---

name: deep-learning description: "Route arcgis.learn geospatial deep-learning workflows, model selection, deployment, and optional dependency checks." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Deep Learning

Use this sub-skill for arcgis.learn work only.

Start here

• Read learn-workflows for the workflow shapes, training loop, and deployment patterns.
• Read model-catalog to choose a model family by data shape and task.
• Read troubleshooting when imports, GPU readiness, data prep, or deployment fail.
• Run python scripts/check_learn_optional_deps.py to probe arcgis.learn, torch, torchvision, and CUDA support without training or downloads.

Use when

• The task is about imagery deep learning, text/NLP deep learning, tabular/time-series ML, point-cloud models, custom model extensibility, or model publishing/inference in arcgis.learn.
• The user needs help deciding between prepare_data, prepare_tabulardata, prepare_textdata, MLModel, ModelExtension, or a specific model family.
• The user wants a safe import/dependency triage for arcgis.learn on a machine that may not have torchvision or GPU support.

Route away when

• Raster analytics without learned models -> imagery-raster-analysis
• Feature/SEDF analysis -> features-dataframes-analysis
• GIS content/admin/auth -> gis-admin-content
• AI utility services, dashboards, knowledge graphs, or app automation -> apps-knowledge-ai-services

Operating rules

• Treat arcgis.learn as optional. If torchvision is missing, stop at the dependency gate and report that the deep-learning stack is incomplete.
• Do not claim GPU training, model downloads, or notebook execution success unless the user provided the required runtime and data.
• Prefer notebook-independent guidance that survives outside the original checkout.
• Do not call ArcGIS services or publish models unless the user explicitly asks and has the required credentials and service access.
• Use the workflow references for all detailed API, model, and troubleshooting steps; keep this router terse.