加载模型和处理特征拆分

class Extract_Feature(object):
    def __init__(self, video_model_path = 'savevideo/netvladmodel.ckpt', audio_model_path = 'saveaudio/netvladmodelforaudio.ckpt'):
        self.video_model_path = video_model_path
        self.audio_model_path = audio_model_path
        self.videosess = ''
        self.audiosess = ''
        return
    def __del__(self):
        self.release_model()
        return
    def release_model(self):
        if self.videosess:
            with self.videosess.as_default():
                self.videosess.close()
        elif self.audiosess:
            with self.audiosess.as_default():
                self.audiosess.close()
    def set_model_path(self, video_model_path, audio_model_path):
        self.video_model_path = video_model_path
        self.audio_model_path = audio_model_path
        return

    def load_model_video(self):
        self.video_data = tf.placeholder(dtype=tf.float32, shape=[VIDEO_MAX_SAMPLES,FEATURE_SIZE_VIDEO], name='video_data')
#         tf.reset_default_graph()
        self.video_out = VladNet.network_video(self.video_data)
        self.video_out = tf.nn.softmax(self.video_out)
        self.videorestorer = tf.train.Saver()
        self.videosess = tf.Session()
        with self.videosess.as_default():
            self.videorestorer.restore(self.videosess, self.video_model_path)
        return

    def load_model_audio(self):
        self.audio_data = tf.placeholder(dtype=tf.float32, shape=[AUDIO_MAX_SAMPLES,FEATURE_SIZE_AUDIO], name='audio_data')
#         tf.reset_default_graph()
        self.audio_out = VladNet.network_audio(self.audio_data)
        self.audio_out = tf.nn.softmax(self.audio_out)
        self.audiorestorer = tf.train.Saver()
        self.audiosess = tf.Session()
        with self.audiosess.as_default():
            self.audiorestorer.restore(self.audiosess, self.audio_model_path)
        return


#注释掉这一部分是分开处理的代码
#     def videoclassify(self,video_feature):
#         feed_dict_video = {}
#         feed_dict_video[self.video_data] = video_feature
#         with self.videosess.as_default():
#             video_out = self.videosess.run(self.video_out,feed_dict=feed_dict_video)
#         return video_out
   
#     def audioclassify(self,audio_feature):
#         feed_dict_audio = {}
#         feed_dict_audio[self.audio_data] = audio_feature
#         with self.audiosess.as_default():
#             audio_out = self.audiosess.run(self.audio_out,feed_dict=feed_dict_audio)
#         return audio_out


#这里把功能合成了一个
    def classification(self,video_feature,audio_feature):
        feed_dict_video = {}
        feed_dict_audio = {}
        feed_dict_video[self.video_data] = video_feature
        feed_dict_audio[self.audio_data] = audio_feature
        with self.videosess.as_default():
            video_out = self.videosess.run(self.video_out,feed_dict=feed_dict_video) 
        with self.audiosess.as_default():
            audio_out = self.audiosess.run(self.audio_out,feed_dict=feed_dict_audio)
        fusion_out = np.mean((video_out,audio_out),axis=0)
        class_index = np.argmax(np.array(fusion_out),axis=1)
        class_index = int(class_index)
        class_name = Classes[class_index]
        return class_name,video_out,audio_out