How dataloop-ai-apps/active-learning works
Project type and purpose
• This is an open-source “app” (plugin) for the Dataloop platform that provides three custom Active Learning nodes (data split, create-new-model, compare-models).
• It bundles Python modules (node implementations), JSON pipeline-config templates, and Vue.js panels (UI) to integrate into Dataloop’s pipeline builder and runtime.
File structure
• Top-level files:
– Dockerfile, .dataloop.cfg, requirements.txt: build and package the app as a container for Dataloop runtime.
– modules/ .py files: business logic for each node.
– pipeline_configs/: JSON templates for node parameters and filters.
– panels/dataSplit/: static build of the Vue.js panel for the DataSplit node.
– srcs/dataSplit/: source code (Vue, TypeScript, Vite) for that panel.
– docs/, assets/: documentation and images illustrating workflows.
Additional insights from README
• Each node’s inputs, outputs, parameters, and underlying behavior are documented.
– User & Dataloop Web UI (green)
– Custom Panels (Vue.js SPA) for node configuration (blue)
– Pipeline Orchestrator (Dataloop engine) (orange)
– Node Runtimes (Python modules in Docker) (purple)
– Dataloop Backend Services: Model Registry, Dataset Storage, Metadata Service (gray)
– Pipeline Config Templates (JSON)
1. User interacts with Dataloop Web UI → loads Custom Panel UI from panels/dataSplit (and others).
2. Panel UI reads/writes JSON templates from pipeline_configs to populate form fields.
3. User submits node configuration → Dataloop Orchestrator stores config.
4. At runtime, Orchestrator spins up Docker container (built from Dockerfile) hosting Python modules.
5. Node code calls Dataloop SDK (dl.Model, dl.Dataset) to perform:
– data_split: tag items in storage
– create_new_model: clone & fine-tune model
– model_compare: fetch metrics, compute AUC-PR, output decision
6. Results (tags, new model entity, winning_model) flow back to Pipeline Orchestrator → shown in UI.
c. Architectural patterns & principles
– Plugin architecture: UI panels and Python modules are loosely coupled via the Dataloop platform.
– Separation of concerns:
• Frontend (Vue.js panels) only for config.
• Backend (Python modules) only for execution logic.
– Containerization: each node runs in an isolated Docker environment.
– Extensibility: new nodes can be added by dropping modules, panels, configs.
d. Key technologies & frameworks
– Frontend: Vue.js, TypeScript, Vite
– Backend: Python, Dataloop SDK (dl.*)
– Configuration: JSON templates in pipeline_configs
– Packaging: Docker, .dataloop.cfg
– CI/CD & versioning: GitHub repo, bumpversion, CONTRIBUTING guidelines