Tensorflow模型保存与预测

目录

1. Estimator形式

1.1 模型导出

A. 用到了tf.feature_column接口

B. 没用到tf.feature_column接口，需要自己定义feature_spec

1.2 模型检查

1.3 模型预测

A. 使用CLI指令

B. python加载模型预测

2. Keras Model 训练

1.1 模型导出

1.2 模型预测

A. 使用CLI指令

B. python加载模型预测

---

版本：tensorflow 1.14
保存形式：svaedModel

1. Estimator形式

模型保存后，预测数据以example格式输入

1.1 模型导出

A. 用到了tf.feature_column接口

# feature_columns 
feature_columns = [
     tf.feature_column.numeric_column('dense_features', (10,)),
     ... 
]
serving_feature_spec = tf.feature_column.make_parse_example_spec(feature_columns)
serving_input_receiver_fn = tf.estimator.export.build_parsing_serving_input_receiver_fn(serving_feature_spec)
model_estimator.export_saved_model(export_dir_base=save_model_dir,
                             serving_input_receiver_fn=serving_input_receiver_fn)

• 第4行 make_parse_example_spec() 会根据创建的feature column列表，构建出解析tf.Example所需要的信息，比如：

{
    'feature_1': VarLenFeature(dtype=tf.int64),
    'feature_2': VarLenFeature(dtype=tf.int64),
    'dense_features': FixedLenFeature(shape=(10,), dtype=tf.float32, default_value=None)
}

• 第5行 build_parsing_serving_input_receiver_fn函数 注册接收key为examples的string tensor作为input，依据feature_spec解析后给到模型，源码如下：
  • 本质是：将来模型的输入是一个example形式的序列化string，对其按照feature_spec解析成一个个字段，形成features

def build_parsing_serving_input_receiver_fn(feature_spec, default_batch_size=None):
    def serving_input_receiver_fn():
    """An input_fn that expects a serialized tf.Example."""
        serialized_tf_example = array_ops.placeholder(dtype=dtypes.string,
                                                      shape=[default_batch_size],
                                                      name='input_example_tensor')
        receiver_tensors = {'examples': serialized_tf_example}
        features = parsing_ops.parse_example(serialized_tf_example, feature_spec)
        return ServingInputReceiver(features, receiver_tensors)
    return serving_input_receiver_fn

B. 没用到tf.feature_column接口，需要自己定义feature_spec

feature_spec = {
    'feature_1': tf.io.VarLenFeature(tf.int64), 
    'feature_2': tf.io.VarLenFeature(tf.int64),
    'abc': tf.io.FixedLenFeature([1], tf.int64, default_value=[0]),
    ...
    'dense_features': tf.io.FixedLenFeature(shape=(10,), dtype=tf.float32, default_value=[0] * 10)
}

• 可以自己重写build函数（其实本质和原来的一样...）

def my_build_serving_input_receiver_fn(cols_description):
    def serving_input_receiver_fn():
        serialized_tf_example = tf.placeholder(dtype=tf.string, shape=None, 
                                            name='input_example_tensor')
        # key (e.g. 'examples') should be same with the inputKey when you 
        # buid the request for prediction
        receiver_tensors = {'input_examples': serialized_tf_example}
        features = tf.parse_example(serialized_tf_example, cols_description)
        return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)
    return serving_input_receiver_fn

最后进行导出

serving_input_receiver_fn = my_build_serving_input_receiver_fn(feature_spec)
model_estimator.export_saved_model(export_dir_base=save_model_dir,
                             serving_input_receiver_fn=serving_input_receiver_fn)

导出的模型目录结构如下

/export/model
└── 1784256271
    ├── saved_model.pb
    └── variables
        ├── variables.data-00000-of-00002
        ├── variables.data-00001-of-00002
        └── variables.index

1.2 模型检查

命令行 使用CLI指令 检查保存的模型

saved_model_cli show --all --dir ./model_path

1.3 模型预测

A. 使用CLI指令

saved_model_cli run --dir ${model_path} \
    --tag_set serve \
    --signature_def="predict" \
    --input_examples='examples=[{"feature_1":[0], "feature_2":[1], "dense_features":[0.0, 0.0, 1.0, 1.5, 1.0, 2.0, 3.0, 0.2, 5.0, 0.0]}]'

结果：

B. python加载模型预测

需要包装成tf.Example序列化成string后喂入

# 将pandas df包装成tf.example
def predict_example(df):
    examples = []
    colNames = df.columns
    dtypes = df.dtypes
    for row in df.iterrows():
        features = collections.OrderedDict()
        for i in range(len(colNames)):
            dtype = dtypes[i]
            colName = colNames[i]
            value = row[1][colName]
            if dtype == "int":
                features[colName] = tf.train.Feature(int64_list=tf.train.Int64List(value=[value])) # 注意，要包装成list
            elif dtype == 'float':
                features[colName] = tf.train.Feature(float_list=tf.train.FloatList(value=[value]))
            else:
                features[colName] = tf.train.Feature(float_list=tf.train.FloatList(value=value)) # 本来就是List，不用包装了
        tf_features = tf.train.Features(feature=features)
        tf_example = tf.train.Example(features=tf_features)
        tf_example = tf_example.SerializeToString()
        examples.append(tf_example)
    return examples
# dataframe case
data = [{
    "feature_1: 0, "feature_2": 1, "dense_features": [0.0]*10
}]
df = pd.DataFrame(data)
# 构建预测器和预测样本
predict_fn = tf.contrib.predictor.from_saved_model(model_path, signature_def_key="predict")
examples = predict_example(df)
predictions = predict_fn({"examples": examples})
print(predictions)

2. Keras Model 训练

1.1 模型导出

定义好模型预测的输入输出节点

save_path = "./keras_model/1"
builder = tf.saved_model.builder.SavedModelBuilder(save_path)
with tf.get_default_graph().as_default():
    # define signature which specify input and out nodes
    tensor_info_output = tf.saved_model.utils.build_tensor_info(model.output)

    prediction_signature = (tf.saved_model.signature_def_utils.build_signature_def(
        inputs={f"input_{x}": tf.saved_model.utils.build_tensor_info(item) for x, item in enumerate(model.input)},
        outputs={"pred": tensor_info_output},
        method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME))

    # add graph and variables
    builder.add_meta_graph_and_variables(
        # tags:SERVING,TRAINING,EVAL,GPU,TPU
        sess=tf.compat.v1.keras.backend.get_session(),
        tags=[tf.saved_model.tag_constants.SERVING],
        signature_def_map={'predict': prediction_signature},
    )
    builder.save()

1.2 模型预测

A. 使用CLI指令

saved_model_cli run --dir ${keras_path} \
    --tag_set serve \
    --signature_def="predict" \
    --input_exprs 'input_0=[[0]];input_1=[[0]];input_3=[[0.0]*10]'

B. python加载模型预测

以字典的形式直接喂入，每个输入值都要用list包起来

predict_fn = tf.contrib.predictor.from_saved_model(model_path, signature_def_key="predict")
predictions = predict_fn({"input_0": [[0]],
                          "input_1": [[0]],
                          "input_3": [[0.0]*10],
                          })