使用 TensorFlow Serving 和 Docker 快速部署机器学习服务

发表于 | 分类于 |

从实验到生产，简单快速部署机器学习模型一直是一个挑战。这个过程要做的就是将训练好的模型对外提供预测服务。在生产中，这个过程需要可重现，隔离和安全。这里，我们使用基于Docker的TensorFlow Serving来简单地完成这个过程。TensorFlow 从1.8版本开始支持Docker部署，包括CPU和GPU，非常方便。

获得训练好的模型

获取模型的第一步当然是训练一个模型，但是这不是本篇的重点，所以我们使用一个已经训练好的模型，比如ResNet。TensorFlow Serving 使用SavedModel这种格式来保存其模型，SavedModel是一种独立于语言的，可恢复，密集的序列化格式，支持使用更高级别的系统和工具来生成，使用和转换TensorFlow模型。这里我们直接下载一个预训练好的模型：

1
2
$ mkdir /tmp/resnet
$ curl -s https://storage.googleapis.com/download.tensorflow.org/models/official/20181001_resnet/savedmodels/resnet_v2_fp32_savedmodel_NHWC_jpg.tar.gz | tar --strip-components=2 -C /tmp/resnet -xvz

如果是使用其他框架比如Keras生成的模型，则需要将模型转换为SavedModel格式，比如：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
from keras.models import Sequential
from keras import backend as K
import tensorflow as tf

model = Sequential()
# 中间省略模型构建

# 模型转换为SavedModel
signature = tf.saved_model.signature_def_utils.predict_signature_def(
    inputs={'input_param': model.input}, outputs={'type': model.output})
builder = tf.saved_model.builder.SavedModelBuilder('/tmp/output_model_path/1/')
builder.add_meta_graph_and_variables(
    sess=K.get_session(),
    tags=[tf.saved_model.tag_constants.SERVING],
    signature_def_map={
        tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
            signature
    })
builder.save()

下载完成后，文件目录树为：

1
2
3
4
5
6
7
$ tree /tmp/resnet
/tmp/resnet
└── 1538687457
    ├── saved_model.pb
    └── variables
        ├── variables.data-00000-of-00001
        └── variables.index

部署模型

第一步是安装Docker CE。这将为您提供运行和管理Docker容器所需的所有工具。现在我们有了我们的模型，使用Docker服务就像拉动最新发布的TensorFlow服务服务环境镜像一样简单，并将其指向模型：

使用Docker部署模型服务：

1
2
3
4
5
6
7
8

$ docker pull tensorflow/serving
$ docker run -p 8501:8501 --name tfserving_resnet \
--mount type=bind,source=/tmp/resnet,target=/models/resnet \
-e MODEL_NAME=resnet -t tensorflow/serving &
…
… main.cc:327] Running ModelServer at 0.0.0.0:8500…
… main.cc:337] Exporting HTTP/REST API at:localhost:8501 …

分解命令行参数，分别是：

  • -p 8501:8501 ：将容器的端口8501（TF服务响应REST API请求）发布到主机的端口8501
  • --name tfserving_resnet ：给容器我们创建名称“tfserving_resnet”，以便我们稍后可以参考它
  • --mount type=bind,source=/tmp/resnet,target=/models/resnet ：在主机（/ models / resnet）上安装主机的本地目录（/ tmp / resnet），以便TF服务可以从容器内部读取模型。
  • -e MODEL_NAME=resnet ：Telling TensorFlow服务加载名为“resnet”的模型
  • -t tensorflow/serving ：根据服务镜像“tensorflow / serving”运行Docker容器

其中，8500端口对于TensorFlow Serving提供的gRPC端口，8501为REST API服务端口。上述命令输出为

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
2019-03-04 02:52:26.610387: I tensorflow_serving/model_servers/server.cc:82] Building single TensorFlow model file config:  model_name: resnet model_base_path: /models/resnet
2019-03-04 02:52:26.618200: I tensorflow_serving/model_servers/server_core.cc:461] Adding/updating models.
2019-03-04 02:52:26.618628: I tensorflow_serving/model_servers/server_core.cc:558]  (Re-)adding model: resnet
2019-03-04 02:52:26.745813: I tensorflow_serving/core/basic_manager.cc:739] Successfully reserved resources to load servable {name: resnet version: 1538687457}
2019-03-04 02:52:26.745901: I tensorflow_serving/core/loader_harness.cc:66] Approving load for servable version {name: resnet version: 1538687457}
2019-03-04 02:52:26.745935: I tensorflow_serving/core/loader_harness.cc:74] Loading servable version {name: resnet version: 1538687457}
2019-03-04 02:52:26.747590: I external/org_tensorflow/tensorflow/contrib/session_bundle/bundle_shim.cc:363] Attempting to load native SavedModelBundle in bundle-shim from: /models/resnet/1538687457
2019-03-04 02:52:26.747705: I external/org_tensorflow/tensorflow/cc/saved_model/reader.cc:31] Reading SavedModel from: /models/resnet/1538687457
2019-03-04 02:52:26.795363: I external/org_tensorflow/tensorflow/cc/saved_model/reader.cc:54] Reading meta graph with tags { serve }
2019-03-04 02:52:26.828614: I external/org_tensorflow/tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-03-04 02:52:26.923902: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:162] Restoring SavedModel bundle.
2019-03-04 02:52:28.098479: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:138] Running MainOp with key saved_model_main_op on SavedModel bundle.
2019-03-04 02:52:28.144510: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:259] SavedModel load for tags { serve }; Status: success. Took 1396689 microseconds.
2019-03-04 02:52:28.146646: I tensorflow_serving/servables/tensorflow/saved_model_warmup.cc:83] No warmup data file found at /models/resnet/1538687457/assets.extra/tf_serving_warmup_requests
2019-03-04 02:52:28.168063: I tensorflow_serving/core/loader_harness.cc:86] Successfully loaded servable version {name: resnet version: 1538687457}
2019-03-04 02:52:28.174902: I tensorflow_serving/model_servers/server.cc:286] Running gRPC ModelServer at 0.0.0.0:8500 ...
[warn] getaddrinfo: address family for nodename not supported
2019-03-04 02:52:28.186724: I tensorflow_serving/model_servers/server.cc:302] Exporting HTTP/REST API at:localhost:8501 ...
[evhttp_server.cc : 237] RAW: Entering the event loop ...

我们可以看到，TensorFlow Serving使用1538687457作为模型的版本号。我们使用curl命令来查看一下启动的服务状态，也可以看到提供服务的模型版本以及模型状态。

1
2
3
4
5
6
7
8
9
10
11
12
13
$ curl http://localhost:8501/v1/models/resnet
{
 "model_version_status": [
  {
   "version": "1538687457",
   "state": "AVAILABLE",
   "status": {
    "error_code": "OK",
    "error_message": ""
   }
  }
 ]
}

查看模型输入输出

很多时候我们需要查看模型的输出和输出参数的具体形式，TensorFlow提供了一个saved_model_cli命令来查看模型的输入和输出参数：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
$ saved_model_cli show --dir /tmp/resnet/1538687457/ --all

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

signature_def['predict']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['image_bytes'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: input_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['classes'] tensor_info:
        dtype: DT_INT64
        shape: (-1)
        name: ArgMax:0
    outputs['probabilities'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1001)
        name: softmax_tensor:0
  Method name is: tensorflow/serving/predict

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['image_bytes'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: input_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['classes'] tensor_info:
        dtype: DT_INT64
        shape: (-1)
        name: ArgMax:0
    outputs['probabilities'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 1001)
        name: softmax_tensor:0
  Method name is: tensorflow/serving/predict

注意到signature_def，inputs的名称，类型和输出，这些参数在接下来的模型预测请求中需要。

使用模型接口预测：REST和gRPC

TensorFlow Serving提供REST API和gRPC两种请求方式，接下来将具体这两种方式。

REST

我们下载一个客户端脚本，这个脚本会下载一张猫的图片，同时使用这张图片来计算服务请求时间。

1
$ curl -o /tmp/resnet/resnet_client.py https://raw.githubusercontent.com/tensorflow/serving/master/tensorflow_serving/example/resnet_client.py

以下脚本使用requests库来请求接口，使用图片的base64编码字符串作为请求内容，返回图片分类，并计算了平均处理时间。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
from __future__ import print_function

import base64
import requests

# The server URL specifies the endpoint of your server running the ResNet
# model with the name "resnet" and using the predict interface.
SERVER_URL = 'http://localhost:8501/v1/models/resnet:predict'

# The image URL is the location of the image we should send to the server
IMAGE_URL = 'https://tensorflow.org/images/blogs/serving/cat.jpg'


def main():
  # Download the image
  dl_request = requests.get(IMAGE_URL, stream=True)
  dl_request.raise_for_status()

  # Compose a JSON Predict request (send JPEG image in base64).
  jpeg_bytes = base64.b64encode(dl_request.content).decode('utf-8')
  predict_request = '{"instances" : [{"b64": "%s"}]}' % jpeg_bytes

  # Send few requests to warm-up the model.
  for _ in range(3):
    response = requests.post(SERVER_URL, data=predict_request)
    response.raise_for_status()

  # Send few actual requests and report average latency.
  total_time = 0
  num_requests = 10
  for _ in range(num_requests):
    response = requests.post(SERVER_URL, data=predict_request)
    response.raise_for_status()
    total_time += response.elapsed.total_seconds()
    prediction = response.json()['predictions'][0]

  print('Prediction class: {}, avg latency: {} ms'.format(
      prediction['classes'], (total_time*1000)/num_requests))


if __name__ == '__main__':
  main()

输出结果为

1
2
$ python resnet_client.py
Prediction class: 286, avg latency: 210.12310000000002 ms

gRPC

让我们下载另一个客户端脚本，这个脚本使用gRPC作为服务，传入图片并获取输出结果。这个脚本需要安装tensorflow-serving-api这个库。

1
2
$ curl -o /tmp/resnet/resnet_client_grpc.py https://raw.githubusercontent.com/tensorflow/serving/master/tensorflow_serving/example/resnet_client_grpc.py
$ pip install tensorflow-serving-api

脚本内容：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
from __future__ import print_function

# This is a placeholder for a Google-internal import.

import grpc
import requests
import tensorflow as tf

from tensorflow_serving.apis import predict_pb2
from tensorflow_serving.apis import prediction_service_pb2_grpc

# The image URL is the location of the image we should send to the server
IMAGE_URL = 'https://tensorflow.org/images/blogs/serving/cat.jpg'

tf.app.flags.DEFINE_string('server', 'localhost:8500',
                           'PredictionService host:port')
tf.app.flags.DEFINE_string('image', '', 'path to image in JPEG format')
FLAGS = tf.app.flags.FLAGS


def main(_):
  if FLAGS.image:
    with open(FLAGS.image, 'rb') as f:
      data = f.read()
  else:
    # Download the image since we weren't given one
    dl_request = requests.get(IMAGE_URL, stream=True)
    dl_request.raise_for_status()
    data = dl_request.content

  channel = grpc.insecure_channel(FLAGS.server)
  stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
  # Send request
  # See prediction_service.proto for gRPC request/response details.
  request = predict_pb2.PredictRequest()
  request.model_spec.name = 'resnet'
  request.model_spec.signature_name = 'serving_default'
  request.inputs['image_bytes'].CopyFrom(
      tf.contrib.util.make_tensor_proto(data, shape=[1]))
  result = stub.Predict(request, 10.0)  # 10 secs timeout
  print(result)


if __name__ == '__main__':
  tf.app.run()

输出的结果可以看到图片的分类，概率和使用的模型信息：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
$ python resnet_client_grpc.py
outputs {
  key: "classes"
  value {
    dtype: DT_INT64
    tensor_shape {
      dim {
        size: 1
      }
    }
    int64_val: 286
  }
}
outputs {
  key: "probabilities"
  value {
    dtype: DT_FLOAT
    tensor_shape {
      dim {
        size: 1
      }
      dim {
        size: 1001
      }
    }
    float_val: 2.4162832232832443e-06
    float_val: 1.9012182974620373e-06
    float_val: 2.7247710022493266e-05
    float_val: 4.426385658007348e-07
    ...(中间省略)
    float_val: 1.4636580090154894e-05
    float_val: 5.812107133351674e-07
    float_val: 6.599806511076167e-05
    float_val: 0.0012952701654285192
  }
}
model_spec {
  name: "resnet"
  version {
    value: 1538687457
  }
  signature_name: "serving_default"
}

性能

通过编译优化的TensorFlow Serving二进制来提高性能

TensorFlows serving有时会有输出如下的日志：

1
Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA

TensorFlow Serving已发布Docker镜像旨在尽可能多地使用CPU架构，因此省略了一些优化以最大限度地提高兼容性。如果你没有看到此消息，则你的二进制文件可能已针对你的CPU进行了优化。根据你的模型执行的操作，这些优化可能会对你的服务性能产生重大影响。幸运的是，编译优化的TensorFlow Serving二进制非常简单。官方已经提供了自动化脚本，分以下两部进行：

1
2
3
4
5
6
7
# 1. 编译开发版本：首先，我们要构建TensorFlow服务的优化版本。最简单的方法是构建官方的Tensorflow服务开发环境Docker镜像。这具有为图像构建的系统自动生成优化的TensorFlow服务二进制文件的良好特性。为了区分我们创建的图像和官方图像，我们将$ USER /添加到镜像名称之前。让我们称这个开发镜像为$ USER / tensorflow-serving-devel

$ docker build -t $USER/tensorflow-serving-devel -f Dockerfile.devel https://github.com/tensorflow/serving.git#:tensorflow_serving/tools/docker

# 2. 生产新的镜像：构建TensorFlow服务开发映像可能需要一段时间，具体取决于计算机的速度。完成后，让我们使用优化的二进制文件构建一个新的服务图像，并将其命名为$ USER / tensorflow-serving：

$ docker build -t $USER/tensorflow-serving --build-arg TF_SERVING_BUILD_IMAGE=$USER/tensorflow-serving-devel https://github.com/tensorflow/serving.git#:tensorflow_serving/tools/docker

现在我们有了新的服务图像，让我们再次启动服务器：:

1
2
3
4
$ docker kill tfserving_resnet
$ docker run -p 8501:8501 --name tfserving_resnet \
  --mount type=bind,source=/tmp/resnet,target=/models/resnet \
  -e MODEL_NAME=resnet -t $USER/tensorflow-serving &

最后运行我们的客户端程序：

1
2
$ python /tmp/resnet/resnet_client.py
Prediction class: 282, avg latency: 84.8849 ms

在我们的机器上，我们看到使用我们的原生优化二进制文件，每次预测平均加速超过100毫秒（119％）。根据您的机器（和型号），您可能会看到不同的结果。

最后，结束TensorFlow Serving容器：

1
$ docker kill tfserving_resnet

参考

0%