What problem does it solve? Porting a PyTorch timm model to Lux.jl by hand is error-prone: cross-correlation vs convolution, padding semantics, norm epsilon defaults, variance corrections, and NCHW vs WHCN axis order all silently break numerical equivalence. This Skill encodes a seven-phase workflow that produces a parity-verified Lux port with pretrained weight loading from HuggingFace. ## Core Features & Use Cases - Parity fixture capture: Guides creation of HDF5 fixtures from timm via Python sidecars, including per-stage and per-block intermediates for localizing divergence. - Numeric convention checklist: Enforces cross_correlation=true, explicit epsilon/affine on norms, corrected=false variance, zero-padded pooling, and Lux.testmode for parity tests. - Weight loading pipeline: Wires .safetensors downloads from HuggingFace Hub through apply_state_dict mappings, with axis-order normalization between SafeTensors and HDF5 layouts. - Use Case: Port ResNet-50 from timm to Lux.jl: dump fixtures, implement @compact layers, load pretrained weights, and verify end-to-end parity within 1e-3 tolerance before adding more variants. ## Quick Start Ask the agent to port a specific timm model such as resnet50 to Lux.jl with parity tests and HuggingFace weight loading, and it will follow the seven-phase workflow.
