anhtld commited on
Commit
9bf30b7
·
verified ·
1 Parent(s): 01214f8

Auto-sync: 2026-06-27 10:00:40 (part 3)

Browse files
scripts/slurm/eval_maniskill_policy_rollout.sbatch CHANGED
@@ -19,9 +19,10 @@ SEED="${SLURM_ARRAY_TASK_ID:-0}"
19
  RUN_ROOT="${RUN_ROOT:-}"
20
  OBJECTIVE="${OBJECTIVE:-lattice_field}"
21
  CHECKPOINT_NAME="${CHECKPOINT_NAME:-best.pt}"
 
22
  if [[ -n "$RUN_ROOT" ]]; then
23
  CHECKPOINT="${CHECKPOINT:-$RUN_ROOT/$OBJECTIVE/seed_$SEED/$CHECKPOINT_NAME}"
24
- OUT="${OUT:-$RUN_ROOT/$OBJECTIVE/seed_$SEED/policy_rollout.json}"
25
  else
26
  CHECKPOINT="${CHECKPOINT:?Set CHECKPOINT, or RUN_ROOT for seed-indexed array runs}"
27
  OUT="${OUT:?Set OUT, or RUN_ROOT for seed-indexed array runs}"
 
19
  RUN_ROOT="${RUN_ROOT:-}"
20
  OBJECTIVE="${OBJECTIVE:-lattice_field}"
21
  CHECKPOINT_NAME="${CHECKPOINT_NAME:-best.pt}"
22
+ OUT_NAME="${OUT_NAME:-policy_rollout.json}"
23
  if [[ -n "$RUN_ROOT" ]]; then
24
  CHECKPOINT="${CHECKPOINT:-$RUN_ROOT/$OBJECTIVE/seed_$SEED/$CHECKPOINT_NAME}"
25
+ OUT="${OUT:-$RUN_ROOT/$OBJECTIVE/seed_$SEED/$OUT_NAME}"
26
  else
27
  CHECKPOINT="${CHECKPOINT:?Set CHECKPOINT, or RUN_ROOT for seed-indexed array runs}"
28
  OUT="${OUT:?Set OUT, or RUN_ROOT for seed-indexed array runs}"
scripts/slurm/summarize_h16_policy_ckpt.sbatch ADDED
@@ -0,0 +1,120 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+ #SBATCH --job-name=sum_h16_policy
3
+ #SBATCH --account=def-yalda
4
+ #SBATCH --time=00:20:00
5
+ #SBATCH --cpus-per-task=1
6
+ #SBATCH --mem=2G
7
+ #SBATCH --output=outputs/hpc/logs/%x_%j.out
8
+ #SBATCH --error=outputs/hpc/logs/%x_%j.err
9
+
10
+ set -euo pipefail
11
+
12
+ PROJECT_DIR="${PROJECT_DIR:-$SLURM_SUBMIT_DIR}"
13
+ RUN_ROOT="${RUN_ROOT:-/scratch/$USER/dovla/experiments/dovla_h16_policy_ckpt_runs}"
14
+ OBJECTIVE="${OBJECTIVE:-base}"
15
+ if [[ -z "${PYTHON:-}" ]]; then
16
+ if [[ -x "$PROJECT_DIR/.venv/bin/python" ]]; then
17
+ PYTHON="$PROJECT_DIR/.venv/bin/python"
18
+ else
19
+ PYTHON="python3"
20
+ fi
21
+ fi
22
+
23
+ cd "$PROJECT_DIR"
24
+ mkdir -p results outputs/hpc/logs
25
+
26
+ "$PYTHON" - <<'PY'
27
+ from __future__ import annotations
28
+
29
+ import json
30
+ import os
31
+ import statistics
32
+ from pathlib import Path
33
+
34
+ run_root = Path(os.environ["RUN_ROOT"])
35
+ objective = os.environ.get("OBJECTIVE", "base")
36
+ base_dir = run_root / objective
37
+ baseline_h4 = 0.2967
38
+ baseline_h16_rank = 0.29739130434782607
39
+
40
+ rows = []
41
+ for result_path in sorted(base_dir.glob("seed_*/policy_rollout.json")):
42
+ data = json.loads(result_path.read_text())
43
+ seed = int(result_path.parent.name.split("_")[-1])
44
+ metrics_path = result_path.parent / "metrics.json"
45
+ best_policy = {}
46
+ if metrics_path.exists():
47
+ best_policy = json.loads(metrics_path.read_text()).get("best_policy", {})
48
+ rows.append(
49
+ {
50
+ "seed": seed,
51
+ "path": str(result_path),
52
+ "num_groups": data.get("num_groups"),
53
+ "policy_rollout_success_rate": data.get("policy_rollout_success_rate", 0.0),
54
+ "policy_rollout_progress": data.get("policy_rollout_progress", 0.0),
55
+ "oracle_success_rate": data.get("oracle_success_rate", 0.0),
56
+ "action_mse_to_best": data.get("action_mse_to_best", 0.0),
57
+ "best_policy_val": best_policy,
58
+ "per_task": data.get("per_task", {}),
59
+ }
60
+ )
61
+
62
+ successes = [row["policy_rollout_success_rate"] for row in rows]
63
+ progresses = [row["policy_rollout_progress"] for row in rows]
64
+ mses = [row["action_mse_to_best"] for row in rows]
65
+ summary = {
66
+ "run_root": str(run_root),
67
+ "objective": objective,
68
+ "num_completed": len(rows),
69
+ "baseline_h4_policy_success": baseline_h4,
70
+ "baseline_h16_rank_checkpoint_success": baseline_h16_rank,
71
+ "mean_success": statistics.mean(successes) if successes else 0.0,
72
+ "std_success": statistics.stdev(successes) if len(successes) > 1 else 0.0,
73
+ "mean_progress": statistics.mean(progresses) if progresses else 0.0,
74
+ "mean_action_mse_to_best": statistics.mean(mses) if mses else 0.0,
75
+ "gain_vs_h4": (statistics.mean(successes) - baseline_h4) if successes else 0.0,
76
+ "gain_vs_h16_rank_checkpoint": (
77
+ statistics.mean(successes) - baseline_h16_rank
78
+ )
79
+ if successes
80
+ else 0.0,
81
+ "rows": rows,
82
+ }
83
+
84
+ json_path = Path("results/h16_policy_ckpt_summary.json")
85
+ json_path.write_text(json.dumps(summary, indent=2))
86
+
87
+ lines = [
88
+ "# h=16 Best-Policy Checkpoint Rollout",
89
+ "",
90
+ f"Run root: `{run_root}`",
91
+ f"Completed seeds: {len(rows)}",
92
+ f"Baseline h=4 policy success: {baseline_h4:.2%}",
93
+ f"Baseline h=16 rank-checkpoint success: {baseline_h16_rank:.2%}",
94
+ "",
95
+ f"Mean success: {summary['mean_success']:.2%} +/- {summary['std_success']:.2%}",
96
+ f"Gain vs h=16 rank checkpoint: {summary['gain_vs_h16_rank_checkpoint']:+.2%}",
97
+ f"Mean progress: {summary['mean_progress']:.2%}",
98
+ f"Mean action MSE to best: {summary['mean_action_mse_to_best']:.3f}",
99
+ "",
100
+ "| seed | success | progress | oracle | action MSE |",
101
+ "|---:|---:|---:|---:|---:|",
102
+ ]
103
+ for row in rows:
104
+ lines.append(
105
+ "| {seed} | {success:.2%} | {progress:.2%} | {oracle:.2%} | {mse:.3f} |".format(
106
+ seed=row["seed"],
107
+ success=row["policy_rollout_success_rate"],
108
+ progress=row["policy_rollout_progress"],
109
+ oracle=row["oracle_success_rate"],
110
+ mse=row["action_mse_to_best"],
111
+ )
112
+ )
113
+
114
+ md_path = Path("results/h16_policy_ckpt_summary.md")
115
+ md_path.write_text("\n".join(lines) + "\n")
116
+
117
+ print(json.dumps({key: value for key, value in summary.items() if key != "rows"}, indent=2))
118
+ print(f"Wrote {json_path}")
119
+ print(f"Wrote {md_path}")
120
+ PY
scripts/slurm/train_dovla_h16_policy_ckpt.sbatch ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+ #SBATCH --job-name=train_h16_policyck
3
+ #SBATCH --account=def-yalda_gpu
4
+ #SBATCH --nodes=1
5
+ #SBATCH --ntasks=1
6
+ #SBATCH --cpus-per-task=4
7
+ #SBATCH --gres=gpu:nvidia_h100_80gb_hbm3_1g.10gb:1
8
+ #SBATCH --mem=28G
9
+ #SBATCH --time=03:00:00
10
+ #SBATCH --array=0-2
11
+ #SBATCH --output=outputs/hpc/logs/%x_%A_%a.out
12
+ #SBATCH --error=outputs/hpc/logs/%x_%A_%a.err
13
+
14
+ set -euo pipefail
15
+
16
+ # Retrain the h=16 DoVLAModel with deployment-aligned checkpointing enabled.
17
+ # The trainer now writes both:
18
+ # - best.pt: best validation field ranking checkpoint
19
+ # - best_policy.pt: lowest validation BC loss checkpoint for online rollout
20
+
21
+ PROJECT_DIR="${PROJECT_DIR:-$SLURM_SUBMIT_DIR}"
22
+ SCRATCH_ROOT="/scratch/$USER/dovla"
23
+ SIF="${SIF:-$SCRATCH_ROOT/containers/pytorch_2.7.1_cuda12.8.sif}"
24
+ CONTAINER_PYTHON="${PYTHON:-$SCRATCH_ROOT/envs/maniskill/bin/python}"
25
+ DATASET="${DATASET:-$SCRATCH_ROOT/experiments/h16_merged_dataset}"
26
+ RUN_ROOT="${RUN_ROOT:-$SCRATCH_ROOT/experiments/dovla_h16_policy_ckpt_runs}"
27
+ OBJECTIVE="${OBJECTIVE:-base}"
28
+ SEED=$SLURM_ARRAY_TASK_ID
29
+ OUT_DIR="$RUN_ROOT/$OBJECTIVE/seed_$SEED"
30
+
31
+ module load StdEnv/2023 apptainer/1.4.5
32
+ cd "$PROJECT_DIR"
33
+ mkdir -p "$OUT_DIR" outputs/hpc/logs
34
+
35
+ export OMP_NUM_THREADS=1
36
+ export OPENBLAS_NUM_THREADS=1
37
+ export MKL_NUM_THREADS=1
38
+ export DOVLA_TORCH_THREADS=1
39
+
40
+ PYTHON_CMD=(
41
+ apptainer exec --nv
42
+ --env "OMP_NUM_THREADS=1,OPENBLAS_NUM_THREADS=1,MKL_NUM_THREADS=1,DOVLA_TORCH_THREADS=1,TRANSFORMERS_OFFLINE=1,HF_HUB_OFFLINE=1,PYTHONDONTWRITEBYTECODE=1"
43
+ -B "$PROJECT_DIR:$PROJECT_DIR"
44
+ -B "/scratch/$USER:/scratch/$USER"
45
+ "$SIF" "$CONTAINER_PYTHON"
46
+ )
47
+
48
+ echo "=================================================="
49
+ echo "Train DoVLAModel h=16 with best_policy checkpoint"
50
+ echo "Seed: $SEED"
51
+ echo "Dataset: $DATASET"
52
+ echo "Output: $OUT_DIR"
53
+ echo "=================================================="
54
+
55
+ "${PYTHON_CMD[@]}" -c "
56
+ from dovla_cil.data.datasets import CILDataset
57
+ import torch
58
+ ds = CILDataset('$DATASET')
59
+ print('groups=', len(ds.group_ids), 'records=', len(ds), 'cuda=', torch.cuda.is_available())
60
+ assert len(ds.group_ids) == 2873
61
+ "
62
+
63
+ "${PYTHON_CMD[@]}" scripts/train_dovla.py \
64
+ --dataset "$DATASET" \
65
+ --out "$OUT_DIR" \
66
+ --epochs 50 \
67
+ --batch-groups 32 \
68
+ --records-per-group 16 \
69
+ --pair-count-per-group 32 \
70
+ --hidden-dim 256 \
71
+ --obs-dim 96 \
72
+ --observation-mode state \
73
+ --lang-dim 64 \
74
+ --action-dim 7 \
75
+ --action-horizon 16 \
76
+ --effect-dim 32 \
77
+ --lr 0.001 \
78
+ --device cuda \
79
+ --seed "$SEED" \
80
+ --objective lattice_field
81
+
82
+ test -f "$OUT_DIR/best.pt"
83
+ test -f "$OUT_DIR/best_policy.pt"
84
+
85
+ echo ""
86
+ echo "Training complete"
87
+ echo "Field checkpoint: $OUT_DIR/best.pt"
88
+ echo "Policy checkpoint: $OUT_DIR/best_policy.pt"
tests/test_maniskill_policy_rollout.py CHANGED
@@ -218,3 +218,26 @@ def test_field_mode_scores_clamped_candidates() -> None:
218
  assert float(actions.max()) <= 0.5
219
  assert float(actions.min()) >= -0.5
220
  assert index.tolist()[0] != 0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
218
  assert float(actions.max()) <= 0.5
219
  assert float(actions.min()) >= -0.5
220
  assert index.tolist()[0] != 0
221
+
222
+
223
+ def test_lattice_mode_selects_best_scored_candidate() -> None:
224
+ import torch
225
+
226
+ mean = torch.zeros(1, 1, 3)
227
+ offset = torch.full_like(mean, 0.4)
228
+ model = _StubModel(torch, mean, best_offset=offset)
229
+ candidates = torch.stack([mean, mean + offset], dim=1)
230
+ actions, index = _select_action_chunk(
231
+ model,
232
+ observations=torch.zeros(1, 3),
233
+ instructions=["a"],
234
+ torch=torch,
235
+ selection_mode="lattice",
236
+ num_candidates=1,
237
+ candidate_sigma=0.0,
238
+ selection_seed=0,
239
+ action_candidates=candidates,
240
+ )
241
+
242
+ assert torch.allclose(actions, mean + offset)
243
+ assert index.tolist() == [1]