Questions about TFX? Join us at Google I/O!

Train and serve a TensorFlow model with TensorFlow Serving

This guide trains a neural network model to classify images of clothing, like sneakers and shirts, saves the trained model, and then serves it with TensorFlow Serving. The focus is on TensorFlow Serving, rather than the modeling and training in TensorFlow, so for a complete example which focuses on the modeling and training see the Basic Classification example.

This guide uses tf.keras, a high-level API to build and train models in TensorFlow.

import sys

# Confirm that we're using Python 3
assert sys.version_info.major is 3, 'Oops, not running Python 3. Use Runtime > Change runtime type'
# TensorFlow and tf.keras
print("Installing dependencies for Colab environment")
!pip install -Uq grpcio==1.26.0

import tensorflow as tf
from tensorflow import keras

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

print('TensorFlow version: {}'.format(tf.__version__))

Create your model

Import the Fashion MNIST dataset

This guide uses the Fashion MNIST dataset which contains 70,000 grayscale images in 10 categories. The images show individual articles of clothing at low resolution (28 by 28 pixels), as seen here:

Fashion MNIST sprite
Figure 1. Fashion-MNIST samples (by Zalando, MIT License).
 

Fashion MNIST is intended as a drop-in replacement for the classic MNIST dataset—often used as the "Hello, World" of machine learning programs for computer vision. You can access the Fashion MNIST directly from TensorFlow, just import and load the data.

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))
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-labels-idx1-ubyte.gz
32768/29515 [=================================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-images-idx3-ubyte.gz
26427392/26421880 [==============================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-labels-idx1-ubyte.gz
8192/5148 [===============================================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-images-idx3-ubyte.gz
4423680/4422102 [==============================] - 0s 0us/step

train_images.shape: (60000, 28, 28, 1), of float64
test_images.shape: (10000, 28, 28, 1), of float64

Train and evaluate your model

Let's use the simplest possible CNN, since we're not focused on the modeling part.

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 [==============================] - 13s 2ms/step - loss: 0.7546 - sparse_categorical_accuracy: 0.7457
Epoch 2/5
1875/1875 [==============================] - 3s 2ms/step - loss: 0.4254 - sparse_categorical_accuracy: 0.8521
Epoch 3/5
1875/1875 [==============================] - 3s 2ms/step - loss: 0.3812 - sparse_categorical_accuracy: 0.8668
Epoch 4/5
1875/1875 [==============================] - 3s 2ms/step - loss: 0.3557 - sparse_categorical_accuracy: 0.8770
Epoch 5/5
1875/1875 [==============================] - 3s 2ms/step - loss: 0.3415 - sparse_categorical_accuracy: 0.8795
313/313 [==============================] - 1s 2ms/step - loss: 0.3699 - sparse_categorical_accuracy: 0.8694

Test accuracy: 0.8694000244140625

Save your model

To load our trained model into TensorFlow Serving we first need to save it in SavedModel format. This will create a protobuf file in a well-defined directory hierarchy, and will include a version number. TensorFlow Serving allows us to select which version of a model, or "servable" we want to use when we make inference requests. Each version will be exported to a different sub-directory under the given path.

# 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 = tempfile.gettempdir()
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/1

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

Saved model:
total 88
drwxr-xr-x 2 kbuilder kbuilder  4096 Mar  9 10:10 assets
-rw-rw-r-- 1 kbuilder kbuilder 78123 Mar  9 10:10 saved_model.pb
drwxr-xr-x 2 kbuilder kbuilder  4096 Mar  9 10:10 variables

Examine your saved model

We'll use the command line utility saved_model_cli to look at the MetaGraphDefs (the models) and SignatureDefs (the methods you can call) in our SavedModel. See this discussion of the SavedModel CLI in the TensorFlow Guide.

saved_model_cli show --dir {export_path} --all
2021-03-09 10:10:12.685464: 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

  Function Name: '_default_save_signature'
    Option #1
      Callable with:
        Argument #1
          Conv1_input: TensorSpec(shape=(None, 28, 28, 1), dtype=tf.float32, name='Conv1_input')

  Function Name: 'call_and_return_all_conditional_losses'
    Option #1
      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 #2
      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
    Option #3
      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 #4
      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

That tells us a lot about our model! In this case we just trained our model, so we already know the inputs and outputs, but if we didn't this would be important information. It doesn't tell us everything, like the fact that this is grayscale image data for example, but it's a great start.

Serve your model with TensorFlow Serving

Add TensorFlow Serving distribution URI as a package source:

We're preparing to install TensorFlow Serving using Aptitude since this Colab runs in a Debian environment. We'll add the tensorflow-model-server package to the list of packages that Aptitude knows about. Note that we're running as root.

import sys
# We need sudo prefix if not on a Google Colab.
if 'google.colab' not in sys.modules:
  SUDO_IF_NEEDED = 'sudo'
else:
  SUDO_IF_NEEDED = ''
# This is the same as you would do from your command line, but without the [arch=amd64], and no sudo
# You would instead do:
# 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 -

!echo "deb http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal" | {SUDO_IF_NEEDED} tee /etc/apt/sources.list.d/tensorflow-serving.list && \
curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | {SUDO_IF_NEEDED} apt-key add -
!{SUDO_IF_NEEDED} apt update
deb http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  2943  100  2943    0     0  15822      0 --:--:-- --:--:-- --:--:-- 15822
OK
Hit:1 http://asia-east1.gce.archive.ubuntu.com/ubuntu bionic InRelease
Hit:2 http://asia-east1.gce.archive.ubuntu.com/ubuntu bionic-updates InRelease
Hit:3 http://asia-east1.gce.archive.ubuntu.com/ubuntu bionic-backports InRelease
Hit:4 http://packages.cloud.google.com/apt google-cloud-logging-wheezy InRelease
Get:5 https://packages.cloud.google.com/apt eip-cloud-bionic InRelease [5419 B]
Hit:6 https://nvidia.github.io/libnvidia-container/stable/ubuntu18.04/amd64  InRelease
Hit:7 https://nvidia.github.io/nvidia-container-runtime/ubuntu18.04/amd64  InRelease
Get:8 https://nvidia.github.io/nvidia-docker/ubuntu18.04/amd64  InRelease [1129 B]
Get:9 http://storage.googleapis.com/tensorflow-serving-apt stable InRelease [3012 B]
Hit:10 http://archive.canonical.com/ubuntu bionic InRelease
Hit:11 http://security.ubuntu.com/ubuntu bionic-security InRelease
Get:12 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server amd64 Packages [340 B]
Get:13 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server-universal amd64 Packages [348 B]
Fetched 10.2 kB in 1s (7051 B/s)



114 packages can be upgraded. Run 'apt list --upgradable' to see them.

Install TensorFlow Serving

This is all you need - one command line!

{SUDO_IF_NEEDED} apt-get install tensorflow-model-server
The following packages were automatically installed and are no longer required:
  adwaita-icon-theme ca-certificates-java dconf-gsettings-backend
  dconf-service default-jre default-jre-headless dkms fonts-dejavu-extra
  freeglut3 freeglut3-dev g++-6 glib-networking glib-networking-common
  glib-networking-services gsettings-desktop-schemas gtk-update-icon-cache
  hicolor-icon-theme humanity-icon-theme java-common libaccinj64-9.1
  libasound2 libasound2-data libasyncns0 libatk-bridge2.0-0
  libatk-wrapper-java libatk-wrapper-java-jni libatk1.0-0 libatk1.0-data
  libatspi2.0-0 libavahi-client3 libavahi-common-data libavahi-common3
  libcairo-gobject2 libcolord2 libcroco3 libcudart9.1 libcufft9.1 libcufftw9.1
  libcups2 libcurand9.1 libcusolver9.1 libcusparse9.1 libdconf1 libdrm-amdgpu1
  libdrm-dev libdrm-intel1 libdrm-nouveau2 libdrm-radeon1 libegl-mesa0 libegl1
  libegl1-mesa libepoxy0 libflac8 libfontenc1 libgbm1 libgdk-pixbuf2.0-0
  libgdk-pixbuf2.0-common libgif7 libgl1 libgl1-mesa-dev libgl1-mesa-dri
  libglapi-mesa libgles1 libgles2 libglu1-mesa libglu1-mesa-dev
  libglvnd-core-dev libglvnd-dev libglvnd0 libglx-mesa0 libglx0 libgtk-3-0
  libgtk-3-common libgtk2.0-0 libgtk2.0-common libice-dev libjansson4
  libjson-glib-1.0-0 libjson-glib-1.0-common liblcms2-2 libllvm9 libnppc9.1
  libnppial9.1 libnppicc9.1 libnppicom9.1 libnppidei9.1 libnppif9.1
  libnppig9.1 libnppim9.1 libnppist9.1 libnppisu9.1 libnppitc9.1 libnpps9.1
  libnvrtc9.1 libnvtoolsext1 libnvvm3 libogg0 libopengl0 libpciaccess0
  libpcsclite1 libproxy1v5 libpthread-stubs0-dev libpulse0 librest-0.7-0
  librsvg2-2 librsvg2-common libsensors4 libsm-dev libsndfile1
  libsoup-gnome2.4-1 libsoup2.4-1 libstdc++-6-dev libthrust-dev libvdpau-dev
  libvdpau1 libvorbis0a libvorbisenc2 libwayland-client0 libwayland-cursor0
  libwayland-egl1 libwayland-server0 libx11-dev libx11-xcb-dev libx11-xcb1
  libxau-dev libxcb-dri2-0 libxcb-dri2-0-dev libxcb-dri3-0 libxcb-dri3-dev
  libxcb-glx0 libxcb-glx0-dev libxcb-present-dev libxcb-present0 libxcb-randr0
  libxcb-randr0-dev libxcb-render0-dev libxcb-shape0 libxcb-shape0-dev
  libxcb-sync-dev libxcb-sync1 libxcb-xfixes0 libxcb-xfixes0-dev libxcb1-dev
  libxcomposite1 libxcursor1 libxdamage-dev libxdamage1 libxdmcp-dev
  libxext-dev libxfixes-dev libxfixes3 libxfont2 libxft2 libxi-dev libxi6
  libxinerama1 libxkbcommon0 libxkbfile1 libxmu-dev libxmu-headers libxnvctrl0
  libxrandr2 libxshmfence-dev libxshmfence1 libxt-dev libxtst6 libxv1
  libxxf86dga1 libxxf86vm-dev libxxf86vm1 linux-gcp-5.3-headers-5.3.0-1030
  linux-gcp-headers-5.0.0-1026 linux-headers-5.3.0-1030-gcp
  linux-image-5.3.0-1030-gcp linux-modules-5.3.0-1030-gcp
  linux-modules-extra-5.3.0-1030-gcp mesa-common-dev ocl-icd-libopencl1
  ocl-icd-opencl-dev opencl-c-headers openjdk-11-jre openjdk-11-jre-headless
  openjdk-8-jre openjdk-8-jre-headless pkg-config policykit-1-gnome
  python3-xkit screen-resolution-extra ubuntu-mono x11-utils x11-xkb-utils
  x11proto-core-dev x11proto-damage-dev x11proto-dev x11proto-fixes-dev
  x11proto-input-dev x11proto-xext-dev x11proto-xf86vidmode-dev
  xorg-sgml-doctools xserver-common xserver-xorg-core-hwe-18.04 xtrans-dev
Use 'sudo apt autoremove' to remove them.
The following NEW packages will be installed:
  tensorflow-model-server
0 upgraded, 1 newly installed, 0 to remove and 114 not upgraded.
Need to get 223 MB of archives.
After this operation, 0 B of additional disk space will be used.
Get:1 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server amd64 tensorflow-model-server all 2.4.1 [223 MB]
Fetched 223 MB in 6s (40.3 MB/s)
Selecting previously unselected package tensorflow-model-server.
(Reading database ... 242337 files and directories currently installed.)
Preparing to unpack .../tensorflow-model-server_2.4.1_all.deb ...
Unpacking tensorflow-model-server (2.4.1) ...
Setting up tensorflow-model-server (2.4.1) ...

Start running TensorFlow Serving

This is where we start running TensorFlow Serving and load our model. After it loads we can start making inference requests using REST. There are some important parameters:

  • rest_api_port: The port that you'll use for REST requests.
  • model_name: You'll use this in the URL of REST requests. It can be anything.
  • model_base_path: This is the path to the directory where you've saved your model.
os.environ["MODEL_DIR"] = MODEL_DIR
nohup tensorflow_model_server \
  --rest_api_port=8501 \
  --model_name=fashion_model \
  --model_base_path="${MODEL_DIR}" >server.log 2>&1
tail server.log

Make a request to your model in TensorFlow Serving

First, let's take a look at a random example from our test data.

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]]))

png

Ok, that looks interesting. How hard is that for you to recognize? Now let's create the JSON object for a batch of three inference requests, and see how well our model recognizes things:

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]]]]}

Make REST requests

Newest version of the servable

We'll send a predict request as a POST to our server's REST endpoint, and pass it three examples. We'll ask our server to give us the latest version of our servable by not specifying a particular version.

!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]))

A particular version of the servable

Now let's specify a particular version of our servable. Since we only have one, let's select version 1. We'll also look at all three results.

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]))

png

png

png