重庆分公司,新征程启航

为企业提供网站建设、域名注册、服务器等服务

如何部署TensorFlowServing

这篇文章主要介绍“如何部署TensorFlow Serving”,在日常操作中,相信很多人在如何部署TensorFlow Serving问题上存在疑惑,小编查阅了各式资料,整理出简单好用的操作方法,希望对大家解答”如何部署TensorFlow Serving”的疑惑有所帮助!接下来,请跟着小编一起来学习吧!

创新互联建站是专业的昭平网站建设公司,昭平接单;提供成都网站建设、成都网站设计,网页设计,网站设计,建网站,PHP网站建设等专业做网站服务;采用PHP框架,可快速的进行昭平网站开发网页制作和功能扩展;专业做搜索引擎喜爱的网站,专业的做网站团队,希望更多企业前来合作!

准备环境

准备好 TensorFlow 环境,导入依赖:

import sys

# Confirm that we're using Python 3
assert sys.version_info.major == 3, 'Oops, not running Python 3. Use Runtime > Change runtime type'
import tensorflow as tf
from tensorflow import keras

# Helper libraries
import numpy as np
import matplotlib.pyplot as plt
import os
import subprocess

print(f'TensorFlow version: {tf.__version__}')
print(f'TensorFlow GPU support: {tf.test.is_built_with_gpu_support()}')

physical_gpus = tf.config.list_physical_devices('GPU')
print(physical_gpus)
for gpu in physical_gpus:
  # memory growth must be set before GPUs have been initialized
  tf.config.experimental.set_memory_growth(gpu, True)
logical_gpus = tf.config.experimental.list_logical_devices('GPU')
print(len(physical_gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
TensorFlow version: 2.4.1
TensorFlow GPU support: True
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
1 Physical GPUs, 1 Logical GPUs

创建模型

载入 Fashion MNIST 数据集:

fashion_mnist = keras.datasets.fashion_mnist
(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()

# scale the values to 0.0 to 1.0
train_images = train_images / 255.0
test_images = test_images / 255.0

# reshape for feeding into the model
train_images = train_images.reshape(train_images.shape[0], 28, 28, 1)
test_images = test_images.reshape(test_images.shape[0], 28, 28, 1)

class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',
               'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']

print('\ntrain_images.shape: {}, of {}'.format(train_images.shape, train_images.dtype))
print('test_images.shape: {}, of {}'.format(test_images.shape, test_images.dtype))
train_images.shape: (60000, 28, 28, 1), of float64
test_images.shape: (10000, 28, 28, 1), of float64

用最简单的 CNN 训练模型,

model = keras.Sequential([
  keras.layers.Conv2D(input_shape=(28,28,1), filters=8, kernel_size=3,
                      strides=2, activation='relu', name='Conv1'),
  keras.layers.Flatten(),
  keras.layers.Dense(10, name='Dense')
])
model.summary()

testing = False
epochs = 5

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=[keras.metrics.SparseCategoricalAccuracy()])
model.fit(train_images, train_labels, epochs=epochs)

test_loss, test_acc = model.evaluate(test_images, test_labels)
print('\nTest accuracy: {}'.format(test_acc))
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
Conv1 (Conv2D)               (None, 13, 13, 8)         80
_________________________________________________________________
flatten (Flatten)            (None, 1352)              0
_________________________________________________________________
Dense (Dense)                (None, 10)                13530
=================================================================
Total params: 13,610
Trainable params: 13,610
Non-trainable params: 0
_________________________________________________________________
Epoch 1/5
1875/1875 [==============================] - 3s 722us/step - loss: 0.7387 - sparse_categorical_accuracy: 0.7449
Epoch 2/5
1875/1875 [==============================] - 1s 793us/step - loss: 0.4561 - sparse_categorical_accuracy: 0.8408
Epoch 3/5
1875/1875 [==============================] - 1s 720us/step - loss: 0.4097 - sparse_categorical_accuracy: 0.8566
Epoch 4/5
1875/1875 [==============================] - 1s 718us/step - loss: 0.3899 - sparse_categorical_accuracy: 0.8636
Epoch 5/5
1875/1875 [==============================] - 1s 719us/step - loss: 0.3673 - sparse_categorical_accuracy: 0.8701
313/313 [==============================] - 0s 782us/step - loss: 0.3937 - sparse_categorical_accuracy: 0.8630

Test accuracy: 0.8629999756813049

保存模型

将模型保存成 SavedModel 格式,路径里加上版本号,以便 TensorFlow Serving 时可选择模型版本。

# Fetch the Keras session and save the model
# The signature definition is defined by the input and output tensors,
# and stored with the default serving key
import tempfile

MODEL_DIR = os.path.join(tempfile.gettempdir(), 'tfx')
version = 1
export_path = os.path.join(MODEL_DIR, str(version))
print('export_path = {}\n'.format(export_path))

tf.keras.models.save_model(
    model,
    export_path,
    overwrite=True,
    include_optimizer=True,
    save_format=None,
    signatures=None,
    options=None
)

print('\nSaved model:')
!ls -l {export_path}
export_path = /tmp/tfx/1

INFO:tensorflow:Assets written to: /tmp/tfx/1/assets

Saved model:
total 88
drwxr-xr-x 2 john john  4096 Apr 13 15:10 assets
-rw-rw-r-- 1 john john 78169 Apr 13 15:12 saved_model.pb
drwxr-xr-x 2 john john  4096 Apr 13 15:12 variables

查看模型

使用 saved_model_cli 工具查看模型的 MetaGraphDefs (the models) 和 SignatureDefs (the methods you can call),了解信息。

!saved_model_cli show --dir '/tmp/tfx/1' --all
2021-04-13 15:12:29.433576: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0

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

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['Conv1_input'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 28, 28, 1)
        name: serving_default_Conv1_input:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['Dense'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 10)
        name: StatefulPartitionedCall:0
  Method name is: tensorflow/serving/predict

Defined Functions:
  Function Name: '__call__'
    Option #1
      Callable with:
        Argument #1
          Conv1_input: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='Conv1_input')
        Argument #2
          DType: bool
          Value: False
        Argument #3
          DType: NoneType
          Value: None
    Option #2
      Callable with:
        Argument #1
          inputs: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='inputs')
        Argument #2
          DType: bool
          Value: False
        Argument #3
          DType: NoneType
          Value: None
    Option #3
      Callable with:
        Argument #1
          inputs: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='inputs')
        Argument #2
          DType: bool
          Value: True
        Argument #3
          DType: NoneType
          Value: None
    Option #4
      Callable with:
        Argument #1
          Conv1_input: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='Conv1_input')
        Argument #2
          DType: bool
          Value: True
        Argument #3
          DType: NoneType
          Value: None
  ...

部署模型

安装 Serving

echo "deb [arch=amd64] http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal" | sudo tee /etc/apt/sources.list.d/tensorflow-serving.list && \
curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | sudo apt-key add -

sudo apt update
sudo apt install tensorflow-model-server

开启 Serving

开启 TensorFlow Serving ,提供 REST API :

  • rest_api_port: REST 请求端口。

  • model_name: REST 请求 URL ,自定义的名称。

  • model_base_path: 模型所在目录。

nohup tensorflow_model_server \
  --rest_api_port=8501 \
  --model_name=fashion_model \
  --model_base_path="/tmp/tfx" >server.log 2>&1 &
$ tail server.log
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2021-04-13 15:12:10.706648: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:206] Restoring SavedModel bundle.
2021-04-13 15:12:10.726722: I external/org_tensorflow/tensorflow/core/platform/profile_utils/cpu_utils.cc:112] CPU Frequency: 2599990000 Hz
2021-04-13 15:12:10.756506: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:190] Running initialization op on SavedModel bundle at path: /tmp/tfx/1
2021-04-13 15:12:10.759935: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:277] SavedModel load for tags { serve }; Status: success: OK. Took 110653 microseconds.
2021-04-13 15:12:10.760277: I tensorflow_serving/servables/tensorflow/saved_model_warmup_util.cc:59] No warmup data file found at /tmp/tfx/1/assets.extra/tf_serving_warmup_requests
2021-04-13 15:12:10.760486: I tensorflow_serving/core/loader_harness.cc:87] Successfully loaded servable version {name: fashion_model version: 1}
2021-04-13 15:12:10.763938: I tensorflow_serving/model_servers/server.cc:371] Running gRPC ModelServer at 0.0.0.0:8500 ...
[evhttp_server.cc : 238] NET_LOG: Entering the event loop ...
2021-04-13 15:12:10.765308: I tensorflow_serving/model_servers/server.cc:391] Exporting HTTP/REST API at:localhost:8501 ...

访问服务

随机显示一张测试图:

def show(idx, title):
  plt.figure()
  plt.imshow(test_images[idx].reshape(28,28))
  plt.axis('off')
  plt.title('\n\n{}'.format(title), fontdict={'size': 16})

import random
rando = random.randint(0,len(test_images)-1)
show(rando, 'An Example Image: {}'.format(class_names[test_labels[rando]]))

如何部署TensorFlow Serving

创建 JSON 对象,给到三张要预测的图:

import json
data = json.dumps({"signature_name": "serving_default", "instances": test_images[0:3].tolist()})
print('Data: {} ... {}'.format(data[:50], data[len(data)-52:]))
Data: {"signature_name": "serving_default", "instances": ...  [0.0], [0.0], [0.0], [0.0], [0.0], [0.0], [0.0]]]]}

REST 请求

最新模型版本进行预测:

!pip install -q requests

import requests
headers = {"content-type": "application/json"}
json_response = requests.post('http://localhost:8501/v1/models/fashion_model:predict', data=data, headers=headers)
predictions = json.loads(json_response.text)['predictions']

show(0, 'The model thought this was a {} (class {}), and it was actually a {} (class {})'.format(
  class_names[np.argmax(predictions[0])], np.argmax(predictions[0]), class_names[test_labels[0]], test_labels[0]))

如何部署TensorFlow Serving

指定模型版本进行预测:

headers = {"content-type": "application/json"}
json_response = requests.post('http://localhost:8501/v1/models/fashion_model/versions/1:predict', data=data, headers=headers)
predictions = json.loads(json_response.text)['predictions']

for i in range(0,3):
  show(i, 'The model thought this was a {} (class {}), and it was actually a {} (class {})'.format(
    class_names[np.argmax(predictions[i])], np.argmax(predictions[i]), class_names[test_labels[i]], test_labels[i]))

如何部署TensorFlow Serving

如何部署TensorFlow Serving

如何部署TensorFlow Serving

到此,关于“如何部署TensorFlow Serving”的学习就结束了,希望能够解决大家的疑惑。理论与实践的搭配能更好的帮助大家学习,快去试试吧!若想继续学习更多相关知识,请继续关注创新互联网站,小编会继续努力为大家带来更多实用的文章!


新闻标题:如何部署TensorFlowServing
文章起源:http://cqcxhl.cn/article/pcogid.html

其他资讯

在线咨询
服务热线
服务热线:028-86922220
TOP