Select and deploy text classification models

Deploy two text classification models for comparison: DistilBERT Base Cased and BERT Base Uncased. You'll see the differences between these models and deploy them using the optimal instance configuration.

Why these two models

These models show the typical choice customers face in production between performance and cost:

• BERT Base Uncased: Larger, more accurate, but slower and more resource-intensive.

• DistilBERT Base Cased: Smaller, faster, more cost-effective, but potentially less accurate.

This comparison helps you choose the right model for your specific needs.

Understanding model names in the catalog

Text classification model names in the catalog include these components:

• BERT: Bidirectional Encoder Representations from Transformers.

• L-X_H-Y_A-Z: Model structure where:

  • L-X: Number of layers (X).

  • H-Y: Hidden size (Y).

  • A-Z: Number of attention heads (Z).

• Small/Base/Large: Model size and complexity.

• Uncased/Cased - Case sensitivity setting.

Example: Small BERT L-2_H-128_A-2 indicates a small BERT model with:

• 2 layers.

• 128 hidden units.

• 2 attention heads.

Access the JumpStart model catalog

Navigate to the text classification models in JumpStart catalog.

1. Open SageMaker AI Studio

2. In the left navigation pane, choose JumpStart.

3. On the JumpStart page, choose Hugging Face.

4. Choose Text Classification.

You should see a list of available text classification models in the catalog, including DistilBERT and BERT variants.

Deploy DistilBERT Base Cased

Deploy the DistilBERT model using the default configuration.

1. In the model list, find and choose DistilBERT Base Cased (by distilbert).

2. On the model details page, keep the default instance type.

3. Keep all other default settings and choose Deploy.

4. Wait 5-10 minutes for deployment to complete.

5. To verify deployment success, go to Deployments then Endpoints.

6. Confirm the DistilBERT endpoint shows InService status.

Deploy BERT Base Uncased

Deploy the BERT model for comparison with DistilBERT.

1. Return to the Hugging Face text classification models in JumpStart.

2. Find and choose BERT Base Uncased (by google-bert).

3. Keep the default instance type and choose Deploy.

4. To confirm both deployments, check that both endpoints show InService status in the endpoints list.

Both models appear in your endpoints list with InService status.

Important

Copy and save the endpoint names. You'll need them for the evaluation process.

Troubleshooting

If you encounter deployment issues:

• For instance type errors, verify that you're using the default instance type, not CPU instances like ml.m5.large.

• If you can't find models, search using the exact model names, including the publisher in parentheses.

• For failed deployments, check the service health in your Region or try a different Region.

After your model shows InService status, continue to Evaluate and compare model performance to evaluate your deployed model.