创建我们的 numpy 数组。除此之外，它只是一些非常标准的多处理内务处理，并且在其他教程和示例中得到了很好的探索。

import numpy as np
from multiprocessing import Process, shared_memory, Queue, cpu_count
from queue import Empty
import cv2

class STOPFLAG: pass #a simple flag to tell the worker to stop

def worker_process(in_q, shm_name):
    shm = shared_memory.SharedMemory(name=shm_name) #create from the existing one made by the parent process
    img_stack = np.ndarray([4000, 2048, 2048], dtype="uint8", buffer=shm.buf) #attach a numpy array to the memory object
    while True: #until the worker runs out of work
        try:
            task = in_q.get(1) #don't wait forever on anything if you can help it.
        except Empty: #multiprocessing.Queue uses an exception template from the queue library
            print("assuming all tasks are done. worker exiting...") #assume waiting for a while means no more tasks (we shouldn't hit this, but it could prevent problems in the child if a crash happens elsewhere)
            break
        if isinstance(task, STOPFLAG):
            print("got stop flag. worker exiting...")
            break
        
        #process the image slice (no mutexes are needed because no two workers will ever get the same index to work on at the same time)
        img_stack[task] = cv2.GaussianBlur(img_stack[task],(5,5),0) 
        
    shm.close() #cleanup after yourself (close the local copy. This does not close the copy in the other processes)

if __name__ == "__main__": #this is needed with multiprocessing

    #create shared memory space where numpy will work from
    shm = shared_memory.SharedMemory(create=True, size=4000*2048*2048) #OS may have a hard time allocating this memory block because it's so big...
    #create the numpy array from the allocated memory
    img_stack = np.ndarray([4000, 2048, 2048], dtype="uint8", buffer=shm.buf)
    
    #Here is where you would load the image data onto the img_stack array. It will start out with whatever random data was previously in ram similar to numpy.empty.
    
    #create a queue to send workers tasks (image index to work on)
    in_q = Queue()
    
    #create a couple worker processes
    processes = [Process(target=worker_process, args = (in_q, shm.name)) for _ in range(cpu_count())]
    for p in processes:
        p.start()
    
    #fill up the task queue with image indices that need computation
    for i in range(4000):
        in_q.put(i)
        
    #send a stop signal for each worker
    for _ in processes:
        in_q.put(STOPFLAG())
        
    #wait for all children to finish
    for p in processes:
        p.join()
        
    #do something (save?) with the img_stack
    np.save("processed_images.npy", img_stack)
    
    shm.close() #cleanup
    shm.unlink() #unlink is called only once after the last instance has been "close()"d

【讨论】：

你的回答让我震惊。太奇妙了！作为一个菜鸟，我学到了很多东西，我衷心感谢你对我的好意。