Computer Vision Model Selection, Architecture Design & Training Strategy
Design Production-Ready Computer Vision Models Fast
What You Can Do
You'll gain structured workflows to architect computer vision models from scratch, optimize training pipelines for your hardware constraints, and debug common CV issues like vanishing gradients or poor data augmentation. This skill walks you through model selection (CNN, Vision Transformers, YOLO variants), training strategy design, and performance profiling — turning research papers into production-ready systems.
Features
Evaluate CNNs, Vision Transformers, EfficientNets, and specialized architectures (YOLO, Mask R-CNN) against your accuracy-latency tradeoffs
Profile data loaders, batch sizes, learning rates, and augmentation strategies to maximize GPU utilization and convergence speed
Diagnose vanishing gradients, underfitting, overfitting, and data quality issues with step-by-step decision trees and remediation strategies
Decision tree for choosing PyTorch vs TensorFlow vs JAX based on team expertise, deployment target, and research needs
Pre-built strategies for image classification, object detection, semantic segmentation, and multi-task CV workflows
Evaluate fine-tuning vs feature extraction vs domain adaptation based on dataset size and task similarity
Measure inference latency, memory footprint, accuracy on edge devices, and identify optimization bottlenecks
Verify model quantization, ONNX export, batch inference setup, and deployment target compatibility
Example Output
Example 1: Model Selection for Real-Time Object Detection
Input: "I need to detect cars in 1080p video at 30 FPS on an NVIDIA Jetson. Dataset is 50K labeled images."
Output:
Recommended: YOLOv8-nano (PyTorch)
- Inference: ~15ms per frame on Jetson (achieves 67 FPS target)
- Accuracy: mAP 0.37 (sufficient for tracking)
- Memory: 6.5MB (fits on-device)
- Alt: TensorFlow Lite MobileNet-SSD if iOS deployment needed
Training Strategy:
1. Start with pretrained YOLOv8-nano on COCO
2. Augmentation: Random hue/saturation shifts, horizontal flips, mosaic (for varied object scales)
3. Batch size: 64 on V100 (adjust if OOM)
4. Learning rate: 0.01 → 0.001 (cosine annealing over 100 epochs)
5. Validation: Split 10% of data, measure mAP every 10 epochs
Example 2: Debugging Vanishing Gradients
Symptom: Loss plateaus after 5 epochs, weights don't update.
Diagnosis Tree:
✓ Check gradient magnitude at each layer → gradients < 1e-7 in early layers?
✓ Add batch norm before activation functions
✓ Reduce learning rate (0.01 → 0.001) or use warm-up schedule
✓ Verify data normalization (images scaled to [-1, 1] or [0, 1]?)
✓ Switch activation: ReLU → Leaky ReLU for deeper networks
✓ Add skip connections (ResNet-style) if depth > 50 layers
Result: Adding batch norm + skip connections → loss resumes descent
What's Included
- SKILL.md: Complete CV model selection and training workflows with decision trees, validation checklists, and debugging playbooks
- Architecture comparison template: Side-by-side specs (latency, accuracy, memory) for 15+ model families
- Training configuration checklist: Data loading, augmentation, optimization, and hardware setup verification
- Debugging decision tree: Common CV failure modes and remediation strategies
- Framework selection guide: PyTorch vs TensorFlow vs JAX comparison matrix
- Data augmentation recipes: 8+ strategies (class-balanced, geometric, photometric, mixup variants)
- Performance profiling script template: Measure inference time and memory on target hardware
Who It's For
- ML Engineers building production computer vision systems from prototype to deployment
- Computer Vision Specialists designing custom architectures for novel applications
- Data Scientists scaling image classification, detection, or segmentation projects
- AI Product Managers evaluating model feasibility and resource requirements for new features
- Research Scientists transitioning academic models into robust, reproducible training pipelines
Best For
- Model architecture selection for new image classification, detection, or segmentation projects
- Training pipeline optimization to maximize accuracy within latency/memory constraints
- Debugging convergence issues when loss plateaus or gradients vanish
- Framework and library decisions when choosing PyTorch, TensorFlow, or other stacks
- Data augmentation strategy design to improve model robustness with limited data
- Edge device deployment planning for Jetson, mobile, or browser-based inference
- Transfer learning evaluations to decide between fine-tuning pretrained models vs training from scratch
You might also like
You can transform raw geotechnical field data into professional-grade site investigation reports that include soil classification using USCS standards, bearing capacity calculations, settlement predictions, and specific foundation design recommendations. The skill organizes stratigraphic information across multiple boring locations, cross-references laboratory test results, and produces defensible conclusions compliant with IBC, ASCE 7, and local building codes—suitable for building officials, structural engineers, and construction teams.
by az9713
You can conduct comprehensive safety impact assessments for any transportation design intervention—from intersection redesigns to speed reduction corridors. The skill guides you through hazard identification, risk quantification using established methodologies (conflict studies, crash prediction models), and evidence-based mitigation development. You'll produce defensible safety documentation suitable for audits, litigation, and project approvals.
by az9713
Quickly evaluate pavement design scenarios using AASHTO 1993 methodology to calculate required structural thickness for flexible (asphalt) and rigid (concrete) pavements. Analyze traffic loading by computing Equivalent Single Axle Loads (ESALs), assess subgrade conditions through CBR interpretation, and compare multiple design alternatives with different material combinations to optimize cost and performance. Generate preliminary design documentation with technical justification for client presentations and design reports.
by az9713
This skill cross-references your residential design specifications against jurisdiction-specific building codes (IBC, IRC, local amendments) to identify code violations, ambiguous details inspectors will challenge, and compliance documentation gaps. You can validate designs across multiple jurisdictions, evaluate design trade-offs (fire-rating vs. acoustic performance), and generate compliance narratives that support permit applications—all before formal submission.
by az9713
You can input room dimensions, functional requirements, and design constraints to receive optimized furniture layouts with documented traffic flow patterns and compliance verification. Claude analyzes competing spatial demands—usable area, sightlines, egress, zoning—and produces data-driven plans that satisfy functional, regulatory, and aesthetic objectives simultaneously. This accelerates initial space planning and helps identify conflicts before costly revisions.
by az9713
This skill deconstructs your writing samples to identify the specific patterns, personality markers, emotional range, language quirks, and formatting habits that make your voice uniquely yours. You'll receive a detailed voice DNA profile that includes your core communication style, signature phrases, emotional tone, formality level, and what you actively avoid—creating a blueprint that AI systems can use to replicate your authentic voice consistently across projects, platforms, and content types.
by az9713
Design earth retention systems by inputting soil properties, site geometry, and regulatory requirements. Claude calculates lateral earth pressures using Coulomb, Rankine, and Mononobe-Okabe methods, verifies factors of safety for internal and external stability, compares competing retention concepts (gravity walls, cantilever, anchored, sheet-pile, soil-nail), and generates cost-benefit analysis and regulatory documentation packages ready for permit submission.
by az9713
Query your local Screenpipe instance to retrieve screen recordings, audio transcriptions, UI element accessibility trees, keyboard/mouse input logs, and productivity analytics. You can search by keywords, filter by content type (audio, OCR, accessibility, input), set time ranges, and extract structured data about your applications, meetings, and work sessions without sending data to external servers.
by az9713