diffu-htr-speed / app.py

Commit History

8 days ago

  1. Jev in one fold titled with its verdict (picked the reading with fewer errors on K of M lines); each reading's errors and Jev's certainty in percent

    Verified signature
    dceb361
  2. Jev's section in plain words: it picked the better reading on K of the M lines where one was better; rows say better, worse, tie or same reading; bars: Jev's picks and perfect picks

    Verified signature
    5041a61
  3. Examples drawn at random from test pages (TrOCR trained on our dev volumes), timed again; Jev's section: the region's verdict, CER bars and one row per line

    Verified signature
    17d9643
  4. A compact page: the region box and Read beside a three-column examples table, the live bars beside their summary, our server's timings beside the batched rates, each of Jev's lines beside its readings

    Verified signature
    7ca716d
  5. One stream each: TrOCR reads its lines one by one (its one-batch time beside it), the replay starts all three together, our server's vLLM timings under the race, one region box, Jev's rows folded with marked errors

    Verified signature
    b0dacaa
  6. The app starts without the private TrOCR weights (its lane says so) instead of stopping at startup

    Verified signature
    7a31b73
  7. Jev's limits said plainly (AUROC 0.988 against rule-made errors, 0.624 against the reader's own); published from diffu2 by publish_space.sh

    Verified signature
    005f43a
  8. The race as a custom component: each reader's reading replayed at its measured pace on the manuscript (a highlighter along the lines' outlines) and as text, a block per reader pass, lines linked on hover; results streamed as each reader finishes; Jev scores both readers' reading of each line; a 4,096-token cache; no fla

    Verified signature
    8583929
  9. app.py: the startup warm-up catches CUDA and ZeroGPU errors only

    Verified signature
    4ed56b9
  10. app.py: the TrOCR line within 120 characters

    Verified signature
    c05415f
  11. Swedish HTR speed and Jev: TrOCR vs MiniCPM-OCR alone and with its DSpark draft (PyTorch on CUDA graphs), Jev per line

    Verified signature
    f2884d6