Models

Overview

The Dataloop platform provides a comprehensive suite of tools for managing, training, deploying, and monitoring AI models in production workflows. Its model management capabilities are designed to seamlessly integrate with Dataloop's data-centric pipeline, allowing teams to focus on building high-performing machine learning (ML) solutions while reducing operational overhead.

What Can You Do With Models in Dataloop?

• Automate Labeling: Speed up data annotation by letting models make initial predictions that humans can refine.
• Train Models: Use your labeled data to train new models directly in Dataloop or connect external training tools.
• Use Pre-Trained Models: Get started quickly with ready-made models for common tasks like object detection or image classification.
• Deploy Models: Host and run your models on Dataloop to make predictions in real-time, or in bulk.
• Monitor Performance: Keep track of how your models are performing and retrain them when needed using new or updated data.

Use the Data Browser to perform the following actions with models:

• Generate Predictions: Send items to models to create predictions (e.g., pre-annotations or testing), enhancing productivity and efficiency.
• Organize with Subsets: Assign items to subsets like train, test, or validation directly from the dataset browser to streamline model training and evaluation.
• Smart Search: Effortlessly find items linked to specific datasets or subsets using advanced search features for better data management.

Dataloop's Models module allows you to efficiently manage the entire lifecycle of your models:

• Leverage public model architectures or integrate your custom models.
• Train model versions using data items and their annotations.
• Test and fine-tune models with different hyperparameters.
• Evaluate model performance by running inferences on datasets.
• Analyze and compare training and evaluation metrics, including false-positive and false-negative results.
• Deploy models as services within Dataloop applications.
• Create continuous learning pipelines for ongoing model improvement.

To access Models: Select Models from the left-side menu.

Create Your Own Model

Dataloop enables you to bring your own custom model to the platform, integrating it seamlessly with Dataloop’s ecosystem. This process involves setting up a model adapter and publishing it as an application that can be used within your projects. Below are the steps to successfully create, integrate, and use your custom model on Dataloop.

You can create and integrate your model using either the Dataloop UI or SDK, depending on your preference.

Using the Dataloop UI

The UI provides a straightforward way to create an application and connect your custom model through Docker registry integration.

1. Integrate Your Docker Registry: Connect your own AWS ECR or GCP GCR/GAR docker registry to the application. This step ensures your Docker images are accessible from the platform.
2. Create a Model Application and Link Your Docker Registry: Navigate to the Dataloop platform and create a new application.
   1. Link Your Docker Image: Provide the URL of your Docker image and link it to the application while creating the application. This enables Dataloop to use your custom model for tasks such as predictions, training, and workflows. To update the docker image URL: While you create an application, in the App Config section, update the Docker Image field with your docker image URL.

Using the Dataloop SDK

This process involves setting up a model adapter and publishing it as an application that can be used within your projects.

Refer to the Create Your Own Model developer documentation for more information.

1. Create a Model Adapter

• A model adapter acts as the bridge between your custom model and Dataloop's platform.
• It ensures that your model's API endpoints and functionality are compatible with Dataloop's framework.
• The adapter handles tasks like data input/output conversion, prediction processing, and integration with Dataloop pipelines.
  Learn more.

2. Publish the Application in the Dataloop Marketplace

• Once your model adapter is ready, package it as an App entity in Dataloop.
• Publish the app in the Dataloop Marketplace, making it available for installation.
  Learn more.

3. Install the Application into Your Project

• After publishing, go to the Marketplace and install the app into your desired project.
• The app installation links your custom model to the project, allowing it to be used in pipelines, tasks, or manual workflows.
  Learn more.

4. Start Using Your Model

Once installed:

• Available for use within the Dataloop platform.
• Use it for tasks such as running predictions, active learning workflows, or custom pipelines.
• Allows annotators and developers to leverage its capabilities.

Model Creation Flow - Automated Status Assignment

Automated Status Assignment is a mechanism that intelligently applies a status label to a model depending on its current state, such as Created, Pre-trained, Trained, or Deployed. This status reflects the model's readiness for use or further development, based on the presence of certain artifacts or milestones achieved in the model's lifecycle.

Model Creation

When creating a new model, the platform now automatically assigns a status based on the presence of artifacts.

• If artifacts exist, indicating that the model has undergone some form of pre-training, it will be assigned the status Pre-trained.
• If no artifacts are present, the model will be given the status Created.

Model Cloning

• Created: Cloning a model with the status Created will result in a new model also with the status Created. This indicates that the model is in its initial stage without any training.
• Trained: When a model marked as Trained is cloned, the new model will inherit the status Pre-Trained. This reflects that the model has undergone training and possesses learned weights that make it ready for further fine-tuning or deployment.
• Deployed: Cloning a model with the status Deployed similarly results in a new model with the status Pre-Trained. This acknowledges that the model not only has been trained but also successfully deployed, indicating a level of robustness and reliability.
• Pre-Trained: Cloning a model already marked as Pre-trained maintains the status Pre-trained for the new model. This consistency underscores the model's readiness and the transferability of its pre-trained state.