使用自定义COCO数据集+TensorFlow数据生成器训练U-Net报错求助

基于自定义COCO数据集训练U-Net模型时，使用自定义数据生成器函数触发了TensorFlow图执行错误。尝试调整输入/输出形状等方式排查多次无效，推测是逻辑问题，请求协助定位解决。

数据生成器函数 #

def dataGeneratorCoco(images, classes, coco, folder, 
                      input_image_size=(224,224), batch_size=4, mode='train', mask_type='binary'):
    
    img_folder = '{}/{}-data/images'.format(folder,mode)
    dataset_size = len(images)
    catIds = coco.getCatIds(catNms=classes)
    
    
    c = 0
    while(True):
        img = np.zeros((batch_size, input_image_size[0], input_image_size[1], 3)).astype('float')
        mask = np.zeros((batch_size, input_image_size[0], input_image_size[1], 1)).astype('float')

        for i in range(c, c+batch_size): #initially from 0 to batch_size, when c = 0
            imageObj = images[i]
            
            ### Retrieve Image ###
            train_img = getImage(imageObj, img_folder, input_image_size)
                        
            ### Create Mask ###
            if mask_type=="binary":
                train_mask = getBinaryMask(imageObj, coco, catIds, input_image_size)
            
            elif mask_type=="normal":
                train_mask = getNormalMask(imageObj, classes, coco, catIds, input_image_size)                
            
            # Add to respective batch sized arrays
            # print(img.shape,train_img.shape)
            img[i-c] = train_img
            mask[i-c] = train_mask
            
        c+=batch_size
        if(c + batch_size >= dataset_size):
            c=0
            random.shuffle(images)
        yield img, mask

数据增强生成器 #

def augmentationsGenerator(gen, augGeneratorArgs, seed=None):
    # Initialize the image data generator with args provided
    image_gen = ImageDataGenerator(**augGeneratorArgs)
    
    # Remove the brightness argument for the mask. Spatial arguments similar to image.
    augGeneratorArgs_mask = augGeneratorArgs.copy()
    _ = augGeneratorArgs_mask.pop('brightness_range', None)
    # Initialize the mask data generator with modified args
    mask_gen = ImageDataGenerator(**augGeneratorArgs_mask)
    
    np.random.seed(seed if seed is not None else np.random.choice(range(9999)))
    

    while(True):
        for img, mask in gen:
            seed = np.random.choice(range(9999))
            # keep the seeds syncronized otherwise the augmentation of the images 
            # will end up different from the augmentation of the masks
            g_x = image_gen.flow(255*img, 
                                 batch_size = img.shape[0], 
                                 seed = seed, 
                                 shuffle=True)
            g_y = mask_gen.flow(mask, 
                                 batch_size = mask.shape[0], 
                                 seed = seed, 
                                 shuffle=True)

            img_aug = next(g_x)/255.0
            mask_aug = next(g_y)

            yield img_aug, mask_aug

UnknownError: Graph execution error:

2 root error(s) found.
  (0) UNKNOWN:  Exception: input type is not supported.
Traceback (most recent call last):

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/tensorflow/python/ops/script_ops.py", line 270, in __call__
    ret = func(*args)

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/tensorflow/python/autograph/impl/api.py", line 642, in wrapper
    return func(*args, **kwargs)

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/tensorflow/python/data/ops/dataset_ops.py", line 1030, in generator_py_func
    values = next(generator_state.get_iterator(iterator_id))

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/keras/engine/data_adapter.py", line 831, in wrapped_generator
    for data in generator_fn():

  File "/tmp/ipykernel_284939/3839959985.py", line 15, in augmentationsGenerator
    for img, mask in gen:

  File "/tmp/ipykernel_284939/728327123.py", line 25, in dataGeneratorCoco
    train_mask = getNormalMask(imageObj, classes, coco, catIds, input_image_size)

  File "/tmp/ipykernel_284939/2567587607.py", line 29, in getNormalMask
    new_mask = cv2.resize(coco.annToMask(anns[a])*pixel_value, input_image_size)

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/pycocotools/coco.py", line 442, in annToMask
    rle = self.annToRLE(ann)

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/pycocotools/coco.py", line 427, in annToRLE
    rles = maskUtils.frPyObjects(segm, h, w)

  File "pycocotools/_mask.pyx", line 308, in pycocotools._mask.frPyObjects

Exception: input type is not supported.


     [[{{node PyFunc}}]]
     [[IteratorGetNext]]
     [[Shape/_10]]
  (1) UNKNOWN:  Exception: input type is not supported.
Traceback (most recent call last):

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/tensorflow/python/ops/script_ops.py", line 270, in __call__
    ret = func(*args)

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/tensorflow/python/autograph/impl/api.py", line 642, in wrapper
    return func(*args, **kwargs)

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/tensorflow/python/data/ops/dataset_ops.py", line 1030, in generator_py_func
    values = next(generator_state.get_iterator(iterator_id))

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/keras/engine/data_adapter.py", line 831, in wrapped_generator
    for data in generator_fn():

  File "/tmp/ipykernel_284939/3839959985.py", line 15, in augmentationsGenerator
    for img, mask in gen:

  File "/tmp/ipykernel_284939/728327123.py", line 25, in dataGeneratorCoco
    train_mask = getNormalMask(imageObj, classes, coco, catIds, input_image_size)

  File "/tmp/ipykernel_284939/2567587607.py", line 29, in getNormalMask
    new_mask = cv2.resize(coco.annToMask(anns[a])*pixel_value, input_image_size)

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/pycocotools/coco.py", line 442, in annToMask
    rle = self.annToRLE(ann)

  File "/home/navneeth/anaconda3/lib/python3.9/site-packages/pycocotools/coco.py", line 427, in annToRLE
    rles = maskUtils.frPyObjects(segm, h, w)

  File "pycocotools/_mask.pyx", line 308, in pycocotools._mask.frPyObjects

Exception: input type is not supported.


     [[{{node PyFunc}}]]
     [[IteratorGetNext]]
0 successful operations.
0 derived errors ignored. [Op:__inference_train_function_3047]

从报错堆栈看，问题出在getNormalMask函数调用coco.annToMask(anns[a])时，传入的标注格式不符合pycocotools要求，具体排查和修复方向如下：

• 检查标注数据格式：确认自定义COCO数据集中的segmentation字段类型是否合法。pycocotools的frPyObjects仅接受三种输入：单个多边形（list of list of ints）、多个多边形（list of list of list of ints）、RLE格式字典（含counts和size键）。检查anns[a]的segmentation是否为None、字符串或其他非法类型。
• 修复标注遍历逻辑：在getNormalMask中添加格式校验，跳过或处理无效标注：

  for a in range(len(anns)):
      ann = anns[a]
      segm = ann.get('segmentation')
      if not isinstance(segm, (list, dict)):
          print(f"Invalid segmentation for annotation {ann['id']}: {type(segm)}")
          continue
      # 后续mask生成逻辑
  

• 过滤无效图像：检查数据集中是否存在无标注或标注为空的图像，在数据生成器中提前过滤这类样本，避免传入无效标注给pycocotools。
• 兼容TensorFlow图模式：若报错仍在图执行阶段出现，可将pycocotools相关调用用tf.py_function包装，确保Python代码在图模式下能正常执行，但优先解决标注格式问题。

内容的提问来源于stack exchange，提问作者Navneeth S