DATA-SPEC:  Specify data sets for training and testing.

Data-spec writes specifications of data sets to use for training and
testing to the given log file (which must already exist).  If invoked
with just a log file as argument, the program prints the data
specifications already stored in the log file.


    data-spec log-file N-inputs N-targets [ int-target ]
          / train-inputs train-targets [ test-inputs [ test-targets ] ]
        [ / { input-trans } [ / { target-trans } ] ]


    data-spec log-file

The number of "input" and "target" values must be specified.  The
number of inputs may be zero (though other programs may not allow
this), but the number of targets must be at least one.  The user can
require the targets to be integers by including an 'int-target'
argument.  The permitted range of these integer targets is from zero
to 'int-target' minus one.

The training and (optionally) test sets are given in the form of
specifications described in numin.doc.  The default specification for
training and test inputs is "data@1,1".  The training input
specification provides the default for the training target
specification - i.e. the file and line range are used for the inputs
is used for targets if not overridden, and the item indexes for the
targets start by default after the last index for the inputs.
Similarly, the test input specification is used as the default for the
test targets.

The data files are read and checked for errors.  The number of lines
of training inputs must match the number of lines of training targets,
and similarly for test inputs and targets.

The final optional arguments specify the transformations to be applied
to inputs and targets.  A transformation specification has one of the

               [L][+T][xS]  or  [L][-T][xS]  or  I

If "L" is present in the two forms, the raw data is first transformed
by taking the logarithm.  T is then added or subtracted, and the
result multiplied by S.  The "I" form specifies an identity
transformation.  This is the default for any inputs or targets for
which no transformation is specified.

The translation and scaling amounts ("T" and "S") may be "@", in which
case they are computed from the training data, so as to produce a mean
of zero (for translation), or a variance of one (for scaling).

The specifications are written to the log file as a record with type
'D' and index -1.

            Copyright (c) 1995 by Radford M. Neal