python - Keras 模型无法预测是否在线程中调用

Question

我尝试在线程应用程序中使用 keras 和可用模型 VGG16 执行预测。但是，如果我在主线程中调用预测，一切正常。但是，如果我在线程函数内部进行预测(无论我使用 threading、multiprocessing、...)，它只会在预测期间停止:

这是最小的例子:

########################################
# Alter this variable
USE_THREADING = True
########################################

import numpy as np
import cv2
import copy
import threading
import keras
import platform
import tensorflow as tf
from keras.models import model_from_json
from multiprocessing import Process

def inference_handler(model_hash, frame_resized):
    print("multiprocessing: before prediction call")
    model_hash.predict(np.expand_dims(frame_resized, axis=0), batch_size = 1)
    print("multiprocessing: after prediction call")

if __name__ == "__main__":
    print("keras version:", keras.__version__)
    print("tf vresion: ", tf.__version__)
    print("python version:", platform.python_version())
    model_hash = keras.applications.VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3))
    # Perform the demo
    cap = cv2.VideoCapture(0)
    while(True):
        # Capture frame-by-frame
        ret, frame = cap.read()
        # Process the keys
        key = cv2.waitKey(1) & 0xFF
        if key == ord('q'):
            print("quit")
            break
        # Get the proper image for the network
        frame_resized = cv2.resize(frame, (224, 224))
        # show the images
        cv2.imshow('frame',frame)
        cv2.imshow('frame_resized',frame_resized)

        # Predict
        if USE_THREADING:
            p = Process(target=inference_handler, args=(model_hash, frame_resized,))
            p.start()
            p.join()
        else:
            print("main thread: before prediction call")
            model_hash.predict(np.expand_dims(frame_resized, axis=0), batch_size = 1)
            print("main thread: after prediction call")


    # When everything done, release the capture
    cap.release()
    cv2.destroyAllWindows()

USE_THREADING = False 给我:

Using TensorFlow backend.
keras version: 2.2.0
tf vresion:  1.8.0
python version: 3.5.2
2019-02-25 20:47:32.926696: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
main thread: before prediction call
main thread: after prediction call
main thread: before prediction call
main thread: after prediction call
main thread: before prediction call
...

USE_THREADING = True(失败)给我:

Using TensorFlow backend.
keras version: 2.2.0
tf vresion:  1.8.0
python version: 3.5.2
2019-02-25 20:50:34.922696: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
multiprocessing: before prediction call

Answer 1

因此，如果将模型作为子流程的参数给出，那么具有 tensorflow 后端的 Keras 不幸地存在预测期间停止的问题。但是，如果直接在子流程中创建模型，则一切正常。因此，解决方案是将帧通过队列发送到子进程。这是一个可行的解决方案:

import numpy as np
import cv2
import copy
import keras
import platform
import tensorflow as tf
from keras.models import model_from_json
from multiprocessing import Process, Queue

def inference_handler(frame_queue):
    model_hash = keras.applications.VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3))
    model_hash._make_predict_function()
    while True:
        print("multiprocessing: before queue")
        frame_resized = frame_queue.get(block=True, timeout=None)
        print("multiprocessing: before prediction call")
        model_hash.predict(np.expand_dims(frame_resized, axis=0), batch_size = 1)
        print("multiprocessing: after prediction call")

if __name__ == "__main__":
    print("keras version:", keras.__version__)
    print("tf version: ", tf.__version__)
    print("python version:", platform.python_version())
    frame_queue = Queue(maxsize=1)
    p = Process(target=inference_handler, args=(frame_queue,))
    p.start()
    # p.join()
    cap = cv2.VideoCapture(0)
    while(True):
        # Capture frame-by-frame
        ret, frame = cap.read()
        # Process the keys
        key = cv2.waitKey(1) & 0xFF
        if key == ord('q'):
            print("quit")
            break
        # Get the proper image for the network
        frame_resized = cv2.resize(frame, (224, 224))
        # show the images
        cv2.imshow('frame',frame)
        cv2.imshow('frame_resized',frame_resized)

        # Advertise the frame
        if frame_queue.empty():
            print("Put frame into the queue")
            frame_queue.put_nowait(frame_resized)

    # When everything done, release the capture
    p.terminate()
    cap.release()
    cv2.destroyAllWindows()

这给了我

keras version: 2.2.0
tf version:  1.8.0
python version: 3.5.2
Put frame into the queue
multiprocessing: before queue
multiprocessing: before prediction call
Put frame into the queue
multiprocessing: after prediction call
multiprocessing: before queue
multiprocessing: before prediction call
Put frame into the queue
...