python - Python apply_async 不会修改 numpy 数组

Question

我正在尝试对大型 numpy 数组执行分类任务。对于上下文，阵列是卫星图像的转换。目标是将每个像素与其周围的邻居进行分类。因此，我在 50x50 补丁上使用 CNN。由于双循环滑动窗口的任务非常慢，我尝试使用多处理池。每个预测的结果都存储在一个名为“classif”的 numpy 中，其中包含每个像素的类别。我知道代码由于多次打印而有效，但它似乎没有修改数组“classif”。实际上，输出始终是一个充满零的数组。

我想我并没有真正修改“classif”，而只是它的一个副本，但我似乎无法用我在网上找到的方法来访问它。

有人可以帮我解决这个问题吗？

X_test 是我的大数组

patch_width = 50               
patch_height = patch_width
pad= 25
NUM_PROCESSES=4



img_height, img_width, band = X_test.shape


N = X_test.shape[0]
P = (NUM_PROCESSES + 1) 
partitions = list(zip(np.linspace(0, N, P, dtype=int)[:-1], np.linspace(0, N, P, dtype=int)[1:]))
work = partitions[:-1]
work.append((partitions[-1][0], partitions[-1][1]))


M = X_test.shape[1]
partitions2 = list(zip(np.linspace(0, M, P, dtype=int)[:-1],
    np.linspace(0, M, P, dtype=int)[1:]))

# Final range of indices should end +1 past last index for completeness
work2 = partitions2[:-1]
work2.append((partitions2[-1][0], partitions2[-1][1] ))


work3=list()
for i in work:
    for j in work2:
        n= list((j,i))
        work3.append(n)



X_class = np.pad(X_test, ((pad,pad), (pad,pad), (0, 0)), 'reflect')
        
shm = shared_memory.SharedMemory(create=True, size=X_class.nbytes)
X_class_shared = np.ndarray(X_class.shape, dtype=X_class.dtype, buffer=shm.buf)
X_class_shared =X_class.copy()

classif_shape = np.zeros((img_height, img_width), dtype=int)

shm2 = shared_memory.SharedMemory(create=True, size=classif_shape.nbytes)
classif = np.ndarray(classif_shape.shape, dtype=classif_shape.dtype, buffer=shm2.buf)


def predict(height, width):
    for i in range(width[0],width[1]):
        for j in range(height[0],height[1]):
            x=j 
            x_bis = j+patch_height
            y=i
            y_bis=i +patch_width 
            patch = X_class_shared[x:x_bis,y:y_bis, :]
            X = np.expand_dims(patch, axis=0)
            test = cnn.predict(X)
            rounded = np.argmax(test, axis=1)
            classif[j,i]=rounded



def dispatch_jobs(data, job_number):

    pool =Pool(job_number)
    for i, s in work3:
        pool.apply_async(predict, args=(s,i))
    print(classif)
    pool.close()
    pool.join()

 if __name__ == "__main__":
    dispatch_jobs(work3,NUM_PROCESSES)
    

标签： pythonnumpyclassificationshared-memorymultiprocess