PyTorch Model Recovery

This skill provides guidance for tasks involving PyTorch model architecture recovery from state dictionaries, selective layer training, and TorchScript export.

When to Use This Skill

This skill applies when:

• Reconstructing a model architecture from a state dictionary (.pt or .pth file containing weights)
• Training or fine-tuning specific layers while keeping others frozen
• Converting a recovered model to TorchScript format
• Debugging model loading issues or architecture mismatches

Approach Overview

Model recovery tasks require a systematic, incremental approach with verification at each step. The key phases are:

1. Architecture Analysis - Infer model structure from state dictionary keys
2. Architecture Implementation - Build the model class to match the state dict
3. Verification - Confirm weights load correctly before any training