submission name:

adaptive_range_mask

upload zipped archive.zip

archive.zip

report.txt

=== Evaluation config ===
  batch_size: 16
  device: cuda
  num_threads: 2
  prefetch_queue_depth: 4
  report: /root/comma_video_compression_challenge/submissions/adaptive_range_mask_no_router/report.txt
  seed: 1234
  submission_dir: /root/comma_video_compression_challenge/submissions/adaptive_range_mask_no_router
  uncompressed_dir: /root/comma_video_compression_challenge/videos
  video_names_file: /root/comma_video_compression_challenge/public_test_video_names.txt
=== Evaluation results over 600 samples ===
  Average PoseNet Distortion: 0.00049341
  Average SegNet Distortion: 0.00061248
  Submission file size: 215,735 bytes
  Original uncompressed size: 37,545,489 bytes
  Compression Rate: 0.00574596
  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 0.28

does your submission require gpu for evaluation (inflation)?

yes

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

no

additional comments

This needs a C compiler in the runtime.

I first tried to retrain from #55, and it took wayy too many GPU credits. So I tried optimizing from there. I saw another PR which uses the same strat of compressing the segnet mask. I think both our LLM's landed at the same conclusion hahah. At least right now, this one is marginally better. There is still lots of room for optimizing further from just this model.

For retraining, I used the same pipeline as #62 and #55, and got very similar results without breaking through the 0.33 mark so I just decided to continue from theirs. But with enough GPU time and testing different pipelines, I bet you could get a model that can predict segnet/pose with a smaller compression.

Exact local score: 0.2751402303839512