model_garden/notebooks/official/custom/sdk-custom-image-classification-online.ipynb

In [ ]:

# Copyright 2020 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

Custom training and online prediction

Open in Colab

Open in Colab Enterprise

Open in Workbench

View on GitHub

Overview

This tutorial demonstrates how to use the Vertex AI SDK for Python to train and deploy a custom image classification model for online prediction.

Learn more about Custom training and Vertex AI Prediction.

Objective

In this tutorial, you learn to use Vertex AI Training to create a custom-trained model from a Python script in a Docker container, and learn to use Vertex AI Prediction to do a prediction on the deployed model by sending data. Alternatively, you can create custom-trained models using gcloud command-line tool, or online using the Cloud Console.

This tutorial uses the following Google Cloud ML services:

• Vertex AI Training
• Vertex AI Prediction
• Vertex AI Model resource
• Vertex AI Endpoint resource

The steps performed include:

• Create a Vertex AI custom job for training a TensorFlow model.
• Upload the trained model artifacts to a Model resource.
• Create a serving Endpoint resource.
• Deploy the Model resource to a serving Endpoint resource.
• Make a prediction.
• Undeploy the Model resource.

Dataset

The dataset used for this tutorial is the cifar10 dataset from TensorFlow Datasets. The version of the dataset you will use is built into TensorFlow. The trained model predicts which type of class an image is from ten classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck.

Costs

This tutorial uses billable components of Google Cloud (GCP):

• Vertex AI
• Cloud Storage

Learn about Vertex AI pricing and Cloud Storage pricing, and use the Pricing Calculator to generate a cost estimate based on your projected usage.

Get started

Install Vertex AI SDK for Python and other required packages

In [ ]:

%pip install --upgrade --quiet google-cloud-aiplatform \
                                 google-cloud \
                                 google-cloud-storage \
                                 pillow \
                                 numpy

Authenticate your notebook environment (Colab only)

Authenticate your environment on Google Colab.

In [ ]:

import sys

if "google.colab" in sys.modules:

    from google.colab import auth

    auth.authenticate_user()

Set Google Cloud project information and initialize Vertex AI SDK for Python

To get started using Vertex AI, you must have an existing Google Cloud project and enable the Vertex AI API. Learn more about setting up a project and a development environment.

In [ ]:

PROJECT_ID = "[your-project-id]"  # @param {type:"string"}
LOCATION = "us-central1"  # @param {type:"string"}

Create a Cloud Storage bucket

Create a storage bucket to store intermediate artifacts such as datasets.

In [ ]:

BUCKET_URI = f"gs://your-bucket-name-{PROJECT_ID}-unique"  # @param {type:"string"}

If your bucket doesn't already exist: Run the following cell to create your Cloud Storage bucket.

In [ ]:

! gsutil mb -l $LOCATION -p $PROJECT_ID $BUCKET_URI

Initialize Vertex AI SDK for Python

Initialize the Vertex AI SDK for Python for your project and corresponding bucket.

In [ ]:

from google.cloud import aiplatform

aiplatform.init(project=PROJECT_ID, location=LOCATION, staging_bucket=BUCKET_URI)

Import libraries and define constants

In [ ]:

import os

import numpy as np
from PIL import Image

Set pre-built containers

Vertex AI provides pre-built containers to run training and prediction.

For the latest list, see Pre-built containers for training and Pre-built containers for prediction

In [ ]:

TRAIN_VERSION = "tf-cpu.2-9"
DEPLOY_VERSION = "tf2-cpu.2-9"

TRAIN_IMAGE = "us-docker.pkg.dev/vertex-ai/training/{}:latest".format(TRAIN_VERSION)
DEPLOY_IMAGE = "us-docker.pkg.dev/vertex-ai/prediction/{}:latest".format(DEPLOY_VERSION)

Tutorial

Now you are ready to start creating your own custom-trained model with CIFAR10.

Train a model

There are two ways you can train a custom model using a container image:

• Use a Google Cloud prebuilt container. If you use a prebuilt container, you will additionally specify a Python package to install into the container image. This Python package contains your code for training a custom model.

• Use your own custom container image. If you use your own container, the container needs to contain your code for training a custom model.

Define the command args for the training script

Prepare the command-line arguments to pass to your training script.

• args: The command line arguments to pass to the corresponding Python module. In this example, they will be:
  • "--epochs=" + EPOCHS: The number of epochs for training.
  • "--steps=" + STEPS: The number of steps (batches) per epoch.
  • "--distribute=" + TRAIN_STRATEGY" : The training distribution strategy to use for single or distributed training.
    • "single": single device.
    • "mirror": all GPU devices on a single compute instance.
    • "multi": all GPU devices on all compute instances.

In [ ]:

JOB_NAME = "custom_job_unique"
MODEL_DIR = "{}/{}".format(BUCKET_URI, JOB_NAME)

TRAIN_STRATEGY = "single"

EPOCHS = 20
STEPS = 100

CMDARGS = [
    "--epochs=" + str(EPOCHS),
    "--steps=" + str(STEPS),
    "--distribute=" + TRAIN_STRATEGY,
]

Training script

In the next cell, you will write the contents of the training script, task.py. In summary:

• Get the directory where to save the model artifacts from the environment variable AIP_MODEL_DIR. This variable is set by the training service.
• Loads CIFAR10 dataset from TF Datasets (tfds).
• Builds a model using TF.Keras model API.
• Compiles the model (compile()).
• Sets a training distribution strategy according to the argument args.distribute.
• Trains the model (fit()) with epochs and steps according to the arguments args.epochs and args.steps
• Saves the trained model (save(MODEL_DIR)) to the specified model directory.

In [ ]:

%%writefile task.py
# Single, Mirror and Multi-Machine Distributed Training for CIFAR-10

import tensorflow_datasets as tfds
import tensorflow as tf
from tensorflow.python.client import device_lib
import argparse
import os
import sys
tfds.disable_progress_bar()

parser = argparse.ArgumentParser()
parser.add_argument('--lr', dest='lr',
                    default=0.01, type=float,
                    help='Learning rate.')
parser.add_argument('--epochs', dest='epochs',
                    default=10, type=int,
                    help='Number of epochs.')
parser.add_argument('--steps', dest='steps',
                    default=200, type=int,
                    help='Number of steps per epoch.')
parser.add_argument('--distribute', dest='distribute', type=str, default='single',
                    help='distributed training strategy')
args = parser.parse_args()

print('Python Version = {}'.format(sys.version))
print('TensorFlow Version = {}'.format(tf.__version__))
print('TF_CONFIG = {}'.format(os.environ.get('TF_CONFIG', 'Not found')))
print('DEVICES', device_lib.list_local_devices())

# Single Machine, single compute device
if args.distribute == 'single':
    if tf.test.is_gpu_available():
        strategy = tf.distribute.OneDeviceStrategy(device="/gpu:0")
    else:
        strategy = tf.distribute.OneDeviceStrategy(device="/cpu:0")
# Single Machine, multiple compute device
elif args.distribute == 'mirror':
    strategy = tf.distribute.MirroredStrategy()
# Multiple Machine, multiple compute device
elif args.distribute == 'multi':
    strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()

# Multi-worker configuration
print('num_replicas_in_sync = {}'.format(strategy.num_replicas_in_sync))

# Preparing dataset
BUFFER_SIZE = 10000
BATCH_SIZE = 64

def make_datasets_unbatched():
  # Scaling CIFAR10 data from (0, 255] to (0., 1.]
  def scale(image, label):
    image = tf.cast(image, tf.float32)
    image /= 255.0
    return image, label

  datasets, info = tfds.load(name='cifar10',
                            with_info=True,
                            as_supervised=True)
  return datasets['train'].map(scale).cache().shuffle(BUFFER_SIZE).repeat()


# Build the Keras model
def build_and_compile_cnn_model():
  model = tf.keras.Sequential([
      tf.keras.layers.Conv2D(32, 3, activation='relu', input_shape=(32, 32, 3)),
      tf.keras.layers.MaxPooling2D(),
      tf.keras.layers.Conv2D(32, 3, activation='relu'),
      tf.keras.layers.MaxPooling2D(),
      tf.keras.layers.Flatten(),
      tf.keras.layers.Dense(10, activation='softmax')
  ])
  model.compile(
      loss=tf.keras.losses.sparse_categorical_crossentropy,
      optimizer=tf.keras.optimizers.SGD(learning_rate=args.lr),
      metrics=['accuracy'])
  return model

# Train the model
NUM_WORKERS = strategy.num_replicas_in_sync
# Here the batch size scales up by number of workers since
# `tf.data.Dataset.batch` expects the global batch size.
GLOBAL_BATCH_SIZE = BATCH_SIZE * NUM_WORKERS
MODEL_DIR = os.getenv("AIP_MODEL_DIR")

train_dataset = make_datasets_unbatched().batch(GLOBAL_BATCH_SIZE)

with strategy.scope():
  # Creation of dataset, and model building/compiling need to be within
  # `strategy.scope()`.
  model = build_and_compile_cnn_model()

model.fit(x=train_dataset, epochs=args.epochs, steps_per_epoch=args.steps)
model.save(MODEL_DIR)

Train the model

Define your custom training job on Vertex AI.

Use the CustomTrainingJob class to define the job, which takes the following parameters:

• display_name: The user-defined name of this training pipeline.
• script_path: The local path to the training script.
• container_uri: The URI of the training container image.
• requirements: The list of Python package dependencies of the script.
• model_serving_container_image_uri: The URI of a container that can serve predictions for your model — either a prebuilt container or a custom container.

Use the run function to start training, which takes the following parameters:

• args: The command line arguments to be passed to the Python script.
• replica_count: The number of worker replicas.
• model_display_name: The display name of the Model if the script produces a managed Model.
• machine_type: The type of machine to use for training.
• accelerator_type: The hardware accelerator type.
• accelerator_count: The number of accelerators to attach to a worker replica.

The run function creates a training pipeline that trains and creates a Model object. After the training pipeline completes, the run function returns the Model object.

In [ ]:

job = aiplatform.CustomTrainingJob(
    display_name=JOB_NAME,
    script_path="task.py",
    container_uri=TRAIN_IMAGE,
    requirements=["tensorflow_datasets"],
    model_serving_container_image_uri=DEPLOY_IMAGE,
)

MODEL_DISPLAY_NAME = "cifar10_unique"

# Start the training
model = job.run(
    model_display_name=MODEL_DISPLAY_NAME,
    args=CMDARGS,
    replica_count=1,
)

Deploy the model

Before you use your model to make predictions, you need to deploy it to an Endpoint. You can do this by calling the deploy function on the Model resource. This will do two things:

1. Create an Endpoint resource for deploying the Model resource to.
2. Deploy the Model resource to the Endpoint resource.

The function takes the following parameters:

• deployed_model_display_name: A human readable name for the deployed model.
• traffic_split: Percent of traffic at the endpoint that goes to this model, which is specified as a dictionary of one or more key/value pairs.
  • If only one model, then specify as { "0": 100 }, where "0" refers to this model being uploaded and 100 means 100% of the traffic.
  • If there are existing models on the endpoint, for which the traffic will be split, then use model_id to specify as { "0": percent, model_id: percent, ... }, where model_id is the model id of an existing model to the deployed endpoint. The percents must add up to 100.
• machine_type: The type of machine to use for training.
• accelerator_type: The hardware accelerator type.
• accelerator_count: The number of accelerators to attach to a worker replica.
• starting_replica_count: The number of compute instances to initially provision.
• max_replica_count: The maximum number of compute instances to scale to. In this tutorial, only one instance is provisioned.

Traffic split

The traffic_split parameter is specified as a Python dictionary. You can deploy more than one instance of your model to an endpoint, and then set the percentage of traffic that goes to each instance.

You can use a traffic split to introduce a new model gradually into production. For example, if you had one existing model in production with 100% of the traffic, you could deploy a new model to the same endpoint, direct 10% of traffic to it, and reduce the original model's traffic to 90%. This allows you to monitor the new model's performance while minimizing the distruption to the majority of users.

Compute instance scaling

You can specify a single instance (or node) to serve your online prediction requests. This tutorial uses a single node, so the variables MIN_NODES and MAX_NODES are both set to 1.

If you want to use multiple nodes to serve your online prediction requests, set MAX_NODES to the maximum number of nodes you want to use. Vertex AI autoscales the number of nodes used to serve your predictions, up to the maximum number you set. Refer to the pricing page to understand the costs of autoscaling with multiple nodes.

Endpoint

The method will block until the model is deployed and eventually return an Endpoint object. If this is the first time a model is deployed to the endpoint, it may take a few additional minutes to complete provisioning of resources.

In [ ]:

DEPLOYED_NAME = "cifar10_deployed_unique"

TRAFFIC_SPLIT = {"0": 100}

MIN_NODES = 1
MAX_NODES = 1


endpoint = model.deploy(
    deployed_model_display_name=DEPLOYED_NAME,
    traffic_split=TRAFFIC_SPLIT,
    min_replica_count=MIN_NODES,
    max_replica_count=MAX_NODES,
)

Make an online prediction request

Send an online prediction request to your deployed model.

Get test data

Download images from the CIFAR dataset and preprocess them.

Download the test images

Download the provided set of images from the CIFAR dataset:

In [ ]:

# Download the images
! gsutil -m cp -r gs://cloud-samples-data/ai-platform-unified/cifar_test_images .

Preprocess the images

Before you can run the data through the endpoint, you need to preprocess it to match the format that your custom model defined in task.py expects.

x_test: Normalize (rescale) the pixel data by dividing each pixel by 255. This replaces each single byte integer pixel with a 32-bit floating point number between 0 and 1.

y_test: You can extract the labels from the image filenames. Each image's filename format is "image_{LABEL}_{IMAGE_NUMBER}.jpg"

In [ ]:

# Load image data
IMAGE_DIRECTORY = "cifar_test_images"

image_files = [file for file in os.listdir(IMAGE_DIRECTORY) if file.endswith(".jpg")]

# Decode JPEG images into numpy arrays
image_data = [
    np.asarray(Image.open(os.path.join(IMAGE_DIRECTORY, file))) for file in image_files
]

# Scale and convert to expected format
x_test = [(image / 255.0).astype(np.float32).tolist() for image in image_data]

# Extract labels from image name
y_test = [int(file.split("_")[1]) for file in image_files]

Send the prediction request

Now that you have test images, you can use them to send a prediction request. Use the Endpoint object's predict function, which takes the following parameters:

• instances: A list of image instances. According to your custom model, each image instance should be a 3-dimensional matrix of floats. This was prepared in the previous step.

The predict function returns a list, where each element in the list corresponds to the corresponding image in the request. You will see in the output for each prediction:

• Confidence level for the prediction (predictions), between 0 and 1, for each of the ten classes.

You can then run a quick evaluation on the prediction results:

1. np.argmax: Convert each list of confidence levels to a label
2. Compare the predicted labels to the actual labels
3. Calculate accuracy as correct/total

In [ ]:

predictions = endpoint.predict(instances=x_test)
y_predicted = np.argmax(predictions.predictions, axis=1)

correct = sum(y_predicted == np.array(y_test))
accuracy = len(y_predicted)
print(
    f"Correct predictions = {correct}, Total predictions = {accuracy}, Accuracy = {correct/accuracy}"
)

Undeploy the model

To undeploy your Model resource from the serving Endpoint resource, use the endpoint's undeploy method with the following parameter:

• deployed_model_id: The model deployment identifier returned by the endpoint service when the Model resource was deployed. You can retrieve the deployed models using the endpoint's deployed_models property.

Since this is the only deployed model on the Endpoint resource, you can omit traffic_split.

In [ ]:

deployed_model_id = endpoint.list_models()[0].id
endpoint.undeploy(deployed_model_id=deployed_model_id)

Cleaning up

To clean up all Google Cloud resources used in this project, you can delete the Google Cloud project you used for the tutorial.

Otherwise, you can delete the individual resources you created in this tutorial:

• Training Job
• Model
• Endpoint
• Cloud Storage Bucket

In [ ]:

# Warning: Setting this to true will delete everything in your bucket
delete_bucket = False

# Delete the training job
job.delete()

# Delete the model
model.delete()

# Delete the endpoint
endpoint.delete()

if delete_bucket:
    ! gsutil -m rm -r $BUCKET_URI