v1.3.0

Added

• qc model report (qc model report corpus.jsonl --model model.json --to model.html): an HTML page showing what a ladder model is worth and
  what its corpus holds, of either alone or of both. Findings tell whether
  the model predicts without a probe and without any encode and how far
  from a measured ladder, a model older than its corpus, unfinished
  titles, and a grid that stops before the low rungs of its titles. Then
  the error of each design, the error of each resolution and CRF of the
  grid, the titles the model reads worst (the kinds of content to add),
  the VMAF and the bitrate of the corpus as two surfaces in perspective
  (drag to turn them, point at a corner, pick a title or play them all
  against the mean), the mean rate-quality curve of each resolution, and
  where the titles sit on each feature. qc model train now keeps the
  validation error of each title in the model
  (predict.Validation.Titles).
• Ladder models (qc model corpus, qc model train, qc ladder --ladder-model): a model trained on titles measured on a whole grid of
  resolutions and CRFs completes the probes of a ladder. A build then
  measures 4 to 10 encodes of the digest (the anchors) instead of 15 to 30
  and predicts the others; the level correction, the rung selection and
  the verification of every rung are unchanged. On a 59-minute title the
  model had never seen (x264 fast, a model of two other titles), probing
  went from 3 min 36 s to 52 s with the 10 anchors of the default tolerance
  and to 37 s with 6 (--tolerance 0.8), and the whole build from
  7 min 16 s to 3 min 28 s and 3 min 02 s, for the same rung resolutions
  (but the last rung with 6) and verifications within 1.2 and 0.4 VMAF of
  the predictions; with 4 anchors (--tolerance 1), without analysis nor
  verification (--digest uniform --no-verify), 52 s instead of
  2 min 57 s. Designs of anchors are fitted on digests of a few segments,
  as the anchors of a build are measured: on a corpus of 39 titles the
  error of four anchors went from 0.62 to 0.47 VMAF.
  qc model train validates every set of anchors by leaving each title of
  the corpus out in turn and stores the errors in the model; --tolerance
  picks the fewest anchors whose mean error is within it. A model is for
  one codec and preset and one source height; a build it does not cover
  falls back on fixed probing and says why. Library: packages
  ladder/corpus and ladder/predict, ladder.ProbingModel,
  ladder.Options.Curves, Probe.Predicted, Result.ProbeEncodes.
• Instant ladders (qc ladder --ladder-model model.json --instant): a
  ladder predicted from the analysis of the video, without any probe nor
  verification, in about 4 s for titles of 10 minutes to 1 hour. qc model corpus now also describes every segment (spatial and temporal
  information, luma, camera motion, what x264 ultrafast spends on it at
  360p and the VMAF of those two encodes, what scaling loses at each
  resolution), and qc model train fits and validates a design that
  reads those features and no anchor. Its accuracy is that of its corpus,
  and is reported with the ladder: with three titles, 1.2 VMAF and 8% of
  the bitrate on average on a title never seen; verified, 4 rungs of 6
  and 5 of 8 were right as predicted on two titles, and the two top rungs
  of 8 on the third (CRFs 1.5 to 2 steps high below them). The model
  predicts every segment of the digest and averages them: from the mean
  features of the digest, bitrates came out 13% low on average on a
  corpus of 39 titles (7% of error now, 14% then).
  --probing predicted is the same prediction with the digest and the
  verification of any build. Library: package
  ladder/features, ladder.ProbingPredicted, ladder.WithFeatures,
  encode.FFmpeg.Stats and ScaleLoss. qc model corpus takes folders,
  for every video under them.
• --plain (every command, or QC_PLAIN=1): the progress as plain
  lines on stderr, one per event, in place of the live dashboard and of the
  wizard: for containers, CI jobs and logs. Without it and without a
  terminal, qc is silent until its report, as before.
• --quiet (every command, or QC_QUIET=1): no progress at all, in a
  terminal either: with -f json, nothing but the JSON.
• qc model corpus --crf-count: the number of CRFs a corpus is
  measured at, spread evenly over the codec's range (default: 11 for
  H.264, as before).
• --tolerance (ladder, run): the VMAF error the curves may have
  between probes, for adaptive probing and for ladder models (it was a
  library option only).
• --no-encode (ladder, run, with --instant or --probing predicted): the digest is described without encoding anything, from
  its frames and from the source's own encode, leaving out the two very
  fast encodes a prediction otherwise scores. About 2 s instead of 4 for
  an hour-long title, for about twice the error: 2.2 VMAF and 9.3% of the
  bitrate on 40 titles never seen, against 1.2 and 6.4%. It never falls
  back on probes: without a model that can, the build fails and says why
  (ladder.ErrNoEncode). A model now
  holds a design for each thing a build can read (qc model train
  lists them); one trained before must be trained again for this mode.
  Library: ladder.Options.NoEncode, features.Sets, features.Usable,
  predict.TrainOptions.FeatureSets.

Changed

• Instant ladders are faster (12 s → 4 s for an hour-long title) and
  read the source's own encode. Their digest is 20 s long, ten
  segments, unless --digest-duration says another (twenty segments:
  7.5 s, for about a tenth less error: 1.0 against 1.1 VMAF and 8.8%
  against 9.8% of bitrate on four titles). The
  digest is cut from keyframes, a process per segment (nothing is decoded
  before a segment; encode.DigestSpec.Fast); the two fast encodes are
  scored on three frames spread evenly in each segment instead of three
  clips of four drawn in it (half the frames, and no longer a draw;
  quality.Sample.Even);
  and the mean quantiser of the source, read while the digest is cut
  (8-bit H.264, encode.DigestSpec.Quantisers), joins the features: 7.5%
  → 6.4% of bitrate error on 40 titles. The scale loss, which added
  nothing, is no longer described (models trained with it still get it).
  Corpora are described again by qc model corpus (features version 7,
  the encodes of the grid are kept) and models trained again to read the
  quantiser. predict.Model.Design and Instant take what tells the
  features a build can read.
• An instant ladder never encodes. It measured challenger probes
  where its rungs needed CRFs past the grid of its model (six encodes and
  44 s on an hour-long title in AV1, whose low rungs need CRF 60 to 62
  when the grid stopped at 58); it now places those rungs on what the
  model knows. The corpus grid of AV1 ends on CRF 63, the highest of
  SVT-AV1, whose scale falls too fast at its end to be read from the
  points before it: run qc model corpus again on an AV1 corpus (five
  encodes a title) and train again.
• A ladder no longer inspects its source twice (qc ladder, qc run):
  1.5 s less for an hour of video. Library:
  ladder.Engine.PrepareInspected, analysis.Options.SkipBitstream.

Docker

docker run --rm -v "$PWD:/data" ghcr.io/eko/qc:1.3.0 version --check

Binaries

Self-contained (libvmaf and the VMAF models are built in); ffmpeg is still
needed. Linux binaries are static; the macOS one is not signed
(install).