我遵循了這個很棒的教程并成功地訓練了一個模型（在 CloudML 上）。我的代碼也進行離線預測，但現(xiàn)在我正在嘗試使用 Cloud ML 進行預測并遇到一些問題。為了部署我的模型，我遵循了本教程?，F(xiàn)在我有一個生成TFRecordsvia的代碼，apache_beam.io.WriteToTFRecord我想對這些TFRecords. 為此，我正在關注這篇文章，我的命令如下所示：gcloud ml-engine jobs submit prediction $JOB_ID --model $MODEL --input-paths gs://"$FILE_INPUT".gz --output-path gs://"$OUTPUT"/predictions --region us-west1 --data-format TF_RECORD_GZIP但我只得到錯誤： 'Exception during running the graph: Expected serialized to be a scalar, got shape: [64]似乎它需要不同格式的數(shù)據(jù)。我在這里找到了 JSON 的格式規(guī)范，但找不到如何使用 TFrecords 來實現(xiàn)。更新：這是輸出 saved_model_cli show --all --dirMetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:signature_def['prediction']:  The given SavedModel SignatureDef contains the following input(s):    inputs['example_proto'] tensor_info:    dtype: DT_STRING    shape: unknown_rank    name: input:0  The given SavedModel SignatureDef contains the following output(s):    outputs['probability'] tensor_info:    dtype: DT_FLOAT    shape: (1, 1)    name: probability:0  Method name is: tensorflow/serving/predictsignature_def['serving_default']:  The given SavedModel SignatureDef contains the following input(s):    inputs['example_proto'] tensor_info:    dtype: DT_STRING    shape: unknown_rank    name: input:0  The given SavedModel SignatureDef contains the following output(s):    outputs['probability'] tensor_info:    dtype: DT_FLOAT    shape: (1, 1)    name: probability:0  Method name is: tensorflow/serving/predict