如何将签名名称设置为 serving_default 以避免部署到 GCP AI 平台后出错

Question

我在 AI Platform 上使用 Tensorflow 训练了一个 DNN 模型。然后我在本地复制模型以仔细检查是否可以从相同的模型中获得预测。

gcloud ai-platform local predict --model-dir=/home/jupyter/end-to-end-ml/examples/e2e-ml-model-ex02/app/appbabyweight_trained/export/exporter/1615197796 --json-instances=inputs.json

预测结果带有一些警告。

If the signature defined in the model is not `serving_default` then you must specify it via --signature-name flag, otherwise the command may fail.

（在指定签名名称时可以避免此警告：--signature-name predict）

将模型部署到 AI Platform 后，警告变为错误。服务签名名称必须为serving_default，如以下错误消息所示：

{ "error": "Serving signature name: "serving_default" not found in signature def" }

使用此命令检查保存的模型后：

saved_model_cli show --dir /home/jupyter/end-to-end-ml/examples/e2e-ml-model-ex02/app/appbabyweight_trained2/output-dir/export/exporter/1615439076 --all

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['predict']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['gestation_weeks'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1)
        name: Placeholder_3:0
    inputs['is_male'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: Placeholder:0
    inputs['mother_age'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1)
        name: Placeholder_1:0
    inputs['plurality'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: Placeholder_2:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['predictions'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: dnn/logits/BiasAdd:0
  Method name is: tensorflow/serving/predict

所以，我保存的模型的签名名称是predict。

问题是：如何更改签名名称？

PS：下面是我如何定义 DNN：

# Define feature columns
def get_categorical(name, values):
    return tf.feature_column.indicator_column(
        tf.feature_column.categorical_column_with_vocabulary_list(name, values))

def get_cols():
    # Define column types
    return [\
            get_categorical('is_male', ['True', 'False', 'Unknown']),
            tf.feature_column.numeric_column('mother_age'),
            get_categorical('plurality',
                        ['Single(1)', 'Twins(2)', 'Triplets(3)',
                         'Quadruplets(4)', 'Quintuplets(5)','Multiple(2+)']),
            tf.feature_column.numeric_column('gestation_weeks')
        ]

# Create serving input function to be able to serve predictions later using provided inputs
def serving_input_fn():
    feature_placeholders = {
        'is_male': tf.compat.v1.placeholder(tf.string, [None]),
        'mother_age': tf.compat.v1.placeholder(tf.float32, [None]),
        'plurality': tf.compat.v1.placeholder(tf.string, [None]),
        'gestation_weeks': tf.compat.v1.placeholder(tf.float32, [None])
    }
    features = {
        key: tf.expand_dims(tensor, -1) for key, tensor in feature_placeholders.items()
    }
    return tf.estimator.export.ServingInputReceiver(features, feature_placeholders)

# Create estimator to train and evaluate
def train_and_evaluate(args):
    
    EVAL_INTERVAL = 30
    run_config = tf.estimator.RunConfig(save_checkpoints_secs = EVAL_INTERVAL, keep_checkpoint_max = 3)

    estimator = tf.estimator.DNNRegressor(
                        model_dir = args['output_dir'],
                        feature_columns = get_cols(),
                        hidden_units = args['nnsize'],
                        config = run_config)
    train_spec = tf.estimator.TrainSpec(
                        input_fn = read_dataset(args['train_data_path'],
                                    mode = tf.estimator.ModeKeys.TRAIN,
                                    batch_size =args['batch_size']),
                        max_steps = TRAIN_STEPS)
    exporter = tf.estimator.LatestExporter('exporter', serving_input_fn)
    eval_spec = tf.estimator.EvalSpec(
                        input_fn = read_dataset(args['eval_data_path'], mode = tf.estimator.ModeKeys.EVAL, batch_size =args['batch_size']),
                        steps = args['eval_steps'],
                        start_delay_secs = 60, # start evaluating after N seconds
                        throttle_secs = EVAL_INTERVAL,  # evaluate every N seconds
                        exporters = exporter)
    tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)

谢谢

Answer 1

将签名 serving_default 添加到现有的已保存模型

import tensorflow as tf
m = tf.saved_model.load("tf2-preview_inception_v3_classification_4")
print(m.signatures) # _SignatureMap({}) - Empty
t_spec = tf.TensorSpec([None,None,None,3], tf.float32)
c_func = m.__call__.get_concrete_function(inputs=t_spec)
signatures = {'serving_default': c_func}
tf.saved_model.save(m, 'tf2-preview_inception_v3_classification_5', signatures=signatures)

# Test new model
m5 = tf.saved_model.load("tf2-preview_inception_v3_classification_5")
print(m5.signatures) # _SignatureMap({'serving_default': <ConcreteFunction signature_wrapper(*, inputs) at 0x17316DC50>})