Python中的多处理，用于图像批处理流和处理

Question

用Python开发了一个多处理代码。 图像批处理流在一个过程中完成，而批处理在另一过程中完成。

一旦批处理流具有预定义数量的图像，请使用MultiProcess.Event（）向批处理循环发出信号。 因此，这两个过程需要在正确的时间进行。

批处理流时间比批处理时间长，因此在处理部分没有图像损失。

在大多数情况下，批处理时间比批处理流时间短。 所以我的批处理方面看起来工作正常。 但有时发现批量流重复了两次，然后进行了批量处理，例如，

batch streaming 2.35
batch processing 2.05
batch streaming 2.25
batch processing 2.05
batch streaming 2.32  repeated
batch streaming 2.36
batch processing 3.25
batch streaming 2.35
batch processing 2.15
batch streaming 2.35
batch processing 2.25

这意味着我在处理部分图像丢失。 我该如何解决这个问题？

我无法发布整个代码。 因此，两个过程如下。

批处理流循环

while (not stopbit.is_set()):
        if not cam_queue.empty():
            #print('Got frame')            
            cmd, val = cam_queue.get()

            # calculate FPS
            '''diffTime = time.time() - lastFTime
            fps = 1 / diffTime
            print(fps)
            lastFTime = time.time()'''

            # if cmd == vs.StreamCommands.RESOLUTION:
            #     pass #print(val)

            if cmd == vs.StreamCommands.FRAME:
                if val is not None:
                    missCount=0
                    image = np.array(val, dtype=np.float32, order='C')
                    image=image.transpose([2, 0, 1])
                    imgrshp=image.reshape(921600)
                    #print(str(val.shape))
                    if (batch1_is_processed == False):
                        batch1_[count] = imgrshp#it is 921,600 flat array
                        batch3_[count] = val
                    else:
                        batch2_[count] = imgrshp
                        batch4_[count] = val
                    count = count + 1
                    if (count >= BATCHSIZE):  # to start process for inference and post processing
                        diffTime = time.time() - lastFTime
                        print("batching time " + str(diffTime))
                        if (batch1_is_processed == False):  # process batch1
                            q.put('batch1')
                            batch1_is_processed = True
                            #print('batch1 is set')

                        else:  # process batch2
                            q.put('batch2')
                            batch1_is_processed = False
                            #print('batch2 is set')

                        e.set()#to signal the buffer is full
                        count = 0
                        lastFTime = time.time()
                else:
                    missCount = missCount + 1
                    print("miss frame after " + str(time.time() - startTime))
                    if(missCount >= 10):
                        q.put('lostframes')
                        e.set()#so that immediately will go back to caller to stop with lostframes option

批量处理

while(self.stopbit is not None):
                self.e.wait()
                batch = self.queue.get()
                lastFTime = time.time()
                if(batch == 'batch1'):#process batch1
                    #print('batch1 is processed')
                    for idx in range(BATCHSIZE):
                        images[idx] = np.frombuffer(self.sharedbatch1[idx], dtype=np.float32)
                        uimg = np.frombuffer(self.sharedbatch3[idx], dtype=np.uint8)
                        uimgs[idx] = uimg.reshape(HEIGHT,WIDTH,CHANNEL)
                elif(batch == 'batch2'):#process batch1
                    #print('batch2 is processed')
                    for idx in range(BATCHSIZE):
                        images[idx]=np.frombuffer(self.sharedbatch2[idx], dtype=np.float32)
                        uimg = np.frombuffer(self.sharedbatch4[idx], dtype=np.uint8)
                        uimgs[idx] = uimg.reshape(HEIGHT,WIDTH,CHANNEL)
                elif(batch == 'lostframes'):
                    self.e.clear()
                    self.stopbit.set()#to stop streaming
                    break
                #do batch processing in Nvidia's TensorRT
                with engine.create_execution_context() as context:
                    inputs, outputs, bindings, stream = common.allocate_buffers(engine)
                    inputs[0].host = np.ascontiguousarray(images, dtype=np.float32)
                    [outputs] = common.do_inference(context, bindings, inputs, outputs, stream, BATCHSIZE)
                    outputs=outputs.reshape((BATCHSIZE, 60, 80, 57))
                    humans=[]
                    for i in range(BATCHSIZE):
                       heat_map=outputs[i, :, :, :19] 
                       puf_map=outputs[i, :, :, 19:]
                       humans.append(self.est.inference(heat_map, puf_map, 4.0))

                       #uimgs[i]=TfPoseEstimatorTRT.draw_humans(uimgs[i], humans[i], imgcopy=False)
                       #cv2.imwrite("images/image_"+str(cnt)+".jpeg", uimgs[i])
                       #cnt=cnt+1 
                    hdp.ProcessHumanData(humans, uimgs)
                    #for i in range(BATCHSIZE):
                    #   cv2.imwrite("images/image_"+str(cnt)+".jpeg", uimgs[i])
                    #   cnt=cnt+1
                       #cv2.imshow('display',uimgs[i])
                       #cv2.waitKey(1)
                    humans.clear()
                    diffTime = time.time() - lastFTime
                    print("batch processing time "+str(diffTime))
                    self.e.clear()

Answer 1

使用“来自多处理程序的锁定”解决了该问题。 有时，使用打印进行调试会产生误导。 打印本身需要一些毫秒。 在调试并行处理代码时需要注意这一点。