usage-trainSpacing

scannedb-ok trainSpacing - Train spacing module.

Usage: scannedb-ok trainSpacing ([-p|--pdf] | [-x|--xml])
                                [-l|--lines-per-page LINES]
                                [--threshold THRESHOLD]
                                [-k|--keep-single-glyphs-lines]
                                [--steps-per-line STEPS] [--iterations ARG]
                                [--rate ARG] [--momentum ARG] [--l2 ARG]
                                TRAININGPDF TRAININGTEXT VALIDATIONPDF
                                VALIDATIONTEXT NETFILE
  scannedb-ok trainSpacing trains an ANN for inter-glyph spacing. This requires
  four input files: TRAININGPDF and TRAININGTXT must contain exactly the same
  text, once in PDF format (or some other format conaining glyph descriptions,
  like pdfminers xml), and once as plaintext with correct spaces. This pair of
  files is used for training the neurons. VALIDATIONPDF and VALIDATIONTXT have
  to be exactly corresponding texts, too. They are used to validate the trained
  network after each learning iteration and the validation result is logged, so
  that the progression can be observed. For a start, you can simply reuse the
  first pair of files as validation texts.

Available options:
  -h,--help                Show this help text
  -p,--pdf                 PDF input data. (Default)
  -x,--xml                 XML input data. An XML representation of the glyphs
                           of a PDF file, like produced with PDFMiner's
                           "pdf2txt.py -t xml ..." command.
  -l,--lines-per-page LINES
                           Lines per page. Lines of a vertically filled page.
                           This does not need to be exact. (default: 42)
  --threshold THRESHOLD    A threshold value important for the identification of
                           lines by the internal clustering algorithm:
                           Fail-OCRed glyphs between the lines may disturb the
                           separation of lines. Instead of 0, the threshold
                           value of is used to separate the clusters. If set to
                           high, short lines may be dropped of. (default: 2)
  -k,--keep-single-glyphs-lines
                           Do not drop glyphs found between the lines. By
                           default, lines with a count of glyphs under THRESHOLD
                           are dropped.
  --steps-per-line STEPS   With the STEPS per line you may tweak the clustering
                           algorithm for the indentification of
                           lines. (default: 5)
  --iterations ARG         number of learning iterations (default: 100)
  --rate ARG               Learning rate (default: 1.0e-2)
  --momentum ARG           Learning momentum (default: 0.9)
  --l2 ARG                 Learning regulizer (default: 5.0e-4)
  TRAININGPDF              Path to PDF (or XML) with training text.
  TRAININGTEXT             Path to plaintext with correct spaces which
                           corresponds exactly to TRAININGPDF.
  VALIDATIONPDF            Path to PDF (or XML) with validation text.
  VALIDATIONTEXT           Path to plaintext with correct spaces which
                           corresponds exactly to VALIDATIONPDF.
  NETFILE                  Path to the file where to dump the trained network
                           in. The file will be in a binary format, so ".dat" is
                           a reasonable suffix.