DeepSite - 🧬 RemixA step-by-step guide to training and deploying image classification models using Google Vertex AI AutoML Vision
This tutorial is designed for developers, data scientists, and students who want to learn how to build image classification models without deep machine learning expertise.
We'll use Google Vertex AI's AutoML Vision, which automates much of the model training process while still delivering high-quality results. No need to write complex neural network architectures!
By the end of this guide, you'll be able to:
You'll need a Google Cloud account with billing enabled. Vertex AI is a paid service, but new users get $300 in free credits.
Create a new project or select an existing one in the Google Cloud Console where you'll enable the Vertex AI API.
Enable the Vertex AI API for your project. This can be done in the "APIs & Services" section of the Cloud Console.
Create a Cloud Storage bucket to store your training data. The bucket should be in the same region where you'll train your model.
Set up a Python environment (3.7+) with the Google Cloud SDK installed. We recommend using a virtual environment.
Set up authentication by creating a service account and downloading the JSON key file. Set the GOOGLE_APPLICATION_CREDENTIALS environment variable.
Install the Google Cloud Vertex AI SDK and other required packages:
pip install google-cloud-aiplatform pandasFor image classification with AutoML Vision, your dataset needs to be structured in a specific way:
gs://your-bucket-name/
├── train/
│ ├── class1/
│ │ ├── image1.jpg
│ │ ├── image2.jpg
│ │ └── ...
│ ├── class2/
│ │ ├── image1.jpg
│ │ ├── image2.jpg
│ │ └── ...
│ └── ...
└── test/
├── class1/
├── class2/
└── ...Use the Google Cloud Console or gsutil command-line tool to upload your dataset:
gsutil -m cp -r /path/to/local/dataset gs://your-bucket-nameNow we'll create a dataset resource in Vertex AI that points to your Cloud Storage data.
from google.cloud import aiplatform
# Initialize the Vertex AI client
aiplatform.init(project="your-project-id", location="us-central1")
# Create an image dataset
dataset = aiplatform.ImageDataset.create(
display_name="flowers-classification",
gcs_source="gs://your-bucket-name/train/**",
import_schema_uri=aiplatform.schema.dataset.ioformat.image.classification.single_label,
)
print(f"Created dataset: {dataset.resource_name}")With your dataset ready, you can now train an AutoML Vision model. This process will automatically:
# Define training job
training_job = aiplatform.AutoMLImageTrainingJob(
display_name="train-flowers-classification",
prediction_type="classification",
multi_label=False,
model_type="CLOUD",
)
# Run the training job
model = training_job.run(
dataset=dataset,
training_fraction_split=0.8,
validation_fraction_split=0.1,
test_fraction_split=0.1,
budget_milli_node_hours=8000, # 8 compute hours
disable_early_stopping=False,
)
print(f"Training completed. Model: {model.resource_name}")After training completes, you'll want to evaluate the model's performance before deployment.
# Get evaluation metrics
evaluation = model.evaluate()
print("Model evaluation metrics:")
print(f"Precision: {evaluation.metrics['precision']}")
print(f"Recall: {evaluation.metrics['recall']}")
print(f"F1 Score: {evaluation.metrics['f1Score']}")
print(f"Confusion Matrix: {evaluation.metrics['confusionMatrix']}")For a more visual evaluation, check the "Evaluate" tab in the Vertex AI Console where you can see:
To make predictions, you need to deploy your model to an endpoint. This creates a scalable service that can handle prediction requests.
# Create an endpoint
endpoint = aiplatform.Endpoint.create(
display_name="flowers-classification-endpoint",
project="your-project-id",
location="us-central1",
)
# Deploy the model to the endpoint
endpoint.deploy(
model=model,
deployed_model_display_name="flowers-classification-model",
traffic_percentage=100,
machine_type="n1-standard-4", # Choose appropriate machine type
min_replica_count=1,
max_replica_count=1,
)
print(f"Model deployed to endpoint: {endpoint.resource_name}")With your model deployed to an endpoint, you can now make predictions on new images.
import base64
# Function to encode image
def encode_image(image_path):
with open(image_path, "rb") as image_file:
return base64.b64encode(image_file.read()).decode("utf-8")
# Example prediction
image_path = "path/to/your/test_image.jpg"
encoded_image = encode_image(image_path)
# Make prediction
prediction = endpoint.predict(
instances=[{"content": encoded_image}],
parameters={"confidenceThreshold": 0.5}, # Minimum confidence score
)
# Process results
for result in prediction.predictions:
print("Predicted classes:")
for i, (label, score) in enumerate(zip(result["displayNames"], result["confidences"])):
print(f"{i+1}. {label}: {score:.2%}")For predicting on many images at once, use batch prediction:
# Create batch prediction job
batch_job = model.batch_predict(
job_display_name="batch-pred-flowers",
gcs_source="gs://your-bucket-name/test/**",
gcs_destination_prefix="gs://your-bucket-name/predictions/",
instances_format="jsonl",
predictions_format="jsonl",
)
print(f"Batch prediction job: {batch_job.resource_name}")