paper-replication_skill

Overview

This skill automates reproducible deep learning paper replication and converts academic descriptions into production-ready PyTorch code. It parses PDFs or text, extracts figures, formulas, tables, and architecture details, then produces audited model designs, visual flows, and runnable implementations. The output emphasizes tensor-shape tracking, modular design, and reproducibility checks.

How this skill works

The skill reads a provided paper (PDF or text), performs a three-phase audit: concise problem/contribution summary, mathematical and architectural decomposition, then generates Mermaid diagrams and PyTorch modules. It enforces tensor-shape notation, type hints, weight initialization, seed setting, and includes validation scripts with random-input checks. Deliverables include hyperparameter tables, LaTeX-ready formula extraction, and module-level documentation.

When to use it

• You have a deep learning paper (PDF or text) to reproduce in PyTorch.
• You need the paper's model architecture converted into runnable, industrial-style code.
• You want a precise mapping of formulas, symbols, and loss decomposition to implementation.
• You require architecture diagrams with explicit tensor shapes for integration.
• You need validation scaffolding and reproducibility settings for experiments.

Best practices

• Provide the full PDF and any supplementary material to maximize fidelity.
• Supply reported hyperparameters or tables when available to match the paper closely.
• Specify target input shapes and expected output tasks (classification, detection, etc.).
• Review and confirm ambiguous notations or missing implementation details before training.
• Use the included seed-setting and deterministic flags when comparing metrics to the paper.

Example use cases

• Convert a vision transformer paper into a modular PyTorch implementation with attention blocks and FFNs.
• Reproduce a new convolutional backbone described only by equations and figures, with full shape annotations.
• Generate a training-ready model for a segmentation head plus loss decomposition from a published method.
• Produce module documentation and Mermaid diagrams for integration into an engineering codebase.
• Validate parameter counts and gradient flow against the numbers reported in a paper.