Tutorial Lengkap Vertex AI: Platform ML Terpadu di Google Cloud

Vertex AI adalah platform machine learning terpadu Google Cloud yang menggabungkan semua layanan ML Google Cloud. Platform ini menyediakan tools untuk membangun, mendeploy, dan menskalakan model ML dengan AutoML dan custom training.

Mengapa Vertex AI?

• Platform terpadu: Semua tools ML dalam satu tempat
• AutoML: Pembuatan model tanpa kode
• Custom training: Kontrol penuh dengan kode custom
• MLOps: Built-in pipelines dan monitoring
• Scalable: Infrastruktur enterprise-grade

• Datasets
• Training (AutoML dan Custom)
• Model Registry
• Endpoints
• Pipelines
• Feature Store
• Experiments

Prerequisites

pip install google-cloud-aiplatform

Autentikasi
gcloud auth login
gcloud config set project your-project-id

Setup

1. Inisialisasi Vertex AI

from google.cloud import aiplatform

aiplatform.init(
    project="your-project-id",
    location="us-central1",
    stagingbucket="gs://your-bucket"

)

gcloud services enable aiplatform.googleapis.com
gcloud services enable compute.googleapis.com
gcloud services enable storage.googleapis.com

from google.cloud import aiplatform

Buat dari BigQuery
dataset = aiplatform.TabularDataset.create(
    displayname="customer-churn-dataset",
    bqsource="bq://project.dataset.table"

)

Buat dari GCS
dataset = aiplatform.TabularDataset.create(
    displayname="customer-churn-dataset",
    gcssource="gs://bucket/data/train.csv"

)

print(f"Dataset dibuat: {dataset.resourcename}")

2. Buat Image Dataset

# Buat image dataset
imagedataset = aiplatform.ImageDataset.create(

    displayname="product-images",
    gcssource="gs://bucket/images/",

    importschemauri=aiplatform.schema.dataset.ioformat.image.singlelabelclassification

)

# Buat text dataset
textdataset = aiplatform.TextDataset.create(
    displayname="sentiment-dataset",

    gcssource="gs://bucket/text/data.jsonl",
    importschemauri=aiplatform.schema.dataset.ioformat.text.singlelabelclassification
)

AutoML Training

1. AutoML Tabular

# Buat AutoML tabular training job
job = aiplatform.AutoMLTabularTrainingJob(
    displayname="churn-automl",

    optimizationpredictiontype="classification",

    optimizationobjective="maximize-au-roc"
)

Train model
model = job.run(
    dataset=dataset,
    targetcolumn="churn",

    trainingfractionsplit=0.8,

    validationfractionsplit=0.1,

    testfractionsplit=0.1,

    budgetmillinodehours=1000,
    modeldisplayname="churn-model"
)

print(f"Model ditraining: {model.resourcename}")

# Buat AutoML image training job
job = aiplatform.AutoMLImageTrainingJob(
    displayname="image-classifier",
    predictiontype="classification",

    multilabel=False
)

Train model
model = job.run(
    dataset=imagedataset,

    trainingfractionsplit=0.8,

    validationfractionsplit=0.1,

    testfractionsplit=0.1,

    budgetmillinodehours=8000,
    modeldisplayname="product-classifier"
)

3. AutoML Text Classification

# Buat AutoML text training job
job = aiplatform.AutoMLTextTrainingJob(
    displayname="sentiment-classifier",

    predictiontype="classification",
    multilabel=False

)

Train model
model = job.run(
    dataset=textdataset,
    trainingfractionsplit=0.8,
    validationfractionsplit=0.1,
    testfractionsplit=0.1,
    modeldisplayname="sentiment-model"
)

Custom Training

1. Custom Training Job

from google.cloud import aiplatform

Definisikan custom training job
job = aiplatform.CustomTrainingJob(
    displayname="custom-sklearn-training",

    scriptpath="train.py",
    containeruri="us-docker.pkg.dev/vertex-ai/training/sklearn-cpu.1-0:latest",

    requirements=["pandas", "scikit-learn"],
    modelservingcontainerimageuri="us-docker.pkg.dev/vertex-ai/prediction/sklearn-cpu.1-0:latest"

)

Jalankan training
model = job.run(
    dataset=dataset,
    modeldisplayname="sklearn-model",

    machinetype="n1-standard-4",
    replicacount=1,

    args=["--epochs", "100", "--batch-size", "32"]
)

# train.py
import argparse
import os
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.modelselection import traintestsplit
from sklearn.metrics import accuracyscore

import joblib
from google.cloud import storage

def main():
    parser = argparse.ArgumentParser()
    parser.addargument("--epochs", type=int, default=100)
    parser.addargument("--batch-size", type=int, default=32)

    args = parser.parseargs()

    # Load data dari environment variable
    trainingdatauri = os.environ.get("AIPTRAININGDATAURI")

    df = pd.readcsv(trainingdatauri)

    X = df.drop("target", axis=1)
    y = df["target"]

    Xtrain, Xtest, ytrain, ytest = traintestsplit(
        X, y, testsize=0.2, randomstate=42
    )

    # Train model
    model = RandomForestClassifier(nestimators=args.epochs)

    model.fit(Xtrain, ytrain)


    # Evaluasi
    predictions = model.predict(Xtest)
    accuracy = accuracyscore(ytest, predictions)
    print(f"Accuracy: {accuracy}")

    # Simpan model
    modeldir = os.environ.get("AIPMODELDIR")

    joblib.dump(model, os.path.join(modeldir, "model.joblib"))

if name == "main":
    main()

3. Custom Container Training

# Definisikan custom container training job
job = aiplatform.CustomContainerTrainingJob(
    displayname="pytorch-training",

    containeruri="gcr.io/your-project/pytorch-training:latest",
    modelservingcontainerimageuri="gcr.io/your-project/pytorch-serving:latest"
)

Jalankan training
model = job.run(
    dataset=dataset,
    modeldisplayname="pytorch-model",
    machinetype="n1-standard-8",

    acceleratortype="NVIDIATESLAV100",
    acceleratorcount=1,

    replicacount=1
)

4. Dockerfile untuk Custom Container

FROM pytorch/pytorch:1.9.0-cuda10.2-cudnn7-runtime

WORKDIR /app

COPY requirements.txt .
RUN pip install -r requirements.txt

COPY train.py .

ENTRYPOINT ["python", "train.py"]

Hyperparameter Tuning

1. Definisikan Hyperparameter Tuning Job

from google.cloud import aiplatform
from google.cloud.aiplatform import hyperparametertuning as hpt


Definisikan hyperparameter spec
parameterspec = {
    "learningrate": hpt.DoubleParameterSpec(min=0.001, max=0.1, scale="log"),

    "numlayers": hpt.DiscreteParameterSpec(values=[2, 4, 6, 8], scale="linear"),
    "dropout": hpt.DoubleParameterSpec(min=0.0, max=0.5, scale="linear")
}

Definisikan metric spec
metricspec = {"accuracy": "maximize"}


Buat hyperparameter tuning job
job = aiplatform.HyperparameterTuningJob(
    displayname="hpt-training",
    customjob=aiplatform.CustomJob(

        displayname="hpt-custom-job",
        workerpoolspecs=[{
            "machinespec": {"machinetype": "n1-standard-4"},
            "replicacount": 1,

            "containerspec": {
                "imageuri": "gcr.io/your-project/training:latest",

                "args": []
            }
        }]
    ),
    metricspec=metricspec,

    parameterspec=parameterspec,

    maxtrialcount=20,

    paralleltrialcount=4,

    searchalgorithm="random"
)

Jalankan job
job.run()

Model Registry

1. Upload Model

# Upload model ke registry
model = aiplatform.Model.upload(
    displayname="sklearn-model",

    artifacturi="gs://bucket/models/sklearn/",
    servingcontainerimageuri="us-docker.pkg.dev/vertex-ai/prediction/sklearn-cpu.1-0:latest"

)

print(f"Model diupload: {model.resourcename}")

2. List Models

# List semua models
models = aiplatform.Model.list()
for m in models:
    print(f"{m.displayname}: {m.resourcename}")

Dapatkan model spesifik
model = aiplatform.Model("projects/123/locations/us-central1/models/456")
print(f"Model: {model.displayname}")

# Upload versi baru
newversion = aiplatform.Model.upload(
    displayname="sklearn-model",

    artifacturi="gs://bucket/models/sklearn-v2/",
    servingcontainerimageuri="us-docker.pkg.dev/vertex-ai/prediction/sklearn-cpu.1-0:latest",

    parentmodel=model.resourcename

)

# Buat endpoint
endpoint = aiplatform.Endpoint.create(
    displayname="prediction-endpoint",
    description="Production endpoint untuk predictions"
)

print(f"Endpoint dibuat: {endpoint.resourcename}")

# Deploy model ke endpoint
endpoint.deploy(
    model=model,
    deployedmodeldisplayname="sklearn-deployed",
    machinetype="n1-standard-4",

    minreplicacount=1,

    maxreplicacount=5,

    trafficpercentage=100,
    sync=True
)

print("Model dideploy")

3. Buat Predictions

# Online prediction
instances = [
    {"feature1": 1.0, "feature2": 2.0, "feature3": 3.0},
    {"feature1": 4.0, "feature2": 5.0, "feature3": 6.0}
]

predictions = endpoint.predict(instances=instances)
print(f"Predictions: {predictions.predictions}")

4. Batch Prediction

# Buat batch prediction job
batchjob = model.batchpredict(
    jobdisplayname="batch-prediction",
    gcssource="gs://bucket/batch-input/",

    gcsdestinationprefix="gs://bucket/batch-output/",

    machinetype="n1-standard-4",
    startingreplicacount=2,
    maxreplicacount=10
)

batchjob.wait()

print(f"Batch prediction selesai: {batchjob.outputinfo}")

# Inisialisasi experiment
aiplatform.init(experiment="my-experiment")

Mulai run
aiplatform.startrun("run-1")

Log parameters
aiplatform.logparams({

    "learningrate": 0.01,
    "epochs": 100,
    "batchsize": 32

})

Log metrics
aiplatform.logmetrics({
    "accuracy": 0.95,
    "f1score": 0.93

})

Akhiri run
aiplatform.endrun()

2. Compare Runs

# Dapatkan experiment
experiment = aiplatform.Experiment("my-experiment")

Dapatkan semua runs
runsdf = experiment.getdataframe()

print(runsdf[["runname", "accuracy", "learningrate"]])

Best Practices

1. Resource Management

# Gunakan context manager untuk cleanup
with aiplatform.init(project="project", location="us-central1"):
    # Kode ML Anda di sini
    pass

Hapus resources setelah selesai
endpoint.undeployall()

endpoint.delete()
model.delete()

# Gunakan preemptible VMs untuk training
job = aiplatform.CustomTrainingJob(
    displayname="cost-optimized-training",
    scriptpath="train.py",

    containeruri="training-container"
)

model = job.run(
    machinetype="n1-standard-4",

    replicacount=1,
    bootdisktype="pd-ssd",
    bootdisksize_gb=100
)

Kesimpulan

Vertex AI menyediakan:

Key takeaways:

• Gunakan AutoML untuk prototyping cepat
• Gunakan custom training untuk kontrol
• Register models untuk versioning
• Deploy ke managed endpoints
• Track experiments secara sistematis