{ "title": "A80 VGGT-guided Regional Budget Sampler", "updated_at_utc8": "2026-08-01 03:10:09 UTC+8", "wandb_project": "https://wandb.ai/tycr/trellis2-vggt-a80", "notes": [ "当前大规模阶段仍训练 VGGTSpatialAdapter / state-anchor controller，尚未切到真正 A80 Region Budget Head。", "数据集为 Toys4K filtered VGGT manifest：可用 meshes 3229，complete pairs 3224；本轮 train512 + val64。", "25-view correction 正在使用独立数据根 /data/jiachen/trellis2_vggt_toys4k_25views，不覆盖旧 16-view 输出。", "训练启动前 waiter 会验证 train512/val64 feature metadata.view_count == 25；训练中每 case 随机采样 2/4/6/8 views。", "数据齐后会并行启动四个 25-view state-anchor ablation：cap=0.05/0.10/0.20 和 state_loss_weight=0.25。", "gpt-5.3-codex subagent 当前账号组不可用；长时间巡检由本地 shell/Python 守护进程承担，避免用 GPT-5.5 高频监控。", "25-view gate watcher: 四组 train512 ablation 完成 val64 后，自动选择 val64 mean loss 最低的 checkpoint 跑 Benchmark100 2/4/6/8。" ], "experiments": [ { "id": "smoke16_completed", "name": "Smoke16 state-anchor ablations", "goal": "确认训练脚本、W&B、Bark 与状态锚定目标能稳定跑通。", "dataset": "Toys4K smoke16", "cases": 16, "gpu": "4,5,6,7", "status": "completed", "progress": "5/5 runs completed", "started_at_utc8": "2026-07-23 earlier UTC+8", "ended_at_utc8": "2026-07-23 earlier UTC+8", "conclusion": "五组 100-step smoke run 已上传到 W&B；可进入 train512。", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter", "wandb_group": "a80_smoke16", "config": "5 组 100-step smoke ablations；state-anchor / cap sweep / no-rescale 对照；W&B group=a80_smoke16。" }, { "id": "train512_materialize", "name": "Materialize Toys4K train512", "goal": "从 source_manifest 链接 mesh 并抽取 512 个训练对象的多视角 PNG。", "dataset": "Toys4K train split", "cases": 512, "gpu": "CPU/IO", "status": "completed", "progress": "512/512 cases", "started_at_utc8": "2026-07-23 15:10:16 UTC+8", "ended_at_utc8": "2026-07-23 16:35:49 UTC+8", "conclusion": "metadata.csv 和 materialization.json 已生成。", "output": "/data/jiachen/trellis2_vggt_toys4k/train512", "config": "materialize_toys4k_vggt_smoke.py --split train --count 512；CPU/IO。" }, { "id": "val64_materialize", "name": "Materialize Toys4K val64", "goal": "从 source_manifest 准备 64 个验证对象，后续用于 sanity validation。", "dataset": "Toys4K val split", "cases": 64, "gpu": "CPU/IO", "status": "completed", "progress": "64/64 cases", "started_at_utc8": "2026-07-23 15:35:04 UTC+8", "ended_at_utc8": "2026-07-23 16:35:49 UTC+8", "conclusion": "metadata.csv 和 materialization.json 已生成。", "output": "/data/jiachen/trellis2_vggt_toys4k/val64", "config": "materialize_toys4k_vggt_smoke.py --split val --count 64；CPU/IO。" }, { "id": "train512_preprocess", "name": "TRELLIS.2 train512 preprocessing", "goal": "生成 mesh dumps、dual grid、shape latent、SS latent、condition renders 和 VGGT/DINO feature cache。", "dataset": "Toys4K train512", "cases": 512, "gpu": "4,5,6,7 where applicable", "status": "completed", "progress": "mesh 512/512; dual 512/512; shape 512/512; ss 512/512; cond 514/512; features 512/512", "started_at_utc8": "2026-07-23 15:35:15 UTC+8", "ended_at_utc8": "2026-07-23 17:45:55 UTC+8", "conclusion": "train512 latent、condition renders 和 VGGT/DINO feature cache 已完成；feature cache 512/512。", "output": "/data/jiachen/trellis2_vggt_toys4k/train512", "config": "dump_mesh, asset_stats, dual_grid@1024, shape latent, SS latent@64, render_cond 16 views, VGGT/DINO feature cache；sharded world_size=4。" }, { "id": "val64_preprocess", "name": "TRELLIS.2 val64 preprocessing", "goal": "准备 64-case 验证数据，与 train512 使用同一套特征/latent schema。", "dataset": "Toys4K val64", "cases": 64, "gpu": "4,5,6,7 where applicable", "status": "completed", "progress": "64/64 fully preprocessed", "started_at_utc8": "2026-07-23 17:43:33 UTC+8", "ended_at_utc8": "2026-07-23 17:48:13 UTC+8", "conclusion": "val64 latent、condition renders 和 VGGT/DINO feature cache 已完成；launcher 收尾处曾出现非关键 shell token 错误，但 64 个 feature 文件和 rank summaries 已验证存在。", "output": "/data/jiachen/trellis2_vggt_toys4k/val64", "config": "与 train512 相同 schema；64-case validation cache；feature cache 使用 GPU 4-7。" }, { "id": "train512_state_anchor_cap010", "name": "Train512 state-anchor cap=0.10", "goal": "把 smoke16 最稳妥的 state-anchor controller 放大到 512-case，观察 loss、state ratio、梯度和耗时。", "dataset": "Toys4K train512", "cases": 512, "gpu": "4", "status": "completed", "progress": "1000/1000 steps", "started_at_utc8": "2026-07-23 17:46:48 UTC+8", "ended_at_utc8": "2026-07-23 18:37:51 UTC+8", "conclusion": "训练完成；最终指标见 train.jsonl / W&B。", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_state_anchor_cap010", "wandb_group": "a80_train512", "config": "train_trellis2_vggt_state_anchor.py --steps 1000 --state-rel-cap 0.10 --accumulation 1；W&B group=a80_train512。", "total_steps": 1000, "last_metrics": { "loss": 0.07862302660942078, "flow_loss": 0.07839387655258179, "state_loss": 0.00022915257432032377, "gradient_norm": 1.5586357116699219, "max_final_state_ratio": 0.0022520862985402346, "step_elapsed_sec": 3.8026952808722854, "total_elapsed_sec": 3101.7745816879906, "peak_vram_bytes": 26166165504 }, "wandb_url": "https://wandb.ai/tycr/trellis2-vggt-a80/runs/1tm36pw5" }, { "id": "train512_state_anchor_cap005", "name": "Train512 state-anchor cap=0.05", "goal": "并行对照更保守的状态投影半径，判断大数据下 loss/ratio 是否更稳定。", "dataset": "Toys4K train512", "cases": 512, "gpu": "5", "status": "completed", "progress": "1000/1000 steps", "started_at_utc8": "2026-07-23 17:46:46 UTC+8", "ended_at_utc8": "2026-07-23 18:38:48 UTC+8", "conclusion": "训练完成；最终指标见 train.jsonl / W&B。", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_state_anchor_cap005", "wandb_group": "a80_train512", "config": "train_trellis2_vggt_state_anchor.py --steps 1000 --state-rel-cap 0.05 --accumulation 1；W&B group=a80_train512。", "total_steps": 1000, "last_metrics": { "loss": 0.07916956394910812, "flow_loss": 0.07892682403326035, "state_loss": 0.00024273787857964635, "gradient_norm": 1.171106219291687, "max_final_state_ratio": 0.0023731866385787725, "step_elapsed_sec": 3.866234004031867, "total_elapsed_sec": 3163.465599251911, "peak_vram_bytes": 26166165504 }, "wandb_url": "https://wandb.ai/tycr/trellis2-vggt-a80/runs/t5ks9tfk" }, { "id": "a80_trajectory_cache_v0", "name": "A80.1 offline trajectory cache v0", "goal": "构造 state-balanced trajectory cache，连接 A65 trajectory 与 A79 proxy states，为 A80.2 Frozen Controller Training 做输入。", "dataset": "Benchmark100 replay/proxy evidence: A65/A78/A79", "cases": 10, "gpu": "CPU", "status": "completed", "progress": "cache built; see summary.json", "started_at_utc8": "2026-07-23 16:00:47 UTC+8", "ended_at_utc8": "2026-07-23 16:02:04 UTC+8", "conclusion": "A80.1 v0 cache 已生成并上传 W&B；当前仍需要更多真实 sampler trajectory 补齐每状态 20-case 覆盖。", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/a80_offline_trajectory_cache_v0", "config": "" }, { "id": "a80_frozen_controller_smoke_v0", "name": "A80.2 frozen controller smoke v0", "goal": "冻结 TRELLIS.2/VGGT 假设下，训练最小 state classifier + region budget head，验证 A80.1 schema、loss、W&B 和 checkpoint 链路。", "dataset": "A80.1 offline trajectory cache v0", "cases": 37, "gpu": "CPU", "status": "completed", "progress": "500/500 steps", "started_at_utc8": "2026-07-23 16:08:43 UTC+8", "ended_at_utc8": "2026-07-23 16:12:13 UTC+8", "conclusion": "A80.2 smoke 训练完成；state/budget loss 已上传 W&B。注意该结果只验证 controller schema 和训练链路，数据覆盖仍不足。", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/a80_frozen_controller_smoke_v0", "total_steps": 500, "last_metrics": { "loss": 0.00822984054684639, "gradient_norm": 0.21749278903007507, "step_elapsed_sec": 0.0031720218248665333, "total_elapsed_sec": 48.78971529798582 }, "config": "" }, { "id": "train512_state_anchor_cap020", "name": "Train512 state-anchor cap=0.20", "goal": "并行对照更宽松的状态投影半径，检查大数据下是否带来更低 flow/state loss 或更高 state ratio 风险。", "dataset": "Toys4K train512", "cases": 512, "config": "train_trellis2_vggt_state_anchor.py --steps 1000 --state-rel-cap 0.20 --accumulation 1；W&B group=a80_train512。", "gpu": "6", "status": "completed", "progress": "1000/1000 steps", "started_at_utc8": "2026-07-23 17:52:42 UTC+8", "ended_at_utc8": "2026-07-23 18:41:47 UTC+8", "conclusion": "训练完成；最终指标见 train.jsonl / W&B。", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_state_anchor_cap020", "total_steps": 1000, "last_metrics": { "loss": 0.07908892631530762, "flow_loss": 0.07885101437568665, "state_loss": 0.00023791106650605798, "gradient_norm": 0.7623661160469055, "max_final_state_ratio": 0.0026638361159712076, "step_elapsed_sec": 3.8509731930680573, "total_elapsed_sec": 3135.0979456673376, "peak_vram_bytes": 26166165504 }, "wandb_url": "https://wandb.ai/tycr/trellis2-vggt-a80/runs/39895drp" }, { "id": "train512_state_anchor_statew025", "name": "Train512 state-anchor state_loss_weight=0.25", "goal": "并行对照较弱状态锚定约束，判断 flow fitting 和 state consistency 的权衡。", "dataset": "Toys4K train512", "cases": 512, "config": "train_trellis2_vggt_state_anchor.py --steps 1000 --state-rel-cap 0.10 --state-loss-weight 0.25 --accumulation 1；W&B group=a80_train512。", "gpu": "7", "status": "completed", "progress": "1000/1000 steps", "started_at_utc8": "2026-07-23 17:50:26 UTC+8", "ended_at_utc8": "2026-07-23 18:42:02 UTC+8", "conclusion": "训练完成；最终指标见 train.jsonl / W&B。", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_state_anchor_statew025", "total_steps": 1000, "last_metrics": { "loss": 0.07861974835395813, "flow_loss": 0.07856206595897675, "state_loss": 0.00023073874763213098, "gradient_norm": 1.770887017250061, "max_final_state_ratio": 0.002095791744068265, "step_elapsed_sec": 3.8629752043634653, "total_elapsed_sec": 3151.0309861521237, "peak_vram_bytes": 26166165504 }, "wandb_url": "https://wandb.ai/tycr/trellis2-vggt-a80/runs/hbfjbbkw" }, { "id": "statew025_benchmark100_protocol2", "name": "Statew025 Benchmark100 protocol 2", "goal": "评估 train512 statew025 checkpoint 在 Benchmark100 上是否相对 TRELLIS.2 baseline 具备 Pareto 优势。", "dataset": "Benchmark100: 50 GSO + 50 Toys4K", "cases": 100, "config": "run_proobjaverse_trellis2.py --mode vggt_state_anchor --adapter-state-rel-cap 0.10 --steps 12；protocol 2 输入视角；4 shards x 25 cases；samples=20000。", "gpu": "4,5,6,7 sharded lanes", "status": "completed", "progress": "generation 100/100; eval 100/100; completed shards 4/4", "started_at_utc8": "2026-07-30 12:20:00 UTC+8", "ended_at_utc8": "2026-07-30 14:28:07 UTC+8", "conclusion": "P2 未通过 Pareto gate: chamfer_mean rel_delta=+0.2944% better=False; chamfer_p95 rel_delta=-0.0479% better=True; hausdorff_approx rel_delta=+0.1617% better=False; fscore_1pct rel_delta=-1.3908% better=False; fscore_5pct rel_delta=-0.1034% better=False; fscore_10pct rel_delta=-0.0332% better=False Aggregate W&B: https://wandb.ai/tycr/trellis2-vggt-a80/runs/qxgcct68", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_state_anchor_statew025/benchmark100_all_protocols", "wandb_group": "a80_statew025_benchmark100", "last_metrics": { "chamfer_mean_rel_delta": "+0.2944%", "chamfer_p95_rel_delta": "-0.0479%", "hausdorff_approx_rel_delta": "+0.1617%", "fscore_1pct_rel_delta": "-1.3908%", "fscore_5pct_rel_delta": "-0.1034%", "fscore_10pct_rel_delta": "-0.0332%" } }, { "id": "statew025_benchmark100_protocol4", "name": "Statew025 Benchmark100 protocol 4", "goal": "评估 train512 statew025 checkpoint 在 Benchmark100 上是否相对 TRELLIS.2 baseline 具备 Pareto 优势。", "dataset": "Benchmark100: 50 GSO + 50 Toys4K", "cases": 100, "config": "run_proobjaverse_trellis2.py --mode vggt_state_anchor --adapter-state-rel-cap 0.10 --steps 12；protocol 4 输入视角；4 shards x 25 cases；samples=20000。", "gpu": "4,5,6,7 sharded lanes", "status": "completed", "progress": "generation 100/100; eval 100/100; completed shards 4/4", "started_at_utc8": "2026-07-30 12:20:00 UTC+8", "ended_at_utc8": "2026-07-30 14:28:07 UTC+8", "conclusion": "P4 未通过 Pareto gate: chamfer_mean rel_delta=+0.3269% better=False; chamfer_p95 rel_delta=+0.2824% better=False; hausdorff_approx rel_delta=-0.1364% better=True; fscore_1pct rel_delta=+0.4688% better=True; fscore_5pct rel_delta=-0.0334% better=False; fscore_10pct rel_delta=-0.0798% better=False Aggregate W&B: https://wandb.ai/tycr/trellis2-vggt-a80/runs/8wfjewdw", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_state_anchor_statew025/benchmark100_all_protocols", "wandb_group": "a80_statew025_benchmark100", "last_metrics": { "chamfer_mean_rel_delta": "+0.3269%", "chamfer_p95_rel_delta": "+0.2824%", "hausdorff_approx_rel_delta": "-0.1364%", "fscore_1pct_rel_delta": "+0.4688%", "fscore_5pct_rel_delta": "-0.0334%", "fscore_10pct_rel_delta": "-0.0798%" } }, { "id": "statew025_benchmark100_protocol6", "name": "Statew025 Benchmark100 protocol 6", "goal": "评估 train512 statew025 checkpoint 在 Benchmark100 上是否相对 TRELLIS.2 baseline 具备 Pareto 优势。", "dataset": "Benchmark100: 50 GSO + 50 Toys4K", "cases": 100, "config": "run_proobjaverse_trellis2.py --mode vggt_state_anchor --adapter-state-rel-cap 0.10 --steps 12；protocol 6 输入视角；4 shards x 25 cases；samples=20000。", "gpu": "4,5,6,7 sharded lanes", "status": "completed", "progress": "generation 100/100; eval 100/100; completed shards 4/4", "started_at_utc8": "2026-07-30 12:20:00 UTC+8", "ended_at_utc8": "2026-07-30 14:28:07 UTC+8", "conclusion": "P6 未通过 Pareto gate: chamfer_mean rel_delta=+0.1775% better=False; chamfer_p95 rel_delta=+0.1094% better=False; hausdorff_approx rel_delta=+0.3160% better=False; fscore_1pct rel_delta=+0.3668% better=True; fscore_5pct rel_delta=-0.3236% better=False; fscore_10pct rel_delta=-0.1030% better=False Aggregate W&B: https://wandb.ai/tycr/trellis2-vggt-a80/runs/qh2211zp", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_state_anchor_statew025/benchmark100_all_protocols", "wandb_group": "a80_statew025_benchmark100", "last_metrics": { "chamfer_mean_rel_delta": "+0.1775%", "chamfer_p95_rel_delta": "+0.1094%", "hausdorff_approx_rel_delta": "+0.3160%", "fscore_1pct_rel_delta": "+0.3668%", "fscore_5pct_rel_delta": "-0.3236%", "fscore_10pct_rel_delta": "-0.1030%" } }, { "id": "statew025_benchmark100_protocol8", "name": "Statew025 Benchmark100 protocol 8", "goal": "评估 train512 statew025 checkpoint 在 Benchmark100 上是否相对 TRELLIS.2 baseline 具备 Pareto 优势。", "dataset": "Benchmark100: 50 GSO + 50 Toys4K", "cases": 100, "config": "run_proobjaverse_trellis2.py --mode vggt_state_anchor --adapter-state-rel-cap 0.10 --steps 12；protocol 8 输入视角；4 shards x 25 cases；samples=20000。", "gpu": "4,5,6,7 sharded lanes", "status": "completed", "progress": "generation 100/100; eval 100/100; completed shards 4/4", "started_at_utc8": "2026-07-30 12:20:00 UTC+8", "ended_at_utc8": "2026-07-30 14:28:07 UTC+8", "conclusion": "P8 未通过 Pareto gate: chamfer_mean rel_delta=+0.3551% better=False; chamfer_p95 rel_delta=+0.1808% better=False; hausdorff_approx rel_delta=+0.1124% better=False; fscore_1pct rel_delta=-0.5274% better=False; fscore_5pct rel_delta=-0.2963% better=False; fscore_10pct rel_delta=-0.1707% better=False Aggregate W&B: https://wandb.ai/tycr/trellis2-vggt-a80/runs/8inhg2o0", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_state_anchor_statew025/benchmark100_all_protocols", "wandb_group": "a80_statew025_benchmark100", "last_metrics": { "chamfer_mean_rel_delta": "+0.3551%", "chamfer_p95_rel_delta": "+0.1808%", "hausdorff_approx_rel_delta": "+0.1124%", "fscore_1pct_rel_delta": "-0.5274%", "fscore_5pct_rel_delta": "-0.2963%", "fscore_10pct_rel_delta": "-0.1707%" } }, { "id": "statew025_val64_eval", "name": "Statew025 Toys4K val64 loss evaluation", "goal": "在 held-out Toys4K val64 上用固定 seed 评估 statew025 checkpoint 的 state-anchor loss，作为 Stage2 gate 的 validation 证据。", "dataset": "Toys4K val64", "cases": 64, "config": "evaluate_state_anchor_val64.py --adapter-checkpoint adapter_step_1000.pt --state-rel-cap 0.10 --state-loss-weight 0.25 --sampler-steps 12；fixed seed 20260730；logs loss/flow_loss/state_loss/state ratio to W&B.", "gpu": "4", "status": "completed", "progress": "64/64 evaluated; retry1 completed and synced to W&B", "started_at_utc8": "", "ended_at_utc8": "2026-07-30 13:40:43 UTC+8", "conclusion": "val64 completed: loss_mean=0.1385539664, flow_loss_mean=0.1373985661, state_loss_mean=0.0046215977; W&B https://wandb.ai/tycr/trellis2-vggt-a80/runs/cqv7qf8f", "output": "outputs/data_survey/vggt_spatial_adapter/train512_state_anchor_statew025/val64_eval" }, { "id": "train512_preprocess_25view_correction", "name": "25-view correction preprocessing train512", "goal": "Generate 25-view renders_cond and VGGT/DINO feature cache without overwriting the completed 16-view root.", "dataset": "Toys4K train512", "cases": 512, "config": "render_cond.py --num_cond_views 25; cache_toys4k_vggt_features.py --render-root <25view>/renders_cond", "gpu": "4,5,6,7", "status": "running", "progress": "25-view train512 renders=512/512 features=512/512", "started_at_utc8": "2026-07-30 13:53:29 UTC+8", "ended_at_utc8": "", "conclusion": "train512 25-view renders_cond and VGGT/DINO feature cache completed.", "output": "/data/jiachen/trellis2_vggt_toys4k_25views/train512" }, { "id": "train512_25view_state_anchor_statew025", "name": "Train512 25-view state-anchor state_loss_weight=0.25", "goal": "25-view correction train512 state-anchor ablation with random 2/4/6/8 views.", "dataset": "Toys4K train512 25-view correction", "cases": 512, "config": "state_rel_cap=0.10; state_loss_weight=0.25; W&B group=a80_train512_25view_ablation", "gpu": "7", "status": "completed", "progress": "1000/1000 steps", "started_at_utc8": "2026-07-30 17:33:32 UTC+8", "ended_at_utc8": "2026-07-30 18:28:35 UTC+8", "conclusion": "25-view ablation training completed; metrics are in train.jsonl and W&B.", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_statew025", "total_steps": 1000, "last_metrics": { "loss": 0.07938607782125473, "flow_loss": 0.07929825782775879, "state_loss": 0.00035128294257447124, "gradient_norm": 2.1223392486572266, "max_final_state_ratio": 0.005648185033351183, "step_elapsed_sec": 1.5206280797719955, "total_elapsed_sec": 3212.787335728295, "peak_vram_bytes": 24473763840 } }, { "id": "train512_25view_state_anchor_cap010", "name": "Train512 25-view state-anchor cap=0.10", "goal": "25-view correction train512 state-anchor ablation with random 2/4/6/8 views.", "dataset": "Toys4K train512 25-view correction", "cases": 512, "config": "state_rel_cap=0.10; state_loss_weight=1.0; W&B group=a80_train512_25view_ablation", "gpu": "4", "status": "completed", "progress": "1000/1000 steps", "started_at_utc8": "2026-07-30 17:33:32 UTC+8", "ended_at_utc8": "2026-07-30 18:29:34 UTC+8", "conclusion": "25-view ablation training completed; metrics are in train.jsonl and W&B.", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_cap010", "total_steps": 1000, "last_metrics": { "loss": 0.08576298505067825, "flow_loss": 0.07753318548202515, "state_loss": 0.008229796774685383, "gradient_norm": 0.2106255441904068, "max_final_state_ratio": 0.06538503617048264, "step_elapsed_sec": 2.645875214599073, "total_elapsed_sec": 3245.677091038786, "peak_vram_bytes": 20786970624 } }, { "id": "train512_25view_state_anchor_cap005", "name": "Train512 25-view state-anchor cap=0.05", "goal": "25-view correction train512 state-anchor ablation with random 2/4/6/8 views.", "dataset": "Toys4K train512 25-view correction", "cases": 512, "config": "state_rel_cap=0.05; state_loss_weight=1.0; W&B group=a80_train512_25view_ablation", "gpu": "5", "status": "completed", "progress": "1000/1000 steps", "started_at_utc8": "2026-07-30 17:33:32 UTC+8", "ended_at_utc8": "2026-07-30 18:31:35 UTC+8", "conclusion": "25-view ablation training completed; metrics are in train.jsonl and W&B.", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_cap005", "total_steps": 1000, "last_metrics": { "loss": 0.10278216749429703, "flow_loss": 0.10154382139444351, "state_loss": 0.0012383447028696537, "gradient_norm": 0.5270354151725769, "max_final_state_ratio": 0.010822024196386337, "step_elapsed_sec": 1.5793226743116975, "total_elapsed_sec": 3330.156869927421, "peak_vram_bytes": 24473763840 } }, { "id": "train512_25view_state_anchor_cap020", "name": "Train512 25-view state-anchor cap=0.20", "goal": "25-view correction train512 state-anchor ablation with random 2/4/6/8 views.", "dataset": "Toys4K train512 25-view correction", "cases": 512, "config": "state_rel_cap=0.20; state_loss_weight=1.0; W&B group=a80_train512_25view_ablation", "gpu": "6", "status": "completed", "progress": "1000/1000 steps", "started_at_utc8": "2026-07-30 17:33:32 UTC+8", "ended_at_utc8": "2026-07-30 18:29:34 UTC+8", "conclusion": "25-view ablation training completed; metrics are in train.jsonl and W&B.", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_cap020", "total_steps": 1000, "last_metrics": { "loss": 0.08434335142374039, "flow_loss": 0.08401740342378616, "state_loss": 0.0003259474760852754, "gradient_norm": 1.1814559698104858, "max_final_state_ratio": 0.010220397263765335, "step_elapsed_sec": 2.9590868996456265, "total_elapsed_sec": 3238.0785004328936, "peak_vram_bytes": 20786970624 } }, { "id": "val64_preprocess_25view_correction", "name": "25-view correction preprocessing val64", "goal": "Generate 25-view renders_cond and VGGT/DINO feature cache without overwriting the completed 16-view root.", "dataset": "Toys4K val64", "cases": 64, "config": "render_cond.py --num_cond_views 25; cache_toys4k_vggt_features.py --render-root <25view>/renders_cond", "gpu": "4,5,6,7", "status": "running", "progress": "25-view val64 renders=64/64 features=64/64", "started_at_utc8": "2026-07-30 17:26:52 UTC+8", "ended_at_utc8": "", "conclusion": "val64 25-view condition renders 和 VGGT/DINO feature cache 已完成；64 个 feature 文件已验证存在。", "output": "/data/jiachen/trellis2_vggt_toys4k_25views/val64" }, { "id": "val64_25view_sidecar", "name": "Val64 25-view sidecar preprocessing", "goal": "提前生成 val64 的 25-view renders_cond 和 VGGT/DINO feature cache，减少 train512 后的串行等待。", "dataset": "Toys4K val64 25-view correction", "cases": 64, "config": "render_cond.py --num_cond_views 25 --world_size 4 --max_workers 1; feature cache handled by incremental 4-GPU cacher; W&B not applicable for preprocessing.", "gpu": "4,5,6,7", "status": "running", "progress": "render_cond25/cache sidecar on GPUs 4,5,6,7: renders=64/64 features=64/64", "started_at_utc8": "2026-07-30 14:46:52 UTC+8", "ended_at_utc8": "", "conclusion": "val64 25-view renders_cond and VGGT/DINO feature cache completed ahead of train512 gate.", "output": "/data/jiachen/trellis2_vggt_toys4k_25views/val64" }, { "id": "a80_25view_benchmark_gate", "name": "25-view Benchmark100 gate selector", "goal": "等待四组 25-view train512 ablation 的 val64 结果，选择 val64 mean loss 最低的 checkpoint 跑 Benchmark100 2/4/6/8。", "dataset": "Toys4K val64 + Benchmark100", "cases": "4 train runs + 64 val + 100 benchmark cases x 4 protocols", "config": "select min val/loss mean; launch_state_anchor_benchmark_eval.py shards; W&B group=a80_train512_25view_benchmark100", "gpu": "4,5,6,7", "status": "running", "progress": "selected train512_25view_state_anchor_statew025 val_loss_mean=0.13727556377125438; Benchmark100 running", "started_at_utc8": "", "ended_at_utc8": "", "conclusion": "", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/logs/a80_25view_benchmark_gate" }, { "id": "train512_25view_state_anchor_statew025_val64_eval", "name": "Train512 25-view state-anchor state_loss_weight=0.25 val64 eval", "goal": "固定 val64 验证集评估 25-view state-anchor checkpoint 的 loss/flow/state 指标。", "dataset": "Toys4K val64 25-view correction", "cases": 64, "config": "evaluate_state_anchor_val64.py; state_rel_cap=0.10; state_loss_weight=0.25; W&B group=a80_train512_25view_val64_eval", "gpu": "7", "status": "completed", "progress": "64/64 val cases", "started_at_utc8": "2026-07-30 18:28:37 UTC+8", "ended_at_utc8": "2026-07-30 18:33:22 UTC+8", "conclusion": "val64 evaluation completed; metrics are in val.jsonl/summary.json and W&B.", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_statew025/val64_eval", "last_metrics": { "val_loss_mean": 0.13727556377125438, "val_flow_loss_mean": 0.1361540302314097, "val_state_loss_mean": 0.004486137616140695, "val_max_final_state_ratio_mean": 0.024968334189907182, "val_encoder_elapsed_sec_mean": 1.2303094402886927, "val_step_elapsed_sec_mean": 3.706337501105736 } }, { "id": "train512_25view_state_anchor_cap020_val64_eval", "name": "Train512 25-view state-anchor cap=0.20 val64 eval", "goal": "固定 val64 验证集评估 25-view state-anchor checkpoint 的 loss/flow/state 指标。", "dataset": "Toys4K val64 25-view correction", "cases": 64, "config": "evaluate_state_anchor_val64.py; state_rel_cap=0.20; state_loss_weight=1.0; W&B group=a80_train512_25view_val64_eval", "gpu": "6", "status": "completed", "progress": "64/64 val cases", "started_at_utc8": "2026-07-30 18:29:36 UTC+8", "ended_at_utc8": "2026-07-30 18:33:51 UTC+8", "conclusion": "val64 evaluation completed; metrics are in val.jsonl/summary.json and W&B.", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_cap020/val64_eval", "last_metrics": { "val_loss_mean": 0.14762639024411328, "val_flow_loss_mean": 0.14271528809331357, "val_state_loss_mean": 0.004911101400466578, "val_max_final_state_ratio_mean": 0.022988339856965467, "val_encoder_elapsed_sec_mean": 1.0173219459975371, "val_step_elapsed_sec_mean": 2.712270970660029 } }, { "id": "train512_25view_state_anchor_cap010_val64_eval", "name": "Train512 25-view state-anchor cap=0.10 val64 eval", "goal": "固定 val64 验证集评估 25-view state-anchor checkpoint 的 loss/flow/state 指标。", "dataset": "Toys4K val64 25-view correction", "cases": 64, "config": "evaluate_state_anchor_val64.py; state_rel_cap=0.10; state_loss_weight=1.0; W&B group=a80_train512_25view_val64_eval", "gpu": "4", "status": "completed", "progress": "64/64 val cases", "started_at_utc8": "2026-07-30 18:29:37 UTC+8", "ended_at_utc8": "2026-07-30 18:33:56 UTC+8", "conclusion": "val64 evaluation completed; metrics are in val.jsonl/summary.json and W&B.", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_cap010/val64_eval", "last_metrics": { "val_loss_mean": 0.14898873025958892, "val_flow_loss_mean": 0.14219566686369944, "val_state_loss_mean": 0.0067930627958503464, "val_max_final_state_ratio_mean": 0.025707812683322118, "val_encoder_elapsed_sec_mean": 1.004803258299944, "val_step_elapsed_sec_mean": 2.725034730086918 } }, { "id": "train512_25view_state_anchor_cap005_val64_eval", "name": "Train512 25-view state-anchor cap=0.05 val64 eval", "goal": "固定 val64 验证集评估 25-view state-anchor checkpoint 的 loss/flow/state 指标。", "dataset": "Toys4K val64 25-view correction", "cases": 64, "config": "evaluate_state_anchor_val64.py; state_rel_cap=0.05; state_loss_weight=1.0; W&B group=a80_train512_25view_val64_eval", "gpu": "5", "status": "completed", "progress": "64/64 val cases", "started_at_utc8": "2026-07-30 18:31:35 UTC+8", "ended_at_utc8": "2026-07-30 18:35:35 UTC+8", "conclusion": "val64 evaluation completed; metrics are in val.jsonl/summary.json and W&B.", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_cap005/val64_eval", "last_metrics": { "val_loss_mean": 0.14626489111833507, "val_flow_loss_mean": 0.13960539991239784, "val_state_loss_mean": 0.0066594914039797, "val_max_final_state_ratio_mean": 0.020610111991118174, "val_encoder_elapsed_sec_mean": 1.0078789762337692, "val_step_elapsed_sec_mean": 2.746763859744533 } }, { "id": "a80_25view_benchmark100_protocol2", "name": "25-view selected Benchmark100 protocol 2", "goal": "评估 25-view train512 最佳 val64 checkpoint (train512_25view_state_anchor_statew025) 在 Benchmark100 protocol 2 上是否优于 TRELLIS.2 baseline。", "dataset": "Benchmark100: 50 GSO + 50 Toys4K", "cases": 100, "config": "run_proobjaverse_trellis2.py --mode vggt_state_anchor --steps 12；protocol 2 输入视角；4 shards x 25 cases；samples=20000。", "gpu": "4,5,6,7 sharded lanes", "status": "completed", "progress": "generation 100/100; eval 100/100; completed shards 4/4", "started_at_utc8": "", "ended_at_utc8": "2026-07-30 19:54:26 UTC+8", "conclusion": "P2 未通过 Pareto gate: chamfer_mean rel_delta=+0.2159% better=False; chamfer_p95 rel_delta=-0.2598% better=True; hausdorff_approx rel_delta=-0.0877% better=True; fscore_1pct rel_delta=+0.2876% better=True; fscore_5pct rel_delta=-0.4236% better=False; fscore_10pct rel_delta=-0.0352% better=False Aggregate W&B: https://wandb.ai/tycr/trellis2-vggt-a80/runs/m93re16o", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_statew025/benchmark100_all_protocols_25view", "wandb_group": "a80_train512_25view_benchmark100", "last_metrics": { "chamfer_mean_rel_delta": "+0.2159%", "chamfer_p95_rel_delta": "-0.2598%", "hausdorff_approx_rel_delta": "-0.0877%", "fscore_1pct_rel_delta": "+0.2876%", "fscore_5pct_rel_delta": "-0.4236%", "fscore_10pct_rel_delta": "-0.0352%", "mean_shard_fscore_5pct": 0.5744714620322371 } }, { "id": "a80_25view_benchmark100_protocol4", "name": "25-view selected Benchmark100 protocol 4", "goal": "评估 25-view train512 最佳 val64 checkpoint (train512_25view_state_anchor_statew025) 在 Benchmark100 protocol 4 上是否优于 TRELLIS.2 baseline。", "dataset": "Benchmark100: 50 GSO + 50 Toys4K", "cases": 100, "config": "run_proobjaverse_trellis2.py --mode vggt_state_anchor --steps 12；protocol 4 输入视角；4 shards x 25 cases；samples=20000。", "gpu": "4,5,6,7 sharded lanes", "status": "completed", "progress": "generation 100/100; eval 100/100; completed shards 4/4", "started_at_utc8": "", "ended_at_utc8": "2026-07-30 19:54:26 UTC+8", "conclusion": "P4 未通过 Pareto gate: chamfer_mean rel_delta=+0.1286% better=False; chamfer_p95 rel_delta=+0.0929% better=False; hausdorff_approx rel_delta=-0.3498% better=True; fscore_1pct rel_delta=-0.3355% better=False; fscore_5pct rel_delta=-0.0675% better=False; fscore_10pct rel_delta=-0.0742% better=False Aggregate W&B: https://wandb.ai/tycr/trellis2-vggt-a80/runs/1jdiupul", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_statew025/benchmark100_all_protocols_25view", "wandb_group": "a80_train512_25view_benchmark100", "last_metrics": { "chamfer_mean_rel_delta": "+0.1286%", "chamfer_p95_rel_delta": "+0.0929%", "hausdorff_approx_rel_delta": "-0.3498%", "fscore_1pct_rel_delta": "-0.3355%", "fscore_5pct_rel_delta": "-0.0675%", "fscore_10pct_rel_delta": "-0.0742%", "mean_shard_fscore_5pct": 0.5867304261843318 } }, { "id": "a80_25view_benchmark100_protocol6", "name": "25-view selected Benchmark100 protocol 6", "goal": "评估 25-view train512 最佳 val64 checkpoint (train512_25view_state_anchor_statew025) 在 Benchmark100 protocol 6 上是否优于 TRELLIS.2 baseline。", "dataset": "Benchmark100: 50 GSO + 50 Toys4K", "cases": 100, "config": "run_proobjaverse_trellis2.py --mode vggt_state_anchor --steps 12；protocol 6 输入视角；4 shards x 25 cases；samples=20000。", "gpu": "4,5,6,7 sharded lanes", "status": "completed", "progress": "generation 100/100; eval 100/100; completed shards 4/4", "started_at_utc8": "", "ended_at_utc8": "2026-07-30 19:54:26 UTC+8", "conclusion": "P6 未通过 Pareto gate: chamfer_mean rel_delta=-0.1682% better=True; chamfer_p95 rel_delta=+0.1831% better=False; hausdorff_approx rel_delta=+0.3127% better=False; fscore_1pct rel_delta=+1.0191% better=True; fscore_5pct rel_delta=-0.1119% better=False; fscore_10pct rel_delta=+0.0049% better=True Aggregate W&B: https://wandb.ai/tycr/trellis2-vggt-a80/runs/ndirxe2k", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_statew025/benchmark100_all_protocols_25view", "wandb_group": "a80_train512_25view_benchmark100", "last_metrics": { "chamfer_mean_rel_delta": "-0.1682%", "chamfer_p95_rel_delta": "+0.1831%", "hausdorff_approx_rel_delta": "+0.3127%", "fscore_1pct_rel_delta": "+1.0191%", "fscore_5pct_rel_delta": "-0.1119%", "fscore_10pct_rel_delta": "+0.0049%", "mean_shard_fscore_5pct": 0.5871736773659547 } }, { "id": "a80_25view_benchmark100_protocol8", "name": "25-view selected Benchmark100 protocol 8", "goal": "评估 25-view train512 最佳 val64 checkpoint (train512_25view_state_anchor_statew025) 在 Benchmark100 protocol 8 上是否优于 TRELLIS.2 baseline。", "dataset": "Benchmark100: 50 GSO + 50 Toys4K", "cases": 100, "config": "run_proobjaverse_trellis2.py --mode vggt_state_anchor --steps 12；protocol 8 输入视角；4 shards x 25 cases；samples=20000。", "gpu": "4,5,6,7 sharded lanes", "status": "completed", "progress": "generation 100/100; eval 100/100; completed shards 4/4", "started_at_utc8": "", "ended_at_utc8": "2026-07-30 19:54:26 UTC+8", "conclusion": "P8 未通过 Pareto gate: chamfer_mean rel_delta=+0.1344% better=False; chamfer_p95 rel_delta=-0.0296% better=True; hausdorff_approx rel_delta=-0.0598% better=True; fscore_1pct rel_delta=+0.4114% better=True; fscore_5pct rel_delta=-0.0540% better=False; fscore_10pct rel_delta=-0.0872% better=False Aggregate W&B: https://wandb.ai/tycr/trellis2-vggt-a80/runs/yae5kcq9", "output": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_statew025/benchmark100_all_protocols_25view", "wandb_group": "a80_train512_25view_benchmark100", "last_metrics": { "chamfer_mean_rel_delta": "+0.1344%", "chamfer_p95_rel_delta": "-0.0296%", "hausdorff_approx_rel_delta": "-0.0598%", "fscore_1pct_rel_delta": "+0.4114%", "fscore_5pct_rel_delta": "-0.0540%", "fscore_10pct_rel_delta": "-0.0872%", "mean_shard_fscore_5pct": 0.5901289947657635 } } ], "current_focus": "25-view correction preprocessing is running: train512 render/cache plus val64 sidecar. Four train512 25-view ablations will auto-launch on GPUs 4-7 after train512=512 features and val64=64 features are verified as metadata.view_count=25; successful training runs then launch matching val64 W&B evaluations. A Benchmark100 gate watcher is active and will select the lowest val64 mean loss checkpoint for protocols 2/4/6/8.", "phases": [ { "id": "phase1_smoke", "name": "第一阶段 smoke", "dataset": "Toys4K smoke16", "cases": "16 cases", "goal": "验证 A80 trajectory cache、state label、budget head loss、W&B 和 Bark 链路能跑通。", "config": "100-step state-anchor ablations + A80.1/A80.2 schema smoke；不做最终质量声明。", "estimated_time": "1-3 小时", "status": "completed", "gate": "训练脚本、loss、checkpoint、W&B、Bark 均稳定。" }, { "id": "phase2_train512", "name": "第二阶段正式小规模", "dataset": "Toys4K train512 + val64", "cases": "512 train / 64 val", "goal": "在 512-case 上验证 VGGT/state-anchor 控制器是否具备稳定训练信号，再准备 A80 frozen controller 扩展。", "config": "每 case 随机 2/4/6/8 输入视角，来自已有 25 render views；GPU 4-7 做 cache/预处理；训练队列为 GPU4 cap=0.10、GPU5 cap=0.05、GPU6 cap=0.20、GPU7 state_loss_weight=0.25。", "estimated_time": "VGGT/cache/trajectory 6-12 小时；controller 训练 8-24 小时；第一版有效结果约 1 天。", "status": "running", "gate": "val64 和 Benchmark100 上出现 Pareto 优势，或形成明确失败结论。" }, { "id": "phase3_expand", "name": "第三阶段扩展", "dataset": "Toys4K available 3224 cases", "cases": "1k-3k cases", "goal": "在 512-case 版本有效后扩大数据规模，验证趋势是否稳健。", "config": "复用 phase2 通过的 controller/sampler 配置，再扩大到 1k-3k；未通过 phase2 gate 前不启动。", "estimated_time": "2-4 天", "status": "planned", "gate": "仅当 512-case 版本在 val 和 Benchmark100 上显示 Pareto 优势后启动。" } ], "selected_25view_benchmark": "/home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/train512_25view_state_anchor_statew025/benchmark100_all_protocols_25view/selected_checkpoint.json" }