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 上用 Tensorflow 训练了一个 DNN model。 然后我将 model 复制到本地以仔细检查是否可以从相同的 model 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

预测是通过一些警告获得的。

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.

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

将 model 部署到 AI Platform 后，警告变为错误。 服务签名名称必须为serving_default ，如下面的错误消息所示：

{ “error”：“服务签名名称：在签名定义中找不到”serving_default“”}

使用此命令检查保存的 model 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

因此，我保存的 model 的签名名称是predict 。

问题是：如何更改签名名称？

PS：下面是我定义 DNN 的方式：

# 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 saved_model

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>})