python - 使用 TF 2.0 提供 Tensorflow/Keras 模型时嵌入层的问题

Question

我按照 one of the TF beginner tutorial 中的步骤操作创建一个简单的分类模型。它们是:

from __future__ import absolute_import, division, print_function, unicode_literals
import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow import feature_column
from tensorflow.keras import layers
from sklearn.model_selection import train_test_split

URL = 'https://storage.googleapis.com/applied-dl/heart.csv'
dataframe = pd.read_csv(URL)
dataframe.head()

train, test = train_test_split(dataframe, test_size=0.2)
train, val = train_test_split(train, test_size=0.2)

def df_to_dataset(dataframe, shuffle=True, batch_size=32):
  dataframe = dataframe.copy()
  labels = dataframe.pop('target')
  ds = tf.data.Dataset.from_tensor_slices((dict(dataframe), labels))
  if shuffle:
    ds = ds.shuffle(buffer_size=len(dataframe))
  ds = ds.batch(batch_size)
  return ds

batch_size = 5 # A small batch sized is used for demonstration purposes
train_ds = df_to_dataset(train, batch_size=batch_size)
val_ds = df_to_dataset(val, shuffle=False, batch_size=batch_size)
test_ds = df_to_dataset(test, shuffle=False, batch_size=batch_size)

feature_columns = []
for header in ['age', 'trestbps', 'chol', 'thalach', 'oldpeak', 'slope', 'ca']:
  feature_columns.append(feature_column.numeric_column(header))
thal_embedding = feature_column.embedding_column(thal, dimension=8)
feature_columns.append(thal_embedding)

feature_layer = tf.keras.layers.DenseFeatures(feature_columns)

batch_size = 32
train_ds = df_to_dataset(train, batch_size=batch_size)
val_ds = df_to_dataset(val, shuffle=False, batch_size=batch_size)
test_ds = df_to_dataset(test, shuffle=False, batch_size=batch_size)


model = tf.keras.Sequential([
  feature_layer,
  layers.Dense(128, activation='relu'),
  layers.Dense(128, activation='relu'),
  layers.Dense(1, activation='sigmoid')
])

model.compile(optimizer='adam',
              loss='binary_crossentropy',
              metrics=['accuracy'],
              run_eagerly=True)

model.fit(train_ds,
          validation_data=val_ds,
          epochs=5)

我保存了模型:

model.save("model/", save_format='tf')

然后，我尝试使用这个 TF tutorial 来服务这个模型。 。我执行以下操作:

docker pull tensorflow/serving
docker run -p 8501:8501 --mount type=bind,source=/path/to/model/,target=/models/model -e MODEL_NAME=mo

我尝试这样调用模型:

curl -d '{"inputs": {"age": [0], "trestbps": [0], "chol": [0], "thalach": [0], "oldpeak": [0], "slope": [1], "ca": [0], "exang": [0], "restecg": [0], "fbs": [0], "cp": [0], "sex": [0], "thal": ["normal"], "target": [0] }}' -X POST http://localhost:8501/v1/models/model:predict

我收到以下错误:

{ "error": "indices = 1 is not in [0, 1)\n\t [[{{node StatefulPartitionedCall_51/StatefulPartitionedCall/sequential/dense_features/thal_embedding/thal_embedding_weights/GatherV2}}]]" }

它似乎与“thal”特征的嵌入层有关。但我不知道“indices = 1 is not in [0, 1)”是什么意思以及为什么会发生。

发生错误时，TF docker 服务器记录的内容如下:

2019-09-23 12:50:43.921721: W external/org_tensorflow/tensorflow/core/framework/op_kernel.cc:1502] OP_REQUIRES failed at lookup_table_op.cc:952 : Failed precondition: Table already initialized.

知道错误从何而来以及如何修复它吗？

Python版本:3.6

tensorflow 版本:2.0.0-rc0

最新的 TensorFlow/serving(截至 2019 年 9 月 20 日)

模型签名:

signature_def['__saved_model_init_op']:
  The given SavedModel SignatureDef contains the following input(s):
  The given SavedModel SignatureDef contains the following output(s):
    outputs['__saved_model_init_op'] tensor_info:
        dtype: DT_INVALID
        shape: unknown_rank
        name: NoOp
  Method name is: 

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['age'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_age:0
    inputs['ca'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_ca:0
    inputs['chol'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_chol:0
    inputs['cp'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_cp:0
    inputs['exang'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_exang:0
    inputs['fbs'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_fbs:0
    inputs['oldpeak'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: serving_default_oldpeak:0
    inputs['restecg'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_restecg:0
    inputs['sex'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_sex:0
    inputs['slope'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_slope:0
    inputs['thal'] tensor_info:
        dtype: DT_STRING
        shape: (-1, 1)
        name: serving_default_thal:0
    inputs['thalach'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_thalach:0
    inputs['trestbps'] tensor_info:
        dtype: DT_INT32
        shape: (-1, 1)
        name: serving_default_trestbps:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['output_1'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1)
        name: StatefulPartitionedCall:0
  Method name is: tensorflow/serving/predict

Answer 1

我还尝试提供一个由嵌入层、lstm 层等组成的模型，但我收到了一些其他错误。我什至提出了issue在 TF 上。

无论如何，我在您的代码中看到的问题在于您用于通过 Docker 提供服务的已保存模型的类型。如果您阅读here ，它说以下几点-

A SavedModel 服务

这不是 keras model.save 而是另一个 TF API，here是描述从 keras 训练模型创建 SavedModel 的方法。尝试一下，让我们知道结果。