"""Train/eval input mismatch probe (2026-11-02): GT steps always take augment_inputs (augment_probability 1); eval (head.eval) never does. Compare, on one eval window at dense s1, the clean-path head inputs (points, pixel footprint) with the augment-path GT rows (noise / floaters off): relative point offset, footprint ratio, and which pixels differ. python probe_aug_path.py""" import json, sys import torch from omegaconf import OmegaConf sys.argv += ['--out', '/workspace/free3d_multik_hr/overfit/probe_aug_path.json'] W = '/workspace/free3d_multik_hr' cfg = OmegaConf.load(W + '/config.yaml'); cfg.run.out_dir = W + '/overfit/scratch_probe' cfg.gaussian.free3d.noise_fraction = 0.; cfg.gaussian.free3d.floater_fraction = 0.; cfg.gaussian.free3d.multik.cap_mode = 'fixed' g = cfg.gaussian_train; g.target_strides = [1]; g.source_strides = [1]; g.max_gaussians = 1 << 20; g.relational_render.query_budget = 1 << 20 g.dense_crop.enabled = False; g.density_mixture.enabled = False from arrow_train.qc import train_free3d_gt as TF, train_broad_refiner as B from arrow_train import free3d_gaussians as F3 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) evalcfg = OmegaConf.create(OmegaConf.to_container(cfg)); evalcfg.data.order = cfg.free3d_pretrain.eval_order loader = iter(B.make_loader(evalcfg, evaluation=True)) while True: raw = next(loader) if not raw.get('_load_failed') and raw['dataset'] == 'hypersim':break sample = B.apply_tier(cfg, B.tensor_tree(raw, 'cuda'), policies['hypersim']) CAP = {} orig = F3.augment_inputs def wrap(settings, context, stages, embedding, points_r, *args, **kw): out = orig(settings, context, stages, embedding, points_r, *args, **kw) CAP.update(points_r=points_r.detach().float(), pfp=context['pixel_footprint'].detach().float().clone(), xy=args[1].detach(), frames=args[0].detach()) if out is not None:CAP.update(new=out['points'].detach().float(), pfp_new=out['context']['pixel_footprint'][:len(points_r)].detach().float()) return out F3.augment_inputs = wrap import arrow_train.gaussian_train as GT head.train() with torch.no_grad(): torch.manual_seed(0); TF.GT_cache_clear() _, part = runner(dict(sample, _training_step=4000, _training_micro=0, _free3d_gt_only=True, _free3d_no_augment=False)) p, q = CAP['points_r'], CAP['new']; z = q[:, 2].abs().clamp_min(1e-6) rel = (p - q).norm(dim=-1) / z fr = CAP['pfp_new'] / CAP['pfp'].clamp_min(1e-12) xy = CAP['xy'].float() res = dict(n=len(p), rel_offset_median=float(rel.median()), rel_offset_p90=float(rel.quantile(.9)), rel_offset_max=float(rel.max()), frac_rel_gt_1e3=float((rel > 1e-3).float().mean()), footprint_ratio_median=float(fr.median()), footprint_ratio_p10=float(fr.quantile(.1)), footprint_ratio_p90=float(fr.quantile(.9)), xy_frac=[float((xy - xy.floor()).mean()), float(xy.min()), float(xy.max())], offset_vec_median=[float(v) for v in (p - q).median(0).values], psnr_aug=float(part.get('gs_relational_base_psnr', -1)), relerr_logged=float(part.get('gs_free3d_aug_pred_gt_relerr', -1))) print('PROBE', json.dumps(res), flush=True)