解決多TF模型和多Keras模型同時使用，graph問題。

前言

項目需要，使用tensorflow和keras混合編程。因爲不熟悉tensorflow和keras，導致報錯BUG，修復兩天，沒有好的解決方法。

修改過程

其實在早期，有修復過類似問題（1個TF模型加1個KERAS），在keras外包一層tf.Graph()完美解決衝突。但是這次更加複雜。需要調用N個TF模型，N個KERAS模型。如果我new多次keras模型，導致graph無法畫出。
後來在stackoverflow上找到較好的版本。地址：名字爲Andrew Louw的小哥
舊代碼：

class model(object):
    graph = None

    def __init__(self, model_path=modepath):
        # config = tf.ConfigProto()
        # config.gpu_options.allow_growth = True
        # session = tf.Session(config=config)
        # KTF.set_session(session)
        self.graph = tf.Graph()
        sess = tf.Session(graph=self.graph)
        K.set_session(sess)

新代碼：

import os
from keras import backend as K
import tensorflow as tf
import logging
from keras.models import load_model
import cv2
import numpy as np
import matplotlib.image as mpimg
import keras.backend.tensorflow_backend as KTF


#如果前面是tf代碼不需要加上這幾句。防止keras佔用全部顯卡
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
session = tf.Session(config=config)
KTF.set_session(session)




class Keras_Model:
    def __init__(self,file_name):
        self.session = tf.Session()
        self.graph = tf.get_default_graph()
        self.model = None
        with self.graph.as_default():
            with self.session.as_default():
                logging.info("neural network initialised")
                #self.model = load_model(file_name, custom_objects={'R': self.R, 'P': self.P, 'F': self.F})
    			self.model = load_model(file_name)
    def R(self,y_true, y_pred):
            true_positives = K.sum(K.cast(K.greater(K.clip(y_true * y_pred, 0, 1), 0.20), 'float32'))
            poss_positives = K.sum(K.cast(K.greater(K.clip(y_true, 0, 1), 0.20), 'float32'))
            recall = true_positives / (poss_positives + K.epsilon())
            return recall
            # precision
    def P(self,y_true, y_pred):
            true_positives = K.sum(K.cast(K.greater(K.clip(y_true * y_pred, 0, 1), 0.20), 'float32'))
            pred_positives = K.sum(K.cast(K.greater(K.clip(y_pred, 0, 1), 0.20), 'float32'))
            precision = true_positives / (pred_positives + K.epsilon())
            return precision
    def F(self,y_true, y_pred):
            p_val = self.P(y_true, y_pred)
            r_val = self.R(y_true, y_pred)
            f_val = 2 * p_val * r_val / (p_val + r_val)
            return f_val
    def detect_img(self, path):
        def pre_process_img(path, length=300, width=300, channel=3):
            #img = cv2.imread(path)
            img = path
            img = cv2.resize(img, (length, width), interpolation=cv2.INTER_CUBIC)
            img = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
            img = img.reshape([1, length, width, 1])
            return img

        with self.graph.as_default():
            with self.session.as_default():
                y = self.model.predict(pre_process_img(path))
                y = y[0].argmax()
                #data = np.where(y[:, :1] > y[:, 1:], 0, 1)#輸出標籤。flag可改。0是ok，1是ng
        return y

if __name__ == '__main__':
    a = Keras_Model('_v1.h5')
    pic = '24.bmp'
    print(a.predict(pic))