Intermediate

Artifacts & Model Registry

Version datasets and models, track data lineage from raw data to deployed model, and manage model lifecycle with the W&B Model Registry.

What are Artifacts?

W&B Artifacts are versioned, immutable snapshots of files and directories. They enable:

  • Dataset versioning: Track every version of your training data.
  • Model versioning: Save model checkpoints with full lineage.
  • Lineage tracking: Visualize which data produced which model.
  • Deduplication: Files are content-addressed; identical files are stored once.

Logging Artifacts

Python - Creating and logging artifacts
import wandb

run = wandb.init(project="artifacts-demo")

# Create a dataset artifact
dataset = wandb.Artifact(
    name="cifar10-processed",
    type="dataset",
    description="CIFAR-10 with augmentation applied",
    metadata={"num_classes": 10, "split": "train"}
)

# Add files or directories
dataset.add_file("data/train.npz")
dataset.add_dir("data/augmented/")

# Log the artifact (creates version v0, v1, v2, ...)
run.log_artifact(dataset)

# Create a model artifact
model_artifact = wandb.Artifact(
    name="resnet50-classifier",
    type="model",
    description="ResNet-50 trained on CIFAR-10"
)
model_artifact.add_file("checkpoints/best_model.pt")
model_artifact.add_file("checkpoints/config.json")
run.log_artifact(model_artifact)

Using Artifacts

Python - Downloading and using artifacts
run = wandb.init(project="artifacts-demo")

# Download the latest version
artifact = run.use_artifact("cifar10-processed:latest")
data_dir = artifact.download()

# Or use a specific version
artifact_v2 = run.use_artifact("cifar10-processed:v2")
data_dir = artifact_v2.download()

# Or use an alias
artifact_prod = run.use_artifact("resnet50-classifier:production")
model_dir = artifact_prod.download()

# Load the model
import torch
model = torch.load(f"{model_dir}/best_model.pt")

Model Registry

The W&B Model Registry provides a centralized place to manage model versions, aliases, and stage transitions.

Python - Model Registry workflow
import wandb

run = wandb.init(project="model-registry-demo")

# Log a model artifact
model_artifact = wandb.Artifact("fraud-detector", type="model")
model_artifact.add_file("model.pt")
run.log_artifact(model_artifact)

# Link to the Model Registry
run.link_artifact(
    model_artifact,
    "model-registry/fraud-detector",
    aliases=["latest", "staging"]
)

# Later, promote to production (via UI or API)
api = wandb.Api()
artifact = api.artifact("entity/project/fraud-detector:staging")
artifact.aliases.append("production")
artifact.save()

Lineage Graph

W&B automatically tracks the lineage of artifacts:

Lineage - Data flow visualization
Raw Data (v0)
  └── Preprocessing Run
        └── Processed Dataset (v0)
              └── Training Run (config: lr=0.001, epochs=50)
                    └── Model Checkpoint (v0)
                          └── Evaluation Run
                                └── Metrics Report (v0)

# Each arrow is tracked automatically
# Click any artifact in the UI to see its full history
✅
Use aliases for deployment: Instead of referencing artifacts by version number (v0, v1), use aliases like "production" and "staging." This way, your deployment code always pulls the correct model without code changes. Update aliases when promoting models.

Ready to Go Deeper?

Live instructor-led courses from our partners. Affiliate disclosure.

🧠

AI & ML Courses - 30% Off

Live instructor-led AI, machine learning, data science, and cloud courses for working professionals. Use code Limited30 at checkout.

Edureka
📊

DataCamp - AI & Data Science

Hands-on Python, machine learning, and AI courses with interactive exercises and real projects.

DataCamp
🎓

edX - Top AI Courses

University-level AI courses from MIT, Harvard, Stanford. Earn certificates that employers recognize.

edX