| """Collect the per-direction transport runs into one table.""" |
| import glob, os |
| import numpy as np, pandas as pd |
| ROOT = "/root/crossarch-diagnostics" |
| fs = sorted(glob.glob(f"{ROOT}/results/transport_*__*.csv")) |
| if not fs: |
| print("no transport runs yet"); raise SystemExit |
| df = pd.concat([pd.read_csv(f) for f in fs], ignore_index=True) |
| df.to_csv(f"{ROOT}/results/transport_merge.csv", index=False) |
| rows = [] |
| for (t, d), g in df.groupby(["target", "donor"]): |
| g = g.sort_values("alpha").set_index("alpha") |
| b0, d0 = g.nll.get(0.0, np.nan), g.nll_donor_lang.get(0.0, np.nan) |
| nz = g.drop(index=0.0, errors="ignore") |
| best_t = nz.nll.idxmin() if len(nz) else np.nan |
| best_d = nz.nll_donor_lang.idxmin() if len(nz) else np.nan |
| r = g.iloc[0] |
| rows.append(dict( |
| target=t, donor=d, target_lang=r.target_lang, donor_lang=r.donor_lang, |
| hid_target=int(r.hid_target), hid_donor=int(r.hid_donor), |
| L_target=int(r.L_target), L_donor=int(r.L_donor), |
| rectangular_hidden=bool(r.hid_target != r.hid_donor), |
| depth_mismatch=bool(r.L_target != r.L_donor), |
| n_fused_tensors=int(r.n_fused_tensors), |
| nll_target_lang_a0=b0, nll_target_lang_best=nz.nll.min() if len(nz) else np.nan, |
| best_alpha_target=best_t, |
| delta_target_lang=(nz.nll.min() - b0) if len(nz) else np.nan, |
| nll_donor_lang_a0=d0, nll_donor_lang_best=nz.nll_donor_lang.min() if len(nz) else np.nan, |
| best_alpha_donor=best_d, |
| delta_donor_lang=(nz.nll_donor_lang.min() - d0) if len(nz) else np.nan, |
| |
| |
| alpha_max=float(g.index.max()), |
| nll_target_ot_amax=float(g.nll.iloc[-1]), |
| nll_target_rand_amax=float(g.nll_randplan.iloc[-1]) if "nll_randplan" in g else np.nan, |
| ot_minus_rand_target=(float(g.nll.iloc[-1]) - float(g.nll_randplan.iloc[-1])) |
| if "nll_randplan" in g else np.nan, |
| ot_minus_rand_donor=(float(g.nll_donor_lang.iloc[-1]) - |
| float(g.nll_donor_lang_randplan.iloc[-1])) |
| if "nll_donor_lang_randplan" in g else np.nan, |
| seconds=float(r.seconds))) |
| s = pd.DataFrame(rows) |
| s.to_csv(f"{ROOT}/results/table_transport_summary.csv", index=False) |
| print(s.to_string(index=False)) |
|
|
