auto-sync 2026-07-02T19:07:26Z workspace (part 7)

workspace/scripts/export_positive_tangent_targets.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+PROJECT_ROOT = Path(__file__).resolve().parents[1]
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+
+from dovla_cil.data.datasets import CILDataset  # noqa: E402
+from dovla_cil.generation.tangent_targets import (  # noqa: E402
+    DEFAULT_BASE_CANDIDATE_TYPES,
+    DEFAULT_LABELS,
+    export_positive_tangent_targets,
+)
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(
+        description=(
+            "Export positive/negative same-state residual tangents for the learned "
+            "CIL-Atlas Generator V2/V3 training path."
+        )
+    )
+    parser.add_argument("--dataset", type=Path, required=True)
+    parser.add_argument("--out", type=Path, required=True)
+    parser.add_argument("--epsilon", type=float, default=0.0)
+    parser.add_argument(
+        "--base-candidate-types",
+        default=",".join(DEFAULT_BASE_CANDIDATE_TYPES),
+        help="Comma-separated candidate_type priority used as the base action.",
+    )
+    parser.add_argument(
+        "--labels",
+        default=",".join(DEFAULT_LABELS),
+        help="Comma-separated tangent labels to export: positive, negative, neutral.",
+    )
+    parser.add_argument("--max-groups", type=int, default=None)
+    parser.add_argument(
+        "--summary-out",
+        type=Path,
+        default=None,
+        help="Optional summary JSON path. Defaults to OUT with _summary suffix.",
+    )
+    parser.add_argument("--no-summary", action="store_true")
+    args = parser.parse_args(argv)
+
+    if args.max_groups is not None and args.max_groups <= 0:
+        parser.error("--max-groups must be positive when provided")
+    if args.epsilon < 0:
+        parser.error("--epsilon must be non-negative")
+
+    dataset = CILDataset(args.dataset)
+    payload = export_positive_tangent_targets(
+        dataset.iter_groups(),
+        epsilon=args.epsilon,
+        base_candidate_types=_parse_csv(args.base_candidate_types),
+        labels=_parse_csv(args.labels),
+        max_groups=args.max_groups,
+    )
+    payload["dataset"] = str(args.dataset)
+
+    args.out.parent.mkdir(parents=True, exist_ok=True)
+    args.out.write_text(json.dumps(payload, indent=2) + "\n")
+    summary = {key: value for key, value in payload.items() if key != "targets"}
+    if not args.no_summary:
+        summary_out = args.summary_out or args.out.with_name(f"{args.out.stem}_summary.json")
+        summary_out.parent.mkdir(parents=True, exist_ok=True)
+        summary_out.write_text(json.dumps(summary, indent=2) + "\n")
+    print(json.dumps(summary, indent=2))
+    print(f"Wrote {args.out}")
+    return 0
+
+
+def _parse_csv(value: str) -> tuple[str, ...]:
+    return tuple(item.strip() for item in value.split(",") if item.strip())
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

workspace/scripts/slurm/advance_v1_generator.sbatch

@@ -0,0 +1,34 @@
+#!/bin/bash
+#SBATCH --job-name=advance_v1
+#SBATCH --account=def-yalda
+#SBATCH --time=00:20:00
+#SBATCH --cpus-per-task=1
+#SBATCH --mem=2G
+#SBATCH --output=outputs/hpc/logs/%x_%j.out
+#SBATCH --error=outputs/hpc/logs/%x_%j.err
+
+set -euo pipefail
+
+PROJECT_DIR="${PROJECT_DIR:-$SLURM_SUBMIT_DIR}"
+RUN_ROOT="${RUN_ROOT:-/scratch/$USER/dovla/experiments/dovla_h16_policy_ckpt_runs}"
+DATASET="${DATASET:-/scratch/$USER/dovla/experiments/six_task_h16_collection}"
+OBJECTIVE="${OBJECTIVE:-transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1}"
+SUBMIT_NEXT="${SUBMIT_NEXT:-0}"
+ADVANCE_ROUND="${ADVANCE_ROUND:-0}"
+PYTHON="${PYTHON:-python3}"
+
+cd "$PROJECT_DIR"
+mkdir -p results outputs/hpc/logs
+
+ARGS=(
+  --project-dir "$PROJECT_DIR"
+  --run-root "$RUN_ROOT"
+  --dataset "$DATASET"
+  --objective "$OBJECTIVE"
+  --round "$ADVANCE_ROUND"
+)
+if [[ "$SUBMIT_NEXT" == "1" ]]; then
+  ARGS+=(--submit-next)
+fi
+
+"$PYTHON" scripts/advance_v1_generator.py "${ARGS[@]}"

workspace/scripts/slurm/export_positive_tangent_targets.sbatch

@@ -0,0 +1,36 @@
+#!/bin/bash
+#SBATCH --job-name=export_pos_tangents
+#SBATCH --account=def-yalda
+#SBATCH --time=00:20:00
+#SBATCH --cpus-per-task=1
+#SBATCH --mem=4G
+#SBATCH --output=outputs/hpc/logs/%x_%j.out
+#SBATCH --error=outputs/hpc/logs/%x_%j.err
+
+set -euo pipefail
+
+PROJECT_DIR="${PROJECT_DIR:-$SLURM_SUBMIT_DIR}"
+SCRATCH_ROOT="/scratch/$USER/dovla"
+PYTHON="${PYTHON:-python3}"
+DATASET="${DATASET:-$SCRATCH_ROOT/experiments/six_task_h16_collection}"
+OUT="${OUT:-$PROJECT_DIR/results/generator_v2_positive_tangent_targets.json}"
+EPSILON="${EPSILON:-0.05}"
+BASE_CANDIDATE_TYPES="${BASE_CANDIDATE_TYPES:-policy,policy_residual,anchor,base,expert}"
+LABELS="${LABELS:-positive,negative,neutral}"
+MAX_GROUPS="${MAX_GROUPS:-}"
+
+cd "$PROJECT_DIR"
+mkdir -p outputs/hpc/logs "$(dirname "$OUT")"
+
+ARGS=(
+  --dataset "$DATASET"
+  --out "$OUT"
+  --epsilon "$EPSILON"
+  --base-candidate-types "$BASE_CANDIDATE_TYPES"
+  --labels "$LABELS"
+)
+if [[ -n "$MAX_GROUPS" ]]; then
+  ARGS+=(--max-groups "$MAX_GROUPS")
+fi
+
+"$PYTHON" scripts/export_positive_tangent_targets.py "${ARGS[@]}"

workspace/tests/test_positive_tangent_targets.py

@@ -0,0 +1,102 @@
+from __future__ import annotations
+
+import math
+
+from dovla_cil.data.schema import (
+    CIL_VERSION,
+    ActionChunk,
+    CILRecord,
+    FailureInfo,
+    RewardInfo,
+    StructuredEffect,
+    make_record_id,
+)
+from dovla_cil.generation.tangent_targets import (
+    choose_base_record,
+    export_positive_tangent_targets,
+    spline_tangent_summary,
+)
+
+
+def _record(
+    group_id: str,
+    action_id: str,
+    candidate_type: str,
+    action: list[list[float]],
+    progress: float,
+    *,
+    success: bool = False,
+    failure: FailureInfo | None = None,
+) -> CILRecord:
+    return CILRecord(
+        version=CIL_VERSION,
+        record_id=make_record_id(group_id, action_id, seed=3),
+        group_id=group_id,
+        state_hash="same-state",
+        task_id="PickCube-v1",
+        scene_id=None,
+        instruction="pick the cube",
+        instruction_family={"family": "pick"},
+        observation_ref=None,
+        observation_inline={"state": [0.0, 1.0]},
+        action_chunk=ActionChunk(action_id=action_id, horizon=len(action), values=action),
+        next_observation_ref=None,
+        next_observation_inline={"state": [progress]},
+        structured_effect=StructuredEffect(),
+        reward=RewardInfo(
+            progress=progress,
+            success=success,
+            terminal_success=success,
+            dense_components={"progress": progress},
+        ),
+        regret=None,
+        rank_within_group=None,
+        candidate_type=candidate_type,
+        failure=failure,
+    )
+
+
+def test_export_positive_tangent_targets_labels_same_state_utility_contrasts() -> None:
+    base = _record("g", "base", "policy", [[0.0, 0.0], [0.0, 0.0]], 0.4)
+    positive = _record("g", "repair", "near_miss", [[0.1, 0.0], [0.2, 0.0]], 0.8)
+    negative = _record(
+        "g",
+        "collide",
+        "random_negative",
+        [[-0.1, 0.0], [-0.2, 0.0]],
+        0.2,
+        failure=FailureInfo(type="collision"),
+    )
+
+    payload = export_positive_tangent_targets([[base, positive, negative]], epsilon=0.05)
+
+    assert payload["num_usable_groups"] == 1
+    assert payload["label_counts"] == {"negative": 1, "positive": 1}
+    by_type = {item["candidate_type"]: item for item in payload["targets"]}
+    assert by_type["near_miss"]["tangent_label"] == "positive"
+    assert by_type["random_negative"]["tangent_label"] == "negative"
+    assert by_type["near_miss"]["delta_action"] == [[0.1, 0.0], [0.2, 0.0]]
+    assert by_type["near_miss"]["tangent_summary"]["endpoint_delta"] == [0.2, 0.0]
+
+
+def test_choose_base_record_uses_priority_order() -> None:
+    expert = _record("g", "expert", "expert", [[1.0]], 1.0)
+    policy = _record("g", "policy", "policy", [[0.0]], 0.3)
+
+    assert choose_base_record([expert, policy]).record_id == policy.record_id
+    assert choose_base_record([expert], base_candidate_types=("expert",)).record_id == expert.record_id
+
+
+def test_spline_tangent_summary_exposes_low_dimensional_code() -> None:
+    summary = spline_tangent_summary(
+        [
+            [0.0, 0.0, 0.0, 0.1],
+            [0.1, 0.0, 0.2, 0.3],
+            [0.2, 0.1, 0.4, 0.7],
+        ]
+    )
+
+    assert summary["spline_code"]["endpoint_delta_position"] == [0.2, 0.1, 0.4]
+    assert summary["spline_code"]["gripper_gate_shift"] == 0.7
+    assert math.isclose(summary["spline_code"]["lift_bias"], 0.4)
+    assert summary["temporal_total_variation"] > 0.0