tensorflow.js转换为'ssd_mobilenet_v1_coco'的对象检测模型会产生全部零结果

Question

我已将'ssd_mobilenet_v1_coco_2017_11_17'TF对象检测模型转换为带有tfjs-converter的tensorflow.js格式为tf_frozen_model。不管我提供的输入如何，“ model.executeAsync”始终为detection_scores输出节点生成带有零的数组。我测试了其他输出节点，它们产生相同的结果，全为零。

tensorflowjs_converter \
    --input_format=tf_frozen_model \
    --output_node_names='detection_scores' \
    --saved_model_tags=serve \
    /Tensor_Flow_Models/research/object_detection/ssd_mobilenet_v1_coco_2017_11_17/frozen_inference_graph.pb \
    /Tensor_Flow_Models/research/object_detection/tfjsmodels


async loadMobilenet() 
{

  this.picture = <HTMLImageElement> document.getElementById("pic");

  const MODEL_URL = '/assets/tensorflowjs_model.pb';
  const WEIGHTS_URL = '/assets/weights_manifest.json';

  tf.loadFrozenModel(MODEL_URL, WEIGHTS_URL).then(async (model) => {
    const image_tensor = tf.expandDims(tf.fromPixels(this.picture),0);

    const result = <tf.Tensor> await model.executeAsync({image_tensor: 
     image_tensor});

    console.log(result.dataSync());
  })

我在项目中使用的tfjs版本：    “ @ tensorflow / tfjs”：“ ^ 0.13.5”，    “ @ tensorflow / tfjs-converter”：“ ^ 0.8.4”，