hnerv_qlp

[Bucky789]

submission name:

hnerv_qlp

upload zipped archive.zip

Hosted as a release asset (curl -L works, SHA-256 verified):
https://github.com/Bucky789/comma_video_compression_challenge/releases/download/hnerv-qlp-v1/archive.zip

SHA-256 ebd513903bd598b4d73d699a11c89600bd11747f27aab26737e518096675b813, 176,525 bytes, single ZIP member x (ZIP_STORED).

report.txt

=== Evaluation results over 600 samples ===
  Average PoseNet Distortion: 0.00003000
  Average SegNet Distortion: 0.00056085
  Submission file size: 176,525 bytes
  Original uncompressed size: 37,545,489 bytes
  Compression Rate: 0.00470163
  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 0.190946

(evaluate.py prints the score to 2 dp as 0.19; full precision is 0.190946.)

does your submission require gpu for evaluation (inflation)?

no — inflate is CPU-only (device pinned to CPU); deps torch, numpy, constriction.

did you include the compression script? and want it to be merged?

yes — the full offline encoder is in submissions/hnerv_qlp/encoder/ (extract → polish → pack). Happy for it to be merged.

is this submission competitive or innovative? explain why

Both.

Competitive: CPU score 0.190946 vs the current #1 (rhnerv_comma, 0.191126) — a margin of -0.000180. Local eval reproduces the #1's published metrics to ~1e-5, so the margin is real (~18x that).

	SegNet	PoseNet	bytes	score
#1 rhnerv_comma	0.00056023	0.00002943	177,136	0.191126
hnerv_qlp	0.00056085	0.00003000	176,525	0.190946

Innovative: every top submission reuses PR #95's decoder and latents; the only per-pair latent tuning attempted so far is PR #101's sidecar (one dimension, a small fixed step table). This submission generalizes that to quantization-aware gradient descent over all 28 latent dimensions of every pair, decoder frozen, optimized directly against SegNet/PoseNet through the exact inflate chain. The polished latents reach the #1's distortion without its 607-byte sidecar, so the archive is 611 bytes smaller.

additional comments

No decoder training. Decoder weights (PR #95) and the FEC6 selector (PR #110) are reused bit-for-bit; the ctx container is PR #112's, with the latent section re-encoded and the sidecar dropped. One implementation note that mattered: the SegNet distortion is a discrete argmax, so fp16 and fp32 disagree on borderline pixels — selecting latents on an fp16 scorer overfits the GPU and regresses on the CPU axis (an fp16-selected build scored 0.191955). Evaluation and per-pair selection are done in fp32; training stays fp16 for speed. Full lineage and attribution in submissions/hnerv_qlp/THIRD_PARTY_NOTICES.md (PRs #95/#98/#101/#110/#112).

[github-actions]

Thanks for the submission @Bucky789! 🤏

A maintainer will review your PR shortly.

To run the evaluation, a maintainer will trigger the eval workflow with your PR number.

[github-actions]

Eval Results: hnerv_qlp

=== Evaluation config ===
  batch_size: 16
  device: cpu
  num_threads: 2
  prefetch_queue_depth: 4
  report: submissions/hnerv_qlp/report.txt
  seed: 1234
  submission_dir: submissions/hnerv_qlp
  uncompressed_dir: /home/runner/work/comma_video_compression_challenge/comma_video_compression_challenge/videos
  video_names_file: /home/runner/work/comma_video_compression_challenge/comma_video_compression_challenge/public_test_video_names.txt
=== Evaluation results over 600 samples ===
  Average PoseNet Distortion: 0.00003000
  Average SegNet Distortion: 0.00056082
  Submission file size: 176,525 bytes
  Original uncompressed size: 37,545,489 bytes
  Compression Rate: 0.00470163
  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 0.19

[YassineYousfi]

added to leaderboard, congrats
the code has too much overlap with code that is already merged though, so we won't merge it unless you refactor to reuse bits and pieces that are already there
feel free to reopen

[Bucky789]

I have emailed the email address mentioned as I'm looking for a Full Stack Engineer job.
Will refactor the code so that it can get merged.
Thank You.