Quick parity/export guide

1. Train locally

   • Run your training as usual. It should produce:
     • models/recurrent_model.pth
     • models/best_params.json

2. Export NumPy weights

   • Requires torch CPU in a Python 3.10–3.12 env.

   • Command: python3 export_npz_from_pth.py --pth models/recurrent_model.pth --out models

   • This generates models/recurrent_model_np.npz and bumps its mtime to be newer than the .pth.

3. Parity check (optional, recommended)

   • Command: python3 compare_backends.py

   • Expect near-zero diffs for base_metabolism_kcal and W_adj_pred.

4. Deploy on Render

   • Upload these three files to /app/models/ in the container:

     • recurrent_model_pth
     • recurrent_model_np.npz
     • best_params.json

   • The API will prefer NumPy if the .npz exists and is as new as (or newer than) the .pth.