Missed TensorFlow Dev Summit? Check out the video playlist. Watch recordings

Train and serve a TensorFlow model with TensorFlow Serving

Warning: This notebook is designed to be run in a Google Colab only. It installs packages on the system and requires root access. If you want to run it in a local Jupyter notebook, please proceed with caution.

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

print('Installing TensorFlow')
try:
  # %tensorflow_version only exists in Colab.
  %tensorflow_version 2.x
except Exception:
  pass
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__))
Installing dependencies for Colab environment
     |████████████████████████████████| 2.4MB 4.6MB/s 
[?25hInstalling TensorFlow
TensorFlow 2.x selected.
TensorFlow version: 2.1.0-rc1

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, activation=tf.nn.softmax, name='Softmax')
])
model.summary()

testing = False
epochs = 5

model.compile(optimizer='adam', 
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])
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         
_________________________________________________________________
Softmax (Dense)              (None, 10)                13530     
=================================================================
Total params: 13,610
Trainable params: 13,610
Non-trainable params: 0
_________________________________________________________________
Train on 60000 samples
Epoch 1/5
60000/60000 [==============================] - 11s 185us/sample - loss: 0.5466 - accuracy: 0.8087
Epoch 2/5
60000/60000 [==============================] - 5s 79us/sample - loss: 0.4032 - accuracy: 0.8580
Epoch 3/5
60000/60000 [==============================] - 5s 76us/sample - loss: 0.3613 - accuracy: 0.8712
Epoch 4/5
60000/60000 [==============================] - 5s 75us/sample - loss: 0.3406 - accuracy: 0.8797
Epoch 5/5
60000/60000 [==============================] - 4s 75us/sample - loss: 0.3247 - accuracy: 0.8848
10000/10000 [==============================] - 1s 73us/sample - loss: 0.3510 - accuracy: 0.8747

Test accuracy: 0.8747000098228455

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

WARNING:tensorflow:From /tensorflow-2.1.0/python3.6/tensorflow_core/python/ops/resource_variable_ops.py:1786: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.
Instructions for updating:
If using Keras pass *_constraint arguments to layers.
INFO:tensorflow:Assets written to: /tmp/1/assets

Saved model:
total 84
drwxr-xr-x 2 root root  4096 Jan  7 23:15 assets
-rw-r--r-- 1 root root 74086 Jan  7 23:15 saved_model.pb
drwxr-xr-x 2 root root  4096 Jan  7 23:15 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

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['Softmax'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 10)
        name: StatefulPartitionedCall:0
  Method name is: tensorflow/serving/predict
WARNING:tensorflow:From /tensorflow-2.1.0/python3.6/tensorflow_core/python/ops/resource_variable_ops.py:1786: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.
Instructions for updating:
If using Keras pass *_constraint arguments to layers.

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: True
        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: True
        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: False
        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: False
        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: True
        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
          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
          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

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.

# 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" | tee /etc/apt/sources.list.d/tensorflow-serving.list && \
curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | apt-key add -
!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  11496      0 --:--:-- --:--:-- --:--:-- 11496
OK
Get:1 http://storage.googleapis.com/tensorflow-serving-apt stable InRelease [3,012 B]
Get:2 https://cloud.r-project.org/bin/linux/ubuntu bionic-cran35/ InRelease [3,626 B]
Ign:3 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64  InRelease
Ign:4 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  InRelease
Hit:5 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64  Release
Get:6 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  Release [564 B]
Get:7 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  Release.gpg [833 B]
Hit:8 http://ppa.launchpad.net/graphics-drivers/ppa/ubuntu bionic InRelease
Hit:9 http://archive.ubuntu.com/ubuntu bionic InRelease
Get:10 http://security.ubuntu.com/ubuntu bionic-security InRelease [88.7 kB]
Get:11 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server amd64 Packages [354 B]
Get:12 https://cloud.r-project.org/bin/linux/ubuntu bionic-cran35/ Packages [81.6 kB]
Get:13 http://archive.ubuntu.com/ubuntu bionic-updates InRelease [88.7 kB]
Get:14 http://storage.googleapis.com/tensorflow-serving-apt stable/tensorflow-model-server-universal amd64 Packages [364 B]
Get:15 http://ppa.launchpad.net/marutter/c2d4u3.5/ubuntu bionic InRelease [15.4 kB]
Get:17 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  Packages [30.4 kB]
Get:18 http://archive.ubuntu.com/ubuntu bionic-backports InRelease [74.6 kB]
Get:19 http://ppa.launchpad.net/marutter/c2d4u3.5/ubuntu bionic/main Sources [1,749 kB]
Get:20 http://security.ubuntu.com/ubuntu bionic-security/universe amd64 Packages [796 kB]
Get:21 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 Packages [1,073 kB]
Get:22 http://security.ubuntu.com/ubuntu bionic-security/main amd64 Packages [776 kB]
Get:23 http://security.ubuntu.com/ubuntu bionic-security/restricted amd64 Packages [21.3 kB]
Get:24 http://archive.ubuntu.com/ubuntu bionic-updates/multiverse amd64 Packages [10.8 kB]
Get:25 http://archive.ubuntu.com/ubuntu bionic-updates/universe amd64 Packages [1,324 kB]
Get:26 http://archive.ubuntu.com/ubuntu bionic-updates/restricted amd64 Packages [35.5 kB]
Get:27 http://ppa.launchpad.net/marutter/c2d4u3.5/ubuntu bionic/main amd64 Packages [844 kB]
Fetched 7,019 kB in 4s (1,913 kB/s)
Reading package lists... Done
Building dependency tree       
Reading state information... Done
21 packages can be upgraded. Run 'apt list --upgradable' to see them.

Install TensorFlow Serving

This is all you need - one command line!

apt-get install tensorflow-model-server
Reading package lists... Done
Building dependency tree       
Reading state information... Done
The following package was automatically installed and is no longer required:
  libnvidia-common-430
Use 'apt autoremove' to remove it.
The following NEW packages will be installed:
  tensorflow-model-server
0 upgraded, 1 newly installed, 0 to remove and 21 not upgraded.
Need to get 140 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 2.0.0 [140 MB]
Fetched 140 MB in 2s (78.8 MB/s)
Selecting previously unselected package tensorflow-model-server.
(Reading database ... 145674 files and directories currently installed.)
Preparing to unpack .../tensorflow-model-server_2.0.0_all.deb ...
Unpacking tensorflow-model-server (2.0.0) ...
Setting up tensorflow-model-server (2.0.0) ...

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
%%bash --bg 
nohup tensorflow_model_server \
  --rest_api_port=8501 \
  --model_name=fashion_model \
  --model_base_path="${MODEL_DIR}" >server.log 2>&1

Starting job # 0 in a separate thread.
tail server.log
[warn] getaddrinfo: address family for nodename not supported
[evhttp_server.cc : 238] NET_LOG: Entering the event loop ...

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]))

png

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