Refine DRF preprocessing and body-prior pipeline

configs/drf/drf_scoliosis1k.yaml

@@ -1,6 +1,6 @@
 data_cfg:
   dataset_name: Scoliosis1K
-  dataset_root: /mnt/public/data/Scoliosis1K/Scoliosis1K-drf-pkl-118
+  dataset_root: /mnt/public/data/Scoliosis1K/Scoliosis1K-drf-pkl-118-paper
   dataset_partition: ./datasets/Scoliosis1K/Scoliosis1K_118.json
   num_workers: 1
   remove_no_gallery: false
@@ -10,7 +10,7 @@ evaluator_cfg:
   enable_float16: true
   restore_ckpt_strict: true
   restore_hint: 20000
-  save_name: DRF
+  save_name: DRF_paper
   eval_func: evaluate_scoliosis
   sampler:
     batch_shuffle: false
@@ -19,7 +19,7 @@ evaluator_cfg:
     frames_all_limit: 720
   metric: euc
   transform:
-    - type: BaseSilCuttingTransform
+    - type: BaseSilTransform
     - type: NoOperation
 
 loss_cfg:
@@ -90,7 +90,7 @@ trainer_cfg:
   restore_ckpt_strict: true
   restore_hint: 0
   save_iter: 20000
-  save_name: DRF
+  save_name: DRF_paper
   sync_BN: true
   total_iter: 20000
   sampler:
@@ -102,5 +102,5 @@ trainer_cfg:
     sample_type: fixed_unordered
     type: TripletSampler
   transform:
-    - type: BaseSilCuttingTransform
+    - type: BaseSilTransform
     - type: NoOperation

configs/drf/drf_scoliosis1k_eval_1gpu.yaml

@@ -0,0 +1,106 @@
+data_cfg:
+  dataset_name: Scoliosis1K
+  dataset_root: /mnt/public/data/Scoliosis1K/Scoliosis1K-drf-pkl-118-paper
+  dataset_partition: ./datasets/Scoliosis1K/Scoliosis1K_118.json
+  num_workers: 1
+  remove_no_gallery: false
+  test_dataset_name: Scoliosis1K
+
+evaluator_cfg:
+  enable_float16: true
+  restore_ckpt_strict: true
+  restore_hint: 20000
+  save_name: DRF_paper
+  eval_func: evaluate_scoliosis
+  sampler:
+    batch_shuffle: false
+    batch_size: 1
+    sample_type: all_ordered
+    frames_all_limit: 720
+  metric: euc
+  transform:
+    - type: BaseSilTransform
+    - type: NoOperation
+
+loss_cfg:
+  - loss_term_weight: 1.0
+    margin: 0.2
+    type: TripletLoss
+    log_prefix: triplet
+  - loss_term_weight: 1.0
+    scale: 16
+    type: CrossEntropyLoss
+    log_prefix: softmax
+    log_accuracy: true
+
+model_cfg:
+  model: DRF
+  num_pairs: 8
+  num_metrics: 3
+  backbone_cfg:
+    type: ResNet9
+    block: BasicBlock
+    in_channel: 2
+    channels:
+      - 64
+      - 128
+      - 256
+      - 512
+    layers:
+      - 1
+      - 1
+      - 1
+      - 1
+    strides:
+      - 1
+      - 2
+      - 2
+      - 1
+    maxpool: false
+  SeparateFCs:
+    in_channels: 512
+    out_channels: 256
+    parts_num: 16
+  SeparateBNNecks:
+    class_num: 3
+    in_channels: 256
+    parts_num: 16
+  bin_num:
+    - 16
+
+optimizer_cfg:
+  lr: 0.1
+  momentum: 0.9
+  solver: SGD
+  weight_decay: 0.0005
+
+scheduler_cfg:
+  gamma: 0.1
+  milestones:
+    - 10000
+    - 14000
+    - 18000
+  scheduler: MultiStepLR
+
+trainer_cfg:
+  enable_float16: true
+  fix_BN: false
+  with_test: false
+  log_iter: 100
+  restore_ckpt_strict: true
+  restore_hint: 0
+  save_iter: 20000
+  save_name: DRF_paper
+  sync_BN: true
+  total_iter: 20000
+  sampler:
+    batch_shuffle: true
+    batch_size:
+      - 8
+      - 8
+    frames_num_fixed: 30
+    sample_type: fixed_unordered
+    type: TripletSampler
+  transform:
+    - type: BaseSilTransform
+    - type: NoOperation

configs/drf/drf_scoliosis1k_smoke.yaml

@@ -1,6 +1,6 @@
 data_cfg:
   dataset_name: Scoliosis1K
-  dataset_root: /mnt/public/data/Scoliosis1K/Scoliosis1K-drf-pkl-118
+  dataset_root: /mnt/public/data/Scoliosis1K/Scoliosis1K-drf-pkl-118-paper
   dataset_partition: ./datasets/Scoliosis1K/Scoliosis1K_118.json
   num_workers: 1
   remove_no_gallery: false
@@ -14,7 +14,7 @@ evaluator_cfg:
   enable_float16: true
   restore_ckpt_strict: true
   restore_hint: 0
-  save_name: DRF_smoke
+  save_name: DRF_paper_smoke
   eval_func: evaluate_scoliosis
   sampler:
     batch_shuffle: false
@@ -23,7 +23,7 @@ evaluator_cfg:
     frames_all_limit: 720
   metric: euc
   transform:
-    - type: BaseSilCuttingTransform
+    - type: BaseSilTransform
     - type: NoOperation
 
 loss_cfg:
@@ -97,7 +97,7 @@ trainer_cfg:
   scheduler_reset: false
   restore_hint: 0
   save_iter: 1
-  save_name: DRF_smoke
+  save_name: DRF_paper_smoke
   sync_BN: true
   total_iter: 1
   sampler:
@@ -109,5 +109,5 @@ trainer_cfg:
     sample_type: fixed_unordered
     type: TripletSampler
   transform:
-    - type: BaseSilCuttingTransform
+    - type: BaseSilTransform
     - type: NoOperation

configs/drf/pretreatment_heatmap_drf.yaml

@@ -0,0 +1,26 @@
+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
+  target_body_height: ${norm_args.heatmap_image_height}
+
+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

@@ -99,14 +99,22 @@ The PAV pass is implemented from the paper:
 4. compute vertical, midline, and angular deviations for the 8 symmetric joint pairs
 5. apply IQR filtering per metric
 6. average over time
-7. min-max normalize across the dataset, or across `TRAIN_SET` when `--stats_partition` is provided
+7. min-max normalize across the full dataset (paper default), or across `TRAIN_SET` when `--stats_partition` is provided as an anti-leakage variant
 
 Run:
 
 ```bash
 uv run python datasets/pretreatment_scoliosis_drf.py \
   --pose_data_path=<path_to_pose_pkl> \
-  --output_path=<path_to_drf_pkl> \
+  --output_path=<path_to_drf_pkl>
+```
+
+To reproduce the paper defaults more closely, the script now uses
+`configs/drf/pretreatment_heatmap_drf.yaml` by default, which enables
+summed two-channel skeleton maps and a literal 128-pixel height normalization.
+If you explicitly want train-only PAV min-max statistics, add:
+
+```bash
   --stats_partition=./datasets/Scoliosis1K/Scoliosis1K_118.json
 ```
 

datasets/pretreatment_heatmap.py

@@ -118,8 +118,8 @@ class GeneratePoseTarget:
             ed_x = min(tmp_ed_x + 1, img_w)
             st_y = max(tmp_st_y, 0)
             ed_y = min(tmp_ed_y + 1, img_h)
-            x = np.arange(st_x, ed_x, 1, np.float32)
-            y = np.arange(st_y, ed_y, 1, np.float32)
+            x = np.arange(st_x, ed_x, dtype=np.float32)
+            y = np.arange(st_y, ed_y, dtype=np.float32)
 
             # if the keypoint not in the heatmap coordinate system
             if not (len(x) and len(y)):
@@ -166,8 +166,8 @@ class GeneratePoseTarget:
             min_y = max(tmp_min_y, 0)
             max_y = min(tmp_max_y + 1, img_h)
 
-            x = np.arange(min_x, max_x, 1, np.float32)
-            y = np.arange(min_y, max_y, 1, np.float32)
+            x = np.arange(min_x, max_x, dtype=np.float32)
+            y = np.arange(min_y, max_y, dtype=np.float32)
 
             if not (len(x) and len(y)):
                 continue
@@ -324,9 +324,37 @@ class HeatmapToImage:
         heatmaps = [cv2.resize(x, (neww, newh)) for x in heatmaps]
         return np.ascontiguousarray(np.mean(np.array(heatmaps), axis=-1, keepdims=True).transpose(0,3,1,2))
 
+
+class HeatmapReducer:
+    """Reduce stacked joint/limb heatmaps to a single grayscale channel."""
+
+    def __init__(self, reduction: str = "max") -> None:
+        if reduction not in {"max", "sum"}:
+            raise ValueError(f"Unsupported heatmap reduction: {reduction}")
+        self.reduction = reduction
+
+    def __call__(self, heatmaps: np.ndarray) -> np.ndarray:
+        """
+        heatmaps: (T, C, H, W)
+        return: (T, 1, H, W)
+        """
+        if self.reduction == "max":
+            reduced = np.max(heatmaps, axis=1, keepdims=True)
+            reduced = np.clip(reduced, 0.0, 1.0)
+            return (reduced * 255).astype(np.uint8)
+
+        reduced = np.sum(heatmaps, axis=1, keepdims=True)
+        return (reduced * 255.0).astype(np.float32)
+
 class CenterAndScaleNormalizer:
 
-    def __init__(self, pose_format="coco", use_conf=True, heatmap_image_height=128) -> None:
+    def __init__(
+        self,
+        pose_format="coco",
+        use_conf=True,
+        heatmap_image_height=128,
+        target_body_height=None,
+    ) -> None:
         """
         Parameters:
         - pose_format (str): Specifies the format of the keypoints. 
@@ -334,10 +362,13 @@ class CenterAndScaleNormalizer:
                             The supported formats are "coco" or "openpose-x" where 'x' can be either 18 or 25, indicating the number of keypoints used by the OpenPose model. 
         - use_conf (bool): Indicates whether confidence scores.
         - heatmap_image_height (int): Sets the height (in pixels) for the heatmap images that will be normlization. 
+        - target_body_height (float | None): Optional normalized body height. When omitted,
+          preserve the historical SkeletonGait scaling heuristic.
         """
         self.pose_format = pose_format
         self.use_conf = use_conf
         self.heatmap_image_height = heatmap_image_height
+        self.target_body_height = target_body_height
 
     def __call__(self, data):
         """
@@ -369,7 +400,13 @@ class CenterAndScaleNormalizer:
         # Scale-normalization
         y_max = np.max(pose_seq[:, :, 1], axis=-1) # [t]
         y_min = np.min(pose_seq[:, :, 1], axis=-1) # [t]
-        pose_seq *= ((self.heatmap_image_height // 1.5) / (y_max - y_min)[:, np.newaxis, np.newaxis]) # [t, v, 2]
+        target_body_height = (
+            float(self.target_body_height)
+            if self.target_body_height is not None
+            else float(self.heatmap_image_height // 1.5)
+        )
+        body_height = np.maximum(y_max - y_min, 1e-6)
+        pose_seq *= (target_body_height / body_height)[:, np.newaxis, np.newaxis] # [t, v, 2]
         
         pose_seq += self.heatmap_image_height // 2
         
@@ -523,16 +560,21 @@ class HeatmapAlignment():
         heatmap_imgs: (T, 1, raw_size, raw_size)
         return (T, 1, final_img_size, final_img_size)
         """
-        heatmap_imgs = heatmap_imgs / 255.
-        heatmap_imgs = np.array([self.center_crop(heatmap_img) for heatmap_img in heatmap_imgs]) 
-        return (heatmap_imgs * 255).astype('uint8')
+        original_dtype = heatmap_imgs.dtype
+        heatmap_imgs = heatmap_imgs.astype(np.float32) / 255.0
+        heatmap_imgs = np.array([self.center_crop(heatmap_img) for heatmap_img in heatmap_imgs], dtype=np.float32)
+        heatmap_imgs = heatmap_imgs * 255.0
+        if np.issubdtype(original_dtype, np.integer):
+            return np.clip(heatmap_imgs, 0.0, 255.0).astype(original_dtype)
+        return heatmap_imgs.astype(original_dtype)
 
 def GenerateHeatmapTransform(
     coco18tococo17_args,
     padkeypoints_args,
     norm_args,
     heatmap_generator_args,
-    align_args
+    align_args,
+    reduction="max",
 ):
 
     base_transform = T.Compose([
@@ -545,7 +587,7 @@ def GenerateHeatmapTransform(
     heatmap_generator_args["with_kp"] = False
     transform_bone = T.Compose([
         GeneratePoseTarget(**heatmap_generator_args), 
-        HeatmapToImage(), 
+        HeatmapReducer(reduction=reduction),
         HeatmapAlignment(**align_args) 
     ])
 
@@ -553,7 +595,7 @@ def GenerateHeatmapTransform(
     heatmap_generator_args["with_kp"] = True
     transform_joint = T.Compose([
         GeneratePoseTarget(**heatmap_generator_args), 
-        HeatmapToImage(), 
+        HeatmapReducer(reduction=reduction),
         HeatmapAlignment(**align_args) 
     ])
 

datasets/pretreatment_scoliosis_drf.py

@@ -7,7 +7,7 @@ import pickle
 import sys
 from glob import glob
 from pathlib import Path
-from typing import Any, TypedDict
+from typing import Any, TypedDict, cast
 
 import numpy as np
 import yaml
@@ -63,14 +63,17 @@ def get_args() -> argparse.Namespace:
     _ = parser.add_argument(
         "--heatmap_cfg_path",
         type=str,
-        default="configs/skeletongait/pretreatment_heatmap.yaml",
+        default="configs/drf/pretreatment_heatmap_drf.yaml",
         help="Heatmap preprocessing config used to build the skeleton map branch.",
     )
     _ = parser.add_argument(
         "--stats_partition",
         type=str,
         default=None,
-        help="Optional dataset partition JSON. When set, PAV min/max stats use TRAIN_SET ids only.",
+        help=(
+            "Optional dataset partition JSON. When set, PAV min/max stats use TRAIN_SET ids only. "
+            "Omit it to match the paper's dataset-level min-max normalization."
+        ),
     )
     return parser.parse_args()
 
@@ -79,7 +82,9 @@ def load_heatmap_cfg(cfg_path: str) -> dict[str, Any]:
     with open(cfg_path, "r", encoding="utf-8") as stream:
         cfg = yaml.safe_load(stream)
     replaced = heatmap_prep.replace_variables(cfg, cfg)
-    return dict(replaced)
+    if not isinstance(replaced, dict):
+        raise TypeError(f"Expected heatmap config dict from {cfg_path}, got {type(replaced).__name__}")
+    return cast(dict[str, Any], replaced)
 
 
 def build_pose_transform(cfg: dict[str, Any]) -> T.Compose:
@@ -175,6 +180,7 @@ def main() -> None:
         norm_args=heatmap_cfg["norm_args"],
         heatmap_generator_args=heatmap_cfg["heatmap_generator_args"],
         align_args=heatmap_cfg["align_args"],
+        reduction="sum",
     )
 
     pose_paths = iter_pose_paths(args.pose_data_path)

opengait/modeling/base_model_body.py

@@ -0,0 +1,82 @@
+from __future__ import annotations
+
+from typing import Any, Callable, cast
+
+import numpy as np
+import torch
+from jaxtyping import Float, Int
+
+from .base_model import BaseModel
+from opengait.utils.common import list2var, np2var
+
+
+class BaseModelBody(BaseModel):
+    """Base model variant with a separate sequence-level body-prior input."""
+
+    def inputs_pretreament(
+        self,
+        inputs: tuple[list[np.ndarray], list[int], list[str], list[str], np.ndarray | None],
+    ) -> Any:
+        seqs_batch, labs_batch, typs_batch, vies_batch, seqL_batch = inputs
+        seq_trfs = cast(
+            list[Callable[[Any], Any]],
+            self.trainer_trfs if self.training else self.evaluator_trfs,
+        )
+        if len(seqs_batch) != len(seq_trfs):
+            raise ValueError(
+                "The number of types of input data and transform should be same. "
+                f"But got {len(seqs_batch)} and {len(seq_trfs)}"
+            )
+        if len(seqs_batch) < 2:
+            raise ValueError("BaseModelBody expects one visual input and one body-prior input.")
+
+        requires_grad = bool(self.training)
+        visual_seqs = [
+            np2var(
+                np.asarray([trf(fra) for fra in seq]),
+                requires_grad=requires_grad,
+            ).float()
+            for trf, seq in zip(seq_trfs[:-1], seqs_batch[:-1])
+        ]
+        body_trf = seq_trfs[-1]
+        body_seq = np2var(
+            np.asarray([body_trf(fra) for fra in seqs_batch[-1]]),
+            requires_grad=requires_grad,
+        ).float()
+
+        labs = list2var(labs_batch).long()
+        seqL = np2var(seqL_batch).int() if seqL_batch is not None else None
+
+        body_features = aggregate_body_features(body_seq, seqL)
+
+        if seqL is not None:
+            seqL_sum = int(seqL.sum().data.cpu().numpy())
+            ipts = [_[:, :seqL_sum] for _ in visual_seqs]
+        else:
+            ipts = visual_seqs
+        return ipts, labs, typs_batch, vies_batch, seqL, body_features
+
+
+def aggregate_body_features(
+    sequence_features: Float[torch.Tensor, "..."],
+    seqL: Int[torch.Tensor, "1 batch"] | None,
+) -> Float[torch.Tensor, "batch pairs metrics"]:
+    """Collapse a sampled body-prior sequence back to one vector per sequence."""
+
+    if seqL is None:
+        if sequence_features.ndim < 3:
+            raise ValueError(f"Expected body prior with >=3 dims, got shape {tuple(sequence_features.shape)}")
+        return sequence_features.mean(dim=1)
+
+    if sequence_features.ndim < 4:
+        raise ValueError(f"Expected packed body prior with >=4 dims, got shape {tuple(sequence_features.shape)}")
+
+    lengths = seqL[0].tolist()
+    flattened = sequence_features.squeeze(0)
+    aggregated: list[torch.Tensor] = []
+    start = 0
+    for length in lengths:
+        end = start + int(length)
+        aggregated.append(flattened[start:end].mean(dim=0))
+        start = end
+    return torch.stack(aggregated, dim=0)

opengait/modeling/models/drf.py

@@ -8,7 +8,7 @@ from jaxtyping import Float, Int
 
 from einops import rearrange
 
-from ..base_model import BaseModel
+from ..base_model_body import BaseModelBody
 from ..modules import (
     HorizontalPoolingPyramid,
     PackSequenceWrapper,
@@ -18,7 +18,7 @@ from ..modules import (
 )
 
 
-class DRF(BaseModel):
+class DRF(BaseModelBody):
     """Dual Representation Framework from arXiv:2509.00872v1."""
 
     def build_network(self, model_cfg: dict[str, Any]) -> None:
@@ -43,9 +43,10 @@ class DRF(BaseModel):
             list[str],
             list[str],
             Int[torch.Tensor, "1 batch"] | None,
+            Float[torch.Tensor, "batch pairs metrics"],
         ],
     ) -> dict[str, dict[str, Any]]:
-        ipts, pids, labels, _, seqL = inputs
+        ipts, pids, labels, _, seqL, key_features = inputs
         label_ids = torch.as_tensor(
             [LABEL_MAP[str(label).lower()] for label in labels],
             device=pids.device,
@@ -58,15 +59,12 @@ class DRF(BaseModel):
         else:
             heatmaps = rearrange(heatmaps, "n s c h w -> n c s h w")
 
-        pav_seq = ipts[1]
-        pav = aggregate_sequence_features(pav_seq, seqL)
-
         outs = self.Backbone(heatmaps)
         outs = self.TP(outs, seqL, options={"dim": 2})[0]
 
         feat = self.HPP(outs)
         embed_1 = self.FCs(feat)
-        embed_1 = self.PGA(embed_1, pav)
+        embed_1 = self.PGA(embed_1, key_features)
 
         embed_2, logits = self.BNNecks(embed_1)
         del embed_2
@@ -120,24 +118,6 @@ class PAVGuidedAttention(nn.Module):
         return embeddings * channel_att * spatial_att
 
 
-def aggregate_sequence_features(
-    sequence_features: Float[torch.Tensor, "batch seq pairs metrics"],
-    seqL: Int[torch.Tensor, "1 batch"] | None,
-) -> Float[torch.Tensor, "batch pairs metrics"]:
-    if seqL is None:
-        return sequence_features.mean(dim=1)
-
-    lengths = seqL[0].tolist()
-    flattened = sequence_features.squeeze(0)
-    aggregated = []
-    start = 0
-    for length in lengths:
-        end = start + int(length)
-        aggregated.append(flattened[start:end].mean(dim=0))
-        start = end
-    return torch.stack(aggregated, dim=0)
-
-
 LABEL_MAP: dict[str, int] = {
     "negative": 0,
     "neutral": 1,

scripts/monitor_drf_jobs.py

@@ -0,0 +1,145 @@
+from __future__ import annotations
+
+import argparse
+import os
+import sys
+import time
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import Final
+
+
+ERROR_PATTERNS: Final[tuple[str, ...]] = (
+    "traceback",
+    "runtimeerror",
+    "error:",
+    "exception",
+    "failed",
+    "segmentation fault",
+    "killed",
+)
+
+
+@dataclass(frozen=True)
+class JobSpec:
+    name: str
+    pid: int
+    log_path: Path
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Monitor long-running DRF preprocess/train jobs.")
+    parser.add_argument("--preprocess-pid", type=int, required=True)
+    parser.add_argument("--preprocess-log", type=Path, required=True)
+    parser.add_argument("--launcher-pid", type=int, required=True)
+    parser.add_argument("--launcher-log", type=Path, required=True)
+    parser.add_argument("--sentinel-path", type=Path, required=True)
+    parser.add_argument("--status-log", type=Path, required=True)
+    parser.add_argument("--poll-seconds", type=float, default=30.0)
+    return parser.parse_args()
+
+
+def pid_alive(pid: int) -> bool:
+    return Path(f"/proc/{pid}").exists()
+
+
+def read_tail(path: Path, limit: int = 8192) -> str:
+    if not path.exists():
+        return ""
+    with path.open("rb") as handle:
+        handle.seek(0, os.SEEK_END)
+        size = handle.tell()
+        handle.seek(max(size - limit, 0), os.SEEK_SET)
+        data = handle.read()
+    return data.decode("utf-8", errors="replace")
+
+
+def detect_error(log_text: str) -> str | None:
+    lowered = log_text.lower()
+    for pattern in ERROR_PATTERNS:
+        if pattern in lowered:
+            return pattern
+    return None
+
+
+def append_status(path: Path, line: str) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with path.open("a", encoding="utf-8") as handle:
+        handle.write(f"{datetime.now().isoformat(timespec='seconds')} {line}\n")
+
+
+def monitor_job(job: JobSpec, status_log: Path) -> str | None:
+    tail = read_tail(job.log_path)
+    error = detect_error(tail)
+    if error is not None:
+        append_status(status_log, f"[alert] {job.name}: detected `{error}` in {job.log_path}")
+        return f"{job.name} log shows `{error}`"
+    return None
+
+
+def main() -> int:
+    args = parse_args()
+    preprocess = JobSpec("preprocess", args.preprocess_pid, args.preprocess_log)
+    launcher = JobSpec("launcher", args.launcher_pid, args.launcher_log)
+
+    append_status(
+        args.status_log,
+        (
+            "[start] monitoring "
+            f"preprocess_pid={preprocess.pid} launcher_pid={launcher.pid} "
+            f"sentinel={args.sentinel_path}"
+        ),
+    )
+
+    preprocess_seen_alive = False
+    launcher_seen_alive = False
+
+    while True:
+        preprocess_alive = pid_alive(preprocess.pid)
+        launcher_alive = pid_alive(launcher.pid)
+        preprocess_seen_alive = preprocess_seen_alive or preprocess_alive
+        launcher_seen_alive = launcher_seen_alive or launcher_alive
+
+        preprocess_error = monitor_job(preprocess, args.status_log)
+        if preprocess_error is not None:
+            print(preprocess_error, file=sys.stderr)
+            return 1
+
+        launcher_error = monitor_job(launcher, args.status_log)
+        if launcher_error is not None:
+            print(launcher_error, file=sys.stderr)
+            return 1
+
+        sentinel_ready = args.sentinel_path.exists()
+        append_status(
+            args.status_log,
+            (
+                "[ok] "
+                f"preprocess_alive={preprocess_alive} "
+                f"launcher_alive={launcher_alive} "
+                f"sentinel_ready={sentinel_ready}"
+            ),
+        )
+
+        if preprocess_seen_alive and not preprocess_alive and not sentinel_ready:
+            append_status(args.status_log, "[alert] preprocess exited before sentinel was written")
+            print("preprocess exited before sentinel was written", file=sys.stderr)
+            return 1
+
+        launcher_tail = read_tail(launcher.log_path)
+        train_started = "[start]" in launcher_tail
+
+        if launcher_seen_alive and not launcher_alive:
+            if not train_started and not sentinel_ready:
+                append_status(args.status_log, "[alert] launcher exited before training started")
+                print("launcher exited before training started", file=sys.stderr)
+                return 1
+            append_status(args.status_log, "[done] launcher exited; monitoring complete")
+            return 0
+
+        time.sleep(args.poll_seconds)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())