将图像文件写入tfrecords时，列表索引超出范围

编程问答 2022-06-05

问题描述

我正在为图像创建tfrecords文件，并且出现错误信息

100%|########################################################################################################################################| 116/116 [02:40<00:00,1.39s/it]
 69%|##############################################################################################1                                          | 79/115 [01:58<00:53,1.50s/it]
Traceback (most recent call last):
  File "write_tfrecords_for_inference.py",line 54,in <module>
    video_id = int(seq[0].split('_')[0])
IndexError: list index out of range

上面的代码在这里

if not os.path.isdir(tfrecord_folder):
    os.makedirs(tfrecord_folder)

data_point_names = dpc.get_data_point_names(args.data_dir,in_sequences=True,longest_seq=args.longest_seq)

splits = [[] for _ in range(args.n_divides)]
for i in range(len(data_point_names)):
    splits[i%args.n_divides].append(data_point_names[i])

for i in range(len(splits)):
    with tf.python_io.TFRecordWriter(
        os.path.join(tfrecord_folder,"cameras_%d.tfrecords" % i )) as writer:    
        
        for seq in tqdm(splits[i]):
            camera_features = list()
            feature_map_features = list()
            gazemap_features = list()
            gaze_ps_features = list()
            video_id = int(seq[0].split('_')[0])
            predicted_time_point_features = list()
            weight_features = list()
            
            for j in range(len(seq) - args.n_future_steps):
                # write camera images
                camera = cv2.imread(os.path.join(camera_folder,seq[j]+'.jpg'))              
                camera = cv2.resize(
                  camera,tuple(args.image_size[::-1]),interpolation=cv2.INTER_LINEAR
                )
                camera = camera[:,:,::-1]   # flip bgr

                camera_features.append(_bytes_feature(camera.tostring()))
                
                # write frame names
                time_point = int(seq[j+args.n_future_steps].split('_')[1])
                predicted_time_point_features.append(_int64_feature(time_point))
            
            feature_lists = {'cameras': tf.train.FeatureList(feature=camera_features),'gaze_ps': tf.train.FeatureList(feature=gaze_ps_features),'predicted_time_points': \
                               tf.train.FeatureList(feature=predicted_time_point_features),'weights': tf.train.FeatureList(feature=weight_features)}
            features = {'video_id': tf.train.Feature(int64_list=tf.train.Int64List(value=[video_id]))}
            
            example = tf.train.SequenceExample(
                context=tf.train.Features(feature=features),feature_lists=tf.train.FeatureLists(feature_list=feature_lists))
            writer.write(example.SerializetoString())

此代码适用于小型数据集并创建tfrecord文件，但在处理大型数据集时会引发上述错误。也许是numpy的内存问题，但我无法找到确切的原因。请让我知道我需要在哪里改进代码。

我正在使用numpy1.18.7和tensorflow gpu1.5。

解决方法

暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！

如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。

小编邮箱:dio#foxmail.com (将#修改为@）

numpypythontensorflowtfrecord