"""Per-dataset supervision-mode fractions of the hr224 launch config (2026-11-02): no-grad training-mode GT steps on the first N admitted TRAINING windows (training sampler: strides, density mixture, K by supervision, single-source relational plans). Per window: dataset, static_scene, cross-view mode (gs_xview_mode_*), density mode, stride, K drawn / realised, hi-res factor. Writes /workspace/free3d_multik_hr/probe_modes.json and prints a per-dataset table. python probe_modes.py [N=240]""" import json, sys, collections import torch from omegaconf import OmegaConf def main(): W = '/workspace/free3d_multik_hr'; N = int(sys.argv[1]) if len(sys.argv) > 1 else 240 ONLY = set(sys.argv[2].split(',')) if len(sys.argv) > 2 else None # dataset filter, e.g. c4d_scenes,synthverse TAG = ('_' + sys.argv[3]) if len(sys.argv) > 3 else '' cfg = OmegaConf.load(W + '/config.yaml'); cfg.run.out_dir = W + '/overfit/scratch_probe_modes'; cfg.data.workers = 2 from arrow_train.qc import train_free3d_gt as TF, train_broad_refiner as B model, T, _, _, _ = TF.setup(cfg); head = model.gaussian_head head.load_state_dict(torch.load(W + '/init_from.pt', map_location='cpu', weights_only=False)['head'], strict=False) TF.install_teacher(cfg); policies = json.load(open(cfg.free3d_pretrain.policies)); runner = T.GaussianStep(model, cfg) if ONLY: # filtered copy of the training order lines = open(cfg.data.order).read().splitlines(); keep = [lines[0]] + [l for l in lines[1:] if json.loads(l).get('dataset') in ONLY] hdr = json.loads(keep[0]); hdr['header']['entries'] = len(keep) - 1; keep[0] = json.dumps(hdr) path = W + f'/overfit/order{TAG}.jsonl'; open(path, 'w').write('\n'.join(keep) + '\n'); cfg.data.order = path loader = iter(B.make_loader(cfg, 0)); rows = []; head.train(); step = 0 while len(rows) < N: raw = next(loader) if raw.get('_load_failed') or raw['dataset'] not in policies or (ONLY and raw['dataset'] not in ONLY):continue s = B.apply_tier(cfg, B.tensor_tree(raw, 'cuda'), policies[raw['dataset']]); step += 1 s.update(_training_step=step, _training_micro=0, _free3d_gt_only=True, _free3d_no_augment=False) TF.GT_cache_clear() try: with torch.no_grad():_, part = runner(s) except torch.OutOfMemoryError: torch.cuda.empty_cache(); continue if not part:continue pick = lambda pre: next((k[len(pre):] for k, v in part.items() if k.startswith(pre) and isinstance(v, (int, float)) and v > 0), None) rows.append(dict(dataset=raw['dataset'], static=bool(raw.get('static_scene', False)), xview=pick('gs_xview_mode_'), density=pick('gs_density_mode_'), stride=part.get('gs_query_stride'), cams=part.get('gs_relational_render_cameras_selected'), times=part.get('gs_relational_render_times_selected'), requested=part.get('gs_relational_render_requested'), trusted=part.get('gs_relational_render_trusted_eligible'), reject={k[len('gs_relational_rejection_reasons_'):]: v for k, v in part.items() if k.startswith('gs_relational_rejection_reasons_')}, k_drawn=part.get('gs_free3d_k_drawn'), k=part.get('gs_free3d_k_realised'), r=part.get('gs_hires_factor', 1.))) if len(rows) % 20 == 0:print(len(rows), flush=True) by = collections.defaultdict(list) for r in rows:by[r['dataset']].append(r) table = {} for ds, rs in sorted(by.items()): n = len(rs); c = collections.Counter(r['xview'] for r in rs); d = collections.Counter(r['density'] for r in rs) table[ds] = dict(n=n, static=sum(r['static'] for r in rs) / n, xview={k: round(v / n, 3) for k, v in c.items()}, density={k: round(v / n, 3) for k, v in d.items()}, r_mean=round(sum(r['r'] or 1 for r in rs) / n, 2), k_halved=round(sum((r['k'] or 1) < (r['k_drawn'] or 1) for r in rs) / n, 3)) print('MODES', ds, json.dumps(table[ds]), flush=True) tot = collections.Counter(r['xview'] for r in rows) print('MODES_ALL', json.dumps({k: round(v / len(rows), 3) for k, v in tot.items()}), flush=True) json.dump(dict(table=table, rows=rows), open(W + f'/probe_modes{TAG}.json', 'w'), indent=1); print('PROBE_MODES_DONE', flush=True) import os; os._exit(0) # skip loader-worker teardown (aborted once at shutdown) if __name__ == '__main__': main()