hnerv_stage9: 9-stage curriculum with latent polish + LZMA codec#123
Closed
Bucky789 wants to merge 9 commits into
Closed
hnerv_stage9: 9-stage curriculum with latent polish + LZMA codec#123Bucky789 wants to merge 9 commits into
Bucky789 wants to merge 9 commits into
Conversation
Add av, timm, einops, segmentation-models-pytorch, safetensors alongside brotli. modules.py requires all five to load. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
batch_size=4 underutilizes A100 40GB (same speed as RTX 4060). batch_size=16 fills GPU parallelism, ~4x faster per epoch. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
search_order=False during training evals saves ~60s per eval. With 1466 evals total, this saves ~24h. Greedy search still runs in codec_stage for the final archive build. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Stage 2: 5650→3500, Stage 3: 1500→1000, Stage 5: 9000→5000, Stage 6: 2000→1000, Stage 7: 3000→1500, Stage 8: 10000→6000. Total: 23,500 epochs = ~36.6h = ~194 compute units. Stage 9 (latent polish) kept at 2000 — our key innovation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
S1:2500 S2:2000 S3:500 S4:500 S5:4000 S6:750 S7:750 S8:5000 S9:2000 Total 18,000 epochs = ~28h = ~148 CU. Stage 8 (5k) matches original hnerv_muon. Stage 9 (2k latent polish) kept intact. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…e fix) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
Thanks for the submission @Bucky789! 🤏 A maintainer will review your PR shortly. To run the evaluation, a maintainer will trigger the |
Eval Results:
|
Author
|
Closing - superseding this with a reworked submission. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
submission name:
hnerv_stage9
upload zipped
archive.ziparchive.zip
report.txt
=== Evaluation config === batch_size: 16 device: cpu num_threads: 2 prefetch_queue_depth: 4 report: submissions/hnerv_stage9/report.txt seed: 1234 submission_dir: submissions/hnerv_stage9 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: 0.00010009 Average SegNet Distortion: 0.00097826 Submission file size: 179,539 bytes Original uncompressed size: 37,545,489 bytes Compression Rate: 0.00478191 Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 0.25does your submission require gpu for evaluation (inflation)?
no
did you include the compression script? and want it to be merged?
yes
is this submission competitive or innovative? explain why
Both. Novel Stage 9 latent polish: freeze the decoder after 8-stage Muon curriculum, then gradient-descend all 28
latent dims against the frozen quantized decoder for 2000 epochs, zero archive overhead, pure distortion reduction
not done by any existing submission. Also adds LZMA-compressed latents + greedy Brotli tensor ordering in the codec
and a 31-mode FES1 frame selector (598/600 pairs active). Extends hnerv_muon's curriculum with 60% more Muon
training.
additional comments
Archive: 179,431 bytes. Training: ~34h on A100. Frame selector active on 598/600 pairs.