floppyrj45/cosma-qc
2
0
Fork 0
Code Issues 2 Pull Requests Actions Packages Projects Releases Wiki Activity

dispatcher — window_size adaptatif + skip video courtes <8min

Browse Source 
window_size selon len effectif (frame_count / stride):
  <= 320 -> 16/2   (petit overhead)
  <= 3000 -> 32/8
  > 3000 -> 64/16  (moins de windows sur longues sequences)

Skip auto si video_duration_s < COSMA_QC_MIN_VIDEO_S (default 480s = 8min):
  segments trop courts ne contiennent pas de plongee exploitable.
  Reco user: videos < 8 min sont inutiles pour la reconstruction.
This commit is contained in:
Flag
2026-04-22 21:34:51 +00:00
parent 9dd6a82d08
commit 033abc41c5
1 changed files with 16 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
scripts/dispatcher.py
View File

@@ -349,6 +349,12 @@ def do_extract(job: sqlite3.Row, worker: dict) -> str:
# Persist the measured video duration so the dashboard shows real length (segment_label # Persist the measured video duration so the dashboard shows real length (segment_label
# from ingest is only the timestamp of the first MP4 and lies when a segment spans multiple). # from ingest is only the timestamp of the first MP4 and lies when a segment spans multiple).
set_status(job["id"], video_duration_s=total_duration_s) set_status(job["id"], video_duration_s=total_duration_s)
# Skip segments that are too short to contain a meaningful dive.
min_video_s = int(os.environ.get("COSMA_QC_MIN_VIDEO_S", "480")) # 8 min default
if total_duration_s < min_video_s:
print(f" ↳ job #{job['id']}: video too short ({int(total_duration_s)}s < {min_video_s}s) — marking skipped")
set_status(job["id"], status="skipped", error=f"too short: {int(total_duration_s)}s of video")
raise RuntimeError("skipped_short")
# Drop the hors-eau prefix AND suffix before reconstruction — AUV is out-of-water at both ends. # Drop the hors-eau prefix AND suffix before reconstruction — AUV is out-of-water at both ends.
head, tail, remaining = trim_above_water_prefix(worker, frames_dir) head, tail, remaining = trim_above_water_prefix(worker, frames_dir)
if head or tail: if head or tail:
@@ -382,12 +388,21 @@ def do_reconstruct(job: sqlite3.Row, worker: dict, frames_dir: str) -> tuple[str
# demo.py starts a viser web server after saving the PLY and never exits. # demo.py starts a viser web server after saving the PLY and never exits.
# Wrap it: launch in bg, wait for "PLY saved" marker in the log, kill, exit 0. # Wrap it: launch in bg, wait for "PLY saved" marker in the log, kill, exit 0.
# Match on the unique job frames_dir to identify our demo.py among all children/threads. # Match on the unique job frames_dir to identify our demo.py among all children/threads.
# Adapt window size to sequence length (lingbot-map README recommendation): bigger windows
# reduce overhead on long sequences. Effective frame count = frame_count / stride.
eff = frame_count // max(1, stride) if frame_count else 0
if eff > 3000:
window_size, overlap_size = 64, 16
elif eff > 320:
window_size, overlap_size = 32, 8
else:
window_size, overlap_size = 16, 2
marker = shlex.quote(frames_dir) marker = shlex.quote(frames_dir)
cmd = ( cmd = (
f"cd {shlex.quote(worker['lingbot_path'])} && source .venv/bin/activate && " f"cd {shlex.quote(worker['lingbot_path'])} && source .venv/bin/activate && "
f"setsid python3 demo.py --model_path {shlex.quote(ckpt)} " f"setsid python3 demo.py --model_path {shlex.quote(ckpt)} "
f"--image_folder {shlex.quote(frames_dir)} --port {port} " f"--image_folder {shlex.quote(frames_dir)} --port {port} "
f"--stride {stride} --use_sdpa --mode windowed --window_size 16 --overlap_size 2 --offload_to_cpu " f"--stride {stride} --use_sdpa --mode windowed --window_size {window_size} --overlap_size {overlap_size} --offload_to_cpu "
f"--save_ply {shlex.quote(ply_path)} > {log} 2>&1 & " f"--save_ply {shlex.quote(ply_path)} > {log} 2>&1 & "
f"DEMO_PID=$!; " f"DEMO_PID=$!; "
f"for i in $(seq 1 3600); do " f"for i in $(seq 1 3600); do "