SageMaker调用create_labeling_job()生成标注结果异常求助

通过AWS控制台创建图像Bounding Box标注任务可正常完成，但使用boto3的create_labeling_job()函数创建的任务，虽状态显示COMPLETE、标注已完成，输出manifest文件却无有效标注数据：image_size宽高为0，annotations为空。

用户提供的相关文件 #

1. Boto3代码示例

task_description = "Draw bounding box around the images"
task_keywords = ['Images', 'bounding boxes', 'object detection']
task_title = "Bounding Box task"
job_name = "test-labeling-job"
acs_arn = "arn:aws:lambda:eu-west-1:568282634449:function:ACS-BoundingBox"
prehuman_arn = "arn:aws:lambda:eu-west-1:568282634449:function:PRE-BoundingBox"

human_task_config = {
"AnnotationConsolidationConfig": {
"AnnotationConsolidationLambdaArn": acs_arn,
},
"PreHumanTaskLambdaArn": prehuman_arn,
"MaxConcurrentTaskCount": 1000, # 200 images will be sent at a time to the workteam.
"NumberOfHumanWorkersPerDataObject": 1, # 3 separate workers will be required to label each image.
"TaskAvailabilityLifetimeInSeconds": 864000, # Your workteam has 6 hours to complete all pending tasks.
"TaskDescription": task_description,
"TaskKeywords": task_keywords,
"TaskTimeLimitInSeconds": 3600, # Each image must be labeled within 5 minutes.
"TaskTitle": task_title,
"UiConfig": {
"UiTemplateS3Uri": "s3://bucket-name/bounding-box.liquid.html",
},
"WorkteamArn": "arn:aws:sagemaker:eu-west-1:xxxxxxxxx:workteam/private-crowd/Labellers"
}

ground_truth_request = {
"InputConfig": {
"DataSource": {
"S3DataSource": {
"ManifestS3Uri": "s3://bucket-name/manifest-vic-2.manifest",
}
},
"DataAttributes": {
"ContentClassifiers": []
},
},
"OutputConfig": {
"S3OutputPath": "s3://sagemaker-bucket-name/folder-name-07-12-2022/",
},
"StoppingConditions":{
'MaxPercentageOfInputDatasetLabeled': 100
},
"HumanTaskConfig": human_task_config,
"LabelingJobName": job_name,
"RoleArn": "arn:aws:iam::xxxxxxxxx:role/service-role/AmazonSageMaker-ExecutionRole-xxxxxxxxxx",
"LabelAttributeName": job_name,
"LabelCategoryConfigS3Uri": "s3://bucket-name/label-config.json",
}

sagemaker_client = boto3.client("sagemaker")
sagemaker_client.create_labeling_job(**ground_truth_request)

2. 标签配置文件（label-config.json）

{ 
"document-version": "2018-11-28", 
"labels": [{"label": "animal"}, {"label": "no-animal"}] 
}

3. 输入manifest文件

{'source-ref': 's3://sagemaker-inout-folder/2022-10-10-12:50:59/A.jpg'} 
{'source-ref': 's3://sagemaker-inout-folder/2022-10-10-12:50:59/B.jpg'} 
{'source-ref': 's3://sagemaker-inout-folder/2022-10-10-12:50:59/C.jpg'}

4. UI模板文件（bounding-box.liquid.html）

<script src="https://assets.crowd.aws/crowd-html-elements.js"></script>

<crowd-form>
  <crowd-bounding-box
    name="annotatedResult"
    src="{{ task.input.taskObject | grant_read_access }}"
    header="Draw bounding boxes around all the cats and dogs in this image"
    labels="['Animal', 'No-animal']"
  >
    <full-instructions header="Bounding Box Instructions" >
      <p>Use the bounding box tool to draw boxes around the requested target of interest:</p>
      <ol>
        <li>Draw a rectangle using your mouse over each instance of the target.</li>
        <li>Make sure the box does not cut into the target, leave a 2 - 3 pixel margin</li>
        <li>
          When targets are overlapping, draw a box around each object,
          include all contiguous parts of the target in the box.
          Do not include parts that are completely overlapped by another object.
        </li>
        <li>
          Do not include parts of the target that cannot be seen,
          even though you think you can interpolate the whole shape of the target.
        </li>
        <li>Avoid shadows, they're not considered as a part of the target.</li>
        <li>If the target goes off the screen, label up to the edge of the image.</li>
      </ol>
    </full-instructions>

    <short-instructions>
      Draw boxes around the requested target of interest.
    </short-instructions>
  </crowd-bounding-box>
</crowd-form>

5. 输出manifest示例

{ 
   "source-ref": "s3://sagemaker-inout-folder/2022-10-10-12:50:59/A.jpg", 
   "test-labeling-job": { 
      "image_size": [  
            { "width": 0, "height": 0, "depth": 3 } ], 
      "annotations": [] }, 
   "test-labeling-job-metadata": { 
        "objects": [], 
        "class-map": {}, 
        "type": "groundtruth/object-detection", 
        "human-annotated": "yes", 
        "creation-date": "2022-12-01T08:09:07.933647", 
        "job-name": "labeling-job/test-labeling-job" 
   } 
}

问题排查与解决方法 #

1. 修正Lambda函数ARN（核心问题） #

你使用的Pre/聚合Lambda是自定义ARN，若未正确适配Bounding Box任务逻辑，会导致图片尺寸无法读取、标注无法聚合。建议替换为eu-west-1区域的官方托管Lambda：

• 预任务Lambda：arn:aws:lambda:eu-west-1:432418664414:function:PRE-BoundingBox
• 聚合任务Lambda：arn:aws:lambda:eu-west-1:432418664414:function:ACS-BoundingBox
  若坚持使用自定义Lambda，需检查函数是否正确返回图片尺寸、是否能解析标注结果并生成符合要求的输出结构。

2. 统一标签大小写 #

标签配置文件中标签为全小写（animal/no-animal），但UI模板中标签首字母大写（Animal/No-animal），大小写不匹配会导致标注无法映射到标签体系，最终annotations为空。修改UI模板的labels字段：

labels="['animal', 'no-animal']"

3. 修正输入Manifest格式 #

输入Manifest必须是标准JSON格式，需将单引号替换为双引号：

{"source-ref": "s3://sagemaker-inout-folder/2022-10-10-12:50:59/A.jpg"}
{"source-ref": "s3://sagemaker-inout-folder/2022-10-10-12:50:59/B.jpg"}
{"source-ref": "s3://sagemaker-inout-folder/2022-10-10-12:50:59/C.jpg"}

4. 检查SageMaker执行角色权限 #

确保执行角色拥有以下权限：

• 输入/输出S3桶的读写权限（s3:GetObject/s3:PutObject）
• 调用指定Lambda函数的权限（lambda:InvokeFunction）
• 访问图片存储桶的权限（若图片在独立桶中）

5. 验证UI模板变量映射 #

确认UI模板中的{{ task.input.taskObject | grant_read_access }}正确映射到输入Manifest的source-ref字段，自定义Pre函数需确保返回结构包含taskObject字段，否则无法加载图片。

内容的提问来源于stack exchange，提问作者impromptu_user