feat: add fixed-pool scoliosis partition helper

configs/drf/pretreatment_heatmap_drf_118_aligned.yaml

@@ -0,0 +1,25 @@
+coco18tococo17_args:
+  transfer_to_coco17: False
+
+padkeypoints_args:
+  pad_method: knn
+  use_conf: True
+
+norm_args:
+  pose_format: coco
+  use_conf: ${padkeypoints_args.use_conf}
+  heatmap_image_height: 128
+
+heatmap_generator_args:
+  sigma: 8.0
+  use_score: ${padkeypoints_args.use_conf}
+  img_h: ${norm_args.heatmap_image_height}
+  img_w: ${norm_args.heatmap_image_height}
+  with_limb: null
+  with_kp: null
+
+align_args:
+  align: True
+  final_img_size: 64
+  offset: 0
+  heatmap_image_size: ${norm_args.heatmap_image_height}

datasets/Scoliosis1K/README.md

@@ -69,6 +69,33 @@ python -m torch.distributed.launch --nproc_per_node=4 \
 opengait/main.py --cfgs configs/sconet/sconet_scoliosis1k.yaml --phase test --log_to_file
 ```
 
+### Fixed-pool ratio comparison
+
+If you want to compare `1:1:2` against `1:1:8` without changing the evaluation
+pool, do not compare `Scoliosis1K_112.json` against `Scoliosis1K_118.json`
+directly. Those two files differ substantially in train/test membership.
+
+For a cleaner same-pool comparison, use:
+
+* `datasets/Scoliosis1K/Scoliosis1K_118.json`
+  * original `1:1:8` split
+* `datasets/Scoliosis1K/Scoliosis1K_118_fixedpool_train112.json`
+  * same `TEST_SET` as `118`
+  * same positive/neutral `TRAIN_SET` ids as `118`
+  * downsampled `TRAIN_SET` negatives to `148`, giving train counts
+    `74 positive / 74 neutral / 148 negative`
+
+The helper used to generate that derived partition is:
+
+```bash
+uv run python scripts/build_scoliosis_fixedpool_partition.py \
+  --base-partition datasets/Scoliosis1K/Scoliosis1K_118.json \
+  --dataset-root /mnt/public/data/Scoliosis1K/Scoliosis1K-sil-pkl \
+  --negative-multiplier 2 \
+  --output-path datasets/Scoliosis1K/Scoliosis1K_118_fixedpool_train112.json \
+  --seed 118
+```
+
 ### Modality sanity check
 
 The silhouette and skeleton-map pipelines are different experiments and should not be mixed when you interpret results.

docs/drf_author_checkpoint_compat.md

@@ -96,6 +96,79 @@ Result:
 
 This is the strongest recovered path so far.
 
+### Verified provenance of `Scoliosis1K-drf-pkl-118-aligned`
+
+The `118-aligned` root is no longer just an informed guess. It was verified
+directly against the raw pose source:
+- `/mnt/public/data/Scoliosis1K/Scoliosis1K-pose-pkl`
+
+The matching preprocessing path is:
+- `datasets/pretreatment_scoliosis_drf.py`
+- default heatmap config:
+  - `configs/drf/pretreatment_heatmap_drf.yaml`
+- archived equivalent config:
+  - `configs/drf/pretreatment_heatmap_drf_118_aligned.yaml`
+
+That means the aligned root was produced with:
+- shared `sigma: 8.0`
+- `align: True`
+- `final_img_size: 64`
+- default `heatmap_reduction=upstream`
+- no `--stats_partition`, i.e. dataset-level PAV min-max stats
+
+Equivalent command:
+
+```bash
+uv run python datasets/pretreatment_scoliosis_drf.py \
+  --pose_data_path /mnt/public/data/Scoliosis1K/Scoliosis1K-pose-pkl \
+  --output_path /mnt/public/data/Scoliosis1K/Scoliosis1K-drf-pkl-118-aligned
+```
+
+Verification evidence:
+- a regenerated `0_heatmap.pkl` sample from the raw pose input matched the stored
+  `Scoliosis1K-drf-pkl-118-aligned` sample exactly (`array_equal == True`)
+- a full recomputation of `pav_stats.pkl` from the raw pose input matched the
+  stored `pav_min`, `pav_max`, and `stats_partition=None` exactly
+
+So `118-aligned` is the old default OpenGait-style DRF export, not the later:
+- `118-paper` paper-literal summed-heatmap export
+- `118` train-only-stats splitroot export
+- `sigma15` / `sigma15_joint8` exports
+
+### Targeted preprocessing ablations around the recovered path
+
+After verifying the aligned root provenance, a few focused runtime/data ablations
+were tested against the author checkpoint to see which part of the contract still
+mattered most.
+
+Baseline:
+- `118-aligned`
+- `BaseSilCuttingTransform`
+- result:
+  - `80.24 Acc / 76.73 Prec / 76.40 Rec / 76.56 F1`
+
+Hybrid 1:
+- aligned heatmap + splitroot PAV
+- result:
+  - `77.30 Acc / 73.70 Prec / 73.04 Rec / 73.28 F1`
+
+Hybrid 2:
+- splitroot heatmap + aligned PAV
+- result:
+  - `80.37 Acc / 77.16 Prec / 76.48 Rec / 76.80 F1`
+
+Runtime ablation:
+- `118-aligned` + `BaseSilTransform` (`no-cut`)
+- result:
+  - `49.93 Acc / 50.49 Prec / 51.58 Rec / 47.75 F1`
+
+What these ablations suggest:
+- `BaseSilCuttingTransform` is necessary; `no-cut` breaks the checkpoint badly
+- dataset-level PAV stats (`stats_partition=None`) matter more than the exact
+  aligned-vs-splitroot heatmap writer
+- the heatmap export is still part of the contract, but it is no longer the
+  dominant remaining mismatch
+
 ### Other tested paths
 
 `configs/drf/drf_author_eval_118_splitroot_1gpu.yaml`
@@ -123,6 +196,8 @@ What these results mean:
 - the original “very bad” local eval was mostly a compatibility failure
 - the largest single hidden bug was the class-order mismatch
 - the author checkpoint is also sensitive to which local DRF dataset root is used
+- the recovered runtime is now good enough to make the checkpoint believable, but
+  preprocessing alone did not recover the paper DRF headline row
 
 What they do **not** mean:
 
@@ -130,6 +205,20 @@ What they do **not** mean:
 - the provided YAML is trustworthy as-is
 - the paper’s full DRF claim is fully reproduced here
 
+One practical caveat on `1:1:2` vs `1:1:8` comparisons in this repo:
+- local `Scoliosis1K_112.json` and `Scoliosis1K_118.json` are not the same train/test
+  split with only a different class ratio
+- they differ substantially in membership
+- so local `112` vs `118` results should not be overinterpreted as a pure
+  class-balance ablation unless the train/test pool is explicitly held fixed
+
+To support a clean same-pool comparison, the repo now also includes:
+- `datasets/Scoliosis1K/Scoliosis1K_118_fixedpool_train112.json`
+
+That partition keeps the full `118` `TEST_SET` unchanged and keeps the same
+positive/neutral `TRAIN_SET` ids as `118`, but downsamples `TRAIN_SET` negatives
+to `148` so the train ratio becomes `74 / 74 / 148` (`1:1:2`).
+
 The strongest recovered result:
 - `80.24 / 76.73 / 76.40 / 76.56`
 

docs/sconet-drf-status-and-training.md

@@ -1,8 +1,12 @@
 # ScoNet and DRF: Status, Architecture, and Reproduction Notes
 
-This note records the current Scoliosis1K implementation status in this repo and the main conclusions from the recent reproduction/debugging work.
+This note is the high-level status page for Scoliosis1K work in this repo.
+It records what is implemented, what currently works best in practice, and
+how to interpret the local DRF/ScoNet results.
 
-For a stricter paper-vs-local reproducibility breakdown, see [scoliosis_reproducibility_audit.md](scoliosis_reproducibility_audit.md).
+For the stricter paper-vs-local breakdown, see [scoliosis_reproducibility_audit.md](scoliosis_reproducibility_audit.md).
+For the concrete experiment queue, see [scoliosis_next_experiments.md](scoliosis_next_experiments.md).
+For the author-checkpoint compatibility recovery, see [drf_author_checkpoint_compat.md](drf_author_checkpoint_compat.md).
 For the recommended long-running local launch workflow, see [systemd-run-training.md](systemd-run-training.md).
 
 ## Current status
@@ -12,6 +16,22 @@ For the recommended long-running local launch workflow, see [systemd-run-trainin
 - `opengait/modeling/models/drf.py` is now implemented as a standalone DRF model in this repo.
 - Logging supports TensorBoard and optional Weights & Biases through `opengait/utils/msg_manager.py`.
 
+## Current bottom line
+
+- The current practical winner is the skeleton-map ScoNet path, not DRF.
+- The best verified local checkpoint is:
+  - `ScoNet_skeleton_112_sigma15_joint8_bodyonly_plaince_adamw_cosine_finetune_1gpu_80k`
+  - retained best checkpoint at `27000`
+  - verified full-test result: `92.38 Acc / 90.30 Prec / 87.39 Rec / 88.70 F1`
+- The strongest practical recipe behind that checkpoint is:
+  - split: `1:1:2`
+  - representation: `body-only`
+  - losses: plain CE + triplet
+  - baseline training: `SGD`
+  - later finetune: `AdamW` + cosine decay
+- A local DRF run trained from scratch on the same practical recipe did not improve over the plain skeleton baseline.
+- The author-provided DRF checkpoint is now usable in-tree after compatibility fixes, but only under the recovered `118-aligned` runtime contract.
+
 ## Naming clarification
 
 The name `ScoNet` is overloaded across the paper, config files, and checkpoints. Use the mapping below when reading this repo:
@@ -73,20 +93,47 @@ The main findings so far are:
   - a later full-test rerun confirmed the `body-only + plain CE` `7000` result exactly
   - an `AdamW` cosine finetune from that same plain-CE checkpoint improved the practical best further; the retained `27000` checkpoint reproduced at `92.38%` accuracy and `88.70%` macro-F1 on the full test set
   - a `head-lite + plain CE` variant looked promising on the fixed proxy subset but underperformed on the full test set at `7000` (`78.07%` accuracy, `62.08%` macro-F1)
+- The first practical DRF bridge on that same winning `1:1:2` recipe did not improve on the plain skeleton baseline:
+  - best retained DRF checkpoint (`2000`) on the full test set: `80.21 Acc / 58.92 Prec / 59.23 Rec / 57.84 F1`
+  - practical plain skeleton checkpoint (`7000`) on the full test set: `83.16 Acc / 68.24 Prec / 80.02 Rec / 68.47 F1`
+- The author-provided DRF checkpoint initially looked unusable in this fork, but that turned out to be a compatibility problem, not a pure weight problem.
+  - after recovering the legacy runtime contract, the best compatible path was `Scoliosis1K-drf-pkl-118-aligned`
+  - recovered author-checkpoint result: `80.24 Acc / 76.73 Prec / 76.40 Rec / 76.56 F1`
 
 The current working conclusion is:
 
 - the core ScoNet trainer is not the problem
 - the strong silhouette checkpoint is not evidence that the skeleton-map path works
-- the main remaining suspect is the skeleton-map representation and preprocessing path
+- the biggest historical problem was the skeleton-map/runtime contract, not just the optimizer
 - for practical model development, `1:1:2` is currently the better working split than `1:1:8`
 - for practical model development, the current best skeleton recipe is `body-only + plain CE`, and the current best retained checkpoint comes from a later `AdamW` cosine finetune on `1:1:2`
-- the first practical DRF bridge on that same winning `1:1:2` recipe did not improve on the plain skeleton baseline:
+- for practical use, DRF is still behind the local ScoNet skeleton winner
-  - best retained DRF checkpoint (`2000`) on the full test set: `80.21 Acc / 58.92 Prec / 59.23 Rec / 57.84 F1`
+- for paper-compatibility analysis, the author checkpoint demonstrates that our earlier DRF failure was partly caused by contract mismatch
-  - current best plain skeleton checkpoint (`7000`) on the full test set: `83.16 Acc / 68.24 Prec / 80.02 Rec / 68.47 F1`
 
 For readability in this repo's docs, `ScoNet-MT-ske` refers to the skeleton-map variant that the DRF paper writes as `ScoNet-MT^{ske}`.
 
+## DRF compatibility note
+
+There are now two different DRF stories in this repo:
+
+1. The local-from-scratch DRF branch.
+   - This is the branch trained directly in our fork on the current practical recipe.
+   - It did not beat the plain skeleton baseline.
+
+2. The author-checkpoint compatibility branch.
+   - This uses the author-supplied checkpoint plus in-tree compatibility fixes.
+   - The main recovered issues were:
+     - legacy module naming drift: `attention_layer.*` vs `PGA.*`
+     - class-order mismatch between the author stub and our evaluator assumptions
+     - stale/internally inconsistent author YAML
+     - preprocessing/runtime mismatch, where `118-aligned` matched much better than the paper-literal export
+
+That distinction matters. It means:
+
+- "our DRF training branch underperformed" is true
+- "the author DRF checkpoint is unusable" is false
+- "the author result was drop-in reproducible from the handed-over YAML" is also false
+
 ## Architecture mapping
 
 `ScoNet` in this repo maps to the paper as follows:
@@ -115,12 +162,6 @@ The standard Scoliosis1K ScoNet recipe is:
 
 The skeleton-map control used the same recipe, except for the modality-specific changes listed above.
 
-## Recommended next checks
-
-1. Train a pure silhouette `1:1:8` baseline from the upstream ScoNet config as a clean sanity control.
-2. Treat skeleton-map preprocessing as the primary debugging target until a `ScoNet-MT-ske`-style run gets close to the paper.
-3. Only after the skeleton baseline is credible should DRF/PAV-specific conclusions be treated as decisive.
-
 ## Practical conclusion
 
 For practical use in this repo, the current winning path is:
@@ -143,12 +184,15 @@ So the current local recommendation is:
 - keep `1:1:2` as the main practical split
 - treat DRF as an optional research branch, not the mainline model
 
+If the goal is practical deployment/use, use the retained best skeleton checkpoint family first.
+If the goal is paper audit or author-checkpoint verification, use the dedicated DRF compatibility configs instead.
+
 ## Remaining useful experiments
 
 At this point, there are only a few experiments that still look high-value:
 
 1. one clean `full-body` finetune under the same successful `1:1:2` recipe, just to confirm that `body-only` is really the best practical representation
-2. one DRF rerun on top of the now-stronger practical baseline recipe, only if the goal is to test whether DRF can add value once the skeleton branch is already strong
+2. one DRF warm-start rerun on top of the now-stronger practical baseline recipe, only if the goal is to test whether DRF can add value once the skeleton branch is already strong
 3. a final packaging/evaluation pass around the retained best checkpoints, rather than more broad preprocessing churn
 
 Everything else looks lower value than simply using the retained best `27000` checkpoint.

scripts/build_scoliosis_fixedpool_partition.py

@@ -0,0 +1,121 @@
+from __future__ import annotations
+
+import json
+import random
+from collections import Counter
+from pathlib import Path
+from typing import TypedDict, cast
+
+import click
+
+
+class Partition(TypedDict):
+    TRAIN_SET: list[str]
+    TEST_SET: list[str]
+
+
+def infer_pid_label(dataset_root: Path, pid: str) -> str:
+    pid_root = dataset_root / pid
+    if not pid_root.exists():
+        raise FileNotFoundError(f"PID root not found under dataset root: {pid_root}")
+    label_dirs = sorted([entry.name.lower() for entry in pid_root.iterdir() if entry.is_dir()])
+    if len(label_dirs) != 1:
+        raise ValueError(f"Expected exactly one class dir for pid {pid}, got {label_dirs}")
+    label = label_dirs[0]
+    if label not in {"positive", "neutral", "negative"}:
+        raise ValueError(f"Unexpected label directory for pid {pid}: {label}")
+    return label
+
+
+@click.command()
+@click.option(
+    "--base-partition",
+    type=click.Path(path_type=Path, exists=True, dir_okay=False),
+    required=True,
+    help="Path to the source partition JSON, e.g. datasets/Scoliosis1K/Scoliosis1K_118.json",
+)
+@click.option(
+    "--dataset-root",
+    type=click.Path(path_type=Path, exists=True, file_okay=False),
+    required=True,
+    help="Dataset root used to infer each pid class label, e.g. /mnt/public/data/Scoliosis1K/Scoliosis1K-sil-pkl",
+)
+@click.option(
+    "--negative-multiplier",
+    type=int,
+    required=True,
+    help="Target negative count as a multiple of the positive/neutral count, e.g. 2 for 1:1:2",
+)
+@click.option(
+    "--output-path",
+    type=click.Path(path_type=Path, dir_okay=False),
+    required=True,
+    help="Path to write the derived partition JSON.",
+)
+@click.option(
+    "--seed",
+    type=int,
+    default=118,
+    show_default=True,
+    help="Random seed used when downsampling negatives.",
+)
+def main(
+    base_partition: Path,
+    dataset_root: Path,
+    negative_multiplier: int,
+    output_path: Path,
+    seed: int,
+) -> None:
+    with base_partition.open("r", encoding="utf-8") as handle:
+        partition = cast(Partition, json.load(handle))
+
+    train_ids = list(partition["TRAIN_SET"])
+    test_ids = list(partition["TEST_SET"])
+
+    train_by_label: dict[str, list[str]] = {"positive": [], "neutral": [], "negative": []}
+    for pid in train_ids:
+        label = infer_pid_label(dataset_root, pid)
+        train_by_label[label].append(pid)
+
+    pos_count = len(train_by_label["positive"])
+    neu_count = len(train_by_label["neutral"])
+    neg_count = len(train_by_label["negative"])
+    if pos_count != neu_count:
+        raise ValueError(
+            "This helper assumes equal positive/neutral train counts so that only "
+            + f"negative downsampling changes the ratio. Got positive={pos_count}, neutral={neu_count}."
+        )
+
+    target_negative_count = negative_multiplier * pos_count
+    if target_negative_count > neg_count:
+        raise ValueError(
+            f"Requested {target_negative_count} negatives but only {neg_count} are available "
+            + f"in base partition {base_partition}."
+        )
+
+    rng = random.Random(seed)
+    sampled_negatives = sorted(rng.sample(train_by_label["negative"], target_negative_count))
+    derived_train = (
+        sorted(train_by_label["positive"])
+        + sorted(train_by_label["neutral"])
+        + sampled_negatives
+    )
+
+    derived_partition = {
+        "TRAIN_SET": derived_train,
+        "TEST_SET": test_ids,
+    }
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    with output_path.open("w", encoding="utf-8") as handle:
+        json.dump(derived_partition, handle, indent=2)
+        _ = handle.write("\n")
+
+    train_counts = Counter(infer_pid_label(dataset_root, pid) for pid in derived_train)
+    test_counts = Counter(infer_pid_label(dataset_root, pid) for pid in test_ids)
+    click.echo(f"wrote {output_path}")
+    click.echo(f"train_counts={dict(train_counts)}")
+    click.echo(f"test_counts={dict(test_counts)}")
+
+
+if __name__ == "__main__":
+    main()