如何使用 tf.saved_model 加载模型并调用预测函数 [TENSORFLOW 2.0 API]

Question

我对 tensorflow 非常陌生，尤其是 2.0，因为没有足够的关于该 API 的示例，但它似乎比 1.x 方便得多 到目前为止，我设法使用 tf.estimator api 训练了一个线性模型，然后设法使用 tf.estimator.exporter 保存它。

之后我想使用 tf.saved_model api 加载这个模型，我想我成功了，但我对我的过程有些怀疑，所以这里快速浏览一下我的代码：

所以我有一个使用 tf.feature_column api 创建的功能数组，它看起来像这样：

feature_columns = 
[NumericColumn(key='geoaccuracy', shape=(1,), default_value=None, dtype=tf.float32, normalizer_fn=None),
 NumericColumn(key='longitude', shape=(1,), default_value=None, dtype=tf.float32, normalizer_fn=None),
 NumericColumn(key='latitude', shape=(1,), default_value=None, dtype=tf.float32, normalizer_fn=None),
 NumericColumn(key='bidfloor', shape=(1,), default_value=None, dtype=tf.float32, normalizer_fn=None),
 VocabularyListCategoricalColumn(key='adid', vocabulary_list=('115', '124', '139', '122', '121', '146', '113', '103', '123', '104', '147', '114', '149', '148'), dtype=tf.string, default_value=-1, num_oov_buckets=0),
 VocabularyListCategoricalColumn(key='campaignid', vocabulary_list=('36', '31', '33', '28'), dtype=tf.string, default_value=-1, num_oov_buckets=0),
 VocabularyListCategoricalColumn(key='exchangeid', vocabulary_list=('1241', '823', '1240', '1238'), dtype=tf.string, default_value=-1, num_oov_buckets=0),
...]

之后，我以这种方式使用我的特征列数组定义一个估计器，并对其进行训练。到这里，没问题。

linear_est = tf.estimator.LinearClassifier(feature_columns=feature_columns)

在训练我的模型后，我想保存它，所以这里开始怀疑，这是我如何进行但不确定这是正确的方法：

serving_input_parse = tf.feature_column.make_parse_example_spec(feature_columns=feature_columns)

""" view of the variable : serving_input_parse = 
 {'adid': VarLenFeature(dtype=tf.string),
 'at': VarLenFeature(dtype=tf.string),
 'basegenres': VarLenFeature(dtype=tf.string),
 'bestkw': VarLenFeature(dtype=tf.string),
 'besttopic': VarLenFeature(dtype=tf.string),
 'bidfloor': FixedLenFeature(shape=(1,), dtype=tf.float32, default_value=None),
 'browserid': VarLenFeature(dtype=tf.string),
 'browserlanguage': VarLenFeature(dtype=tf.string)
 ...} """

# exporting the model :
linear_est.export_saved_model(export_dir_base='./saved',
 serving_input_receiver_fn=tf.estimator.export.build_parsing_serving_input_receiver_fn(serving_input_receiver_fn),
 as_text=True)

现在我尝试加载它，但我不知道如何使用加载的模型来调用预测，例如使用来自 pandas 数据帧的原始数据

loaded = tf.saved_model.load('saved/1573144361/')

还有一件事，我试图查看模型的签名，但我无法真正理解我的输入形状发生了什么

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

signature_def['classification']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['inputs'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: input_example_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['classes'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 2)
        name: head/Tile:0
    outputs['scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 2)
        name: head/predictions/probabilities:0
  Method name is: tensorflow/serving/classify

signature_def['predict']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['examples'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: input_example_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['all_class_ids'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 2)
        name: head/predictions/Tile:0
    outputs['all_classes'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 2)
        name: head/predictions/Tile_1:0
    outputs['class_ids'] tensor_info:
        dtype: DT_INT64
        shape: (-1, 1)
        name: head/predictions/ExpandDims:0
    outputs['classes'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 1)
        name: head/predictions/str_classes:0
    outputs['logistic'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: head/predictions/logistic:0
    outputs['logits'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: linear/linear_model/linear/linear_model/linear/linear_model/weighted_sum:0
    outputs['probabilities'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 2)
        name: head/predictions/probabilities:0
  Method name is: tensorflow/serving/predict

signature_def['regression']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['inputs'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: input_example_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['outputs'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: head/predictions/logistic:0
  Method name is: tensorflow/serving/regress

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['inputs'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: input_example_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['classes'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 2)
        name: head/Tile:0
    outputs['scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 2)
        name: head/predictions/probabilities:0
  Method name is: tensorflow/serving/classify

Answer 1

saved_model.load(...)documentation 演示了这样的基本机制：

imported = tf.saved_model.load(path)
f = imported.signatures["serving_default"]
print(f(x=tf.constant([[1.]])))

我自己还是新手，但serving_default 似乎是使用saved_model.save(...) 时的默认签名。

（我的理解是saved_model.save(...)不保存模型，它保存图形。为了解释图形，您需要在图形上显式存储“签名”定义操作。如果你不明确地这样做，那么“serve_default”将是你唯一的签名。）

我在下面提供了一个实现。有几个细节值得注意：

1. 输入需要是张量；所以我需要手动进行转换。
2. 输出是一个字典。文档将其描述为“具有从签名键映射到函数的签名属性的可跟踪对象。”

在我的例子中，字典的键是一个相对任意的“dense_83”。这似乎有点……具体。所以我概括了使用迭代器忽略键的解决方案：

import tensorflow as tf
input_data = tf.constant(input_data, dtype=tf.float32)
prediction_tensors = signature_collection.signatures["serving_default"](input_data)
for _, values in prediction_tensors.items():
    predictions = values.numpy()[0]
    return predictions
raise Exception("Expected a response from predict(...).")

Answer 2

看起来您在最后一部分输出中使用了 saved_model_cli 命令行工具。从那里你有一个“预测”函数，它显示输入类型、列等。当我这样做时，我会看到我所有的输入列。在您的情况下，它仅显示一个输入，即名为示例的字符串。这看起来不正确。

这是$ saved_model_cli show --dir /somedir/export/exporter/123456789 --all 输出的摘录。在输出中，点显示已删除的行，因为它们看起来很相似。

signature_def['predict']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['feature_num_1'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1)
        name: Placeholder_29:0
...
...
 The given SavedModel SignatureDef contains the following output(s):
    outputs['all_class_ids'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 2)
        name: dnn/head/predictions/Tile:0
    outputs['all_classes'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 2)
        name: dnn/head/predictions/Tile_1:0
...
...
  Method name is: tensorflow/serving/predict