Rwpics reads a set of PCASYS "matrix" format files each of size 14 x 15. The input files either should be a set of points in regional weights space (in particular, basepoints produced by the optimize regional weights command optrws), or they should be a set of estimated gradients of the activation error (also produced by optrws). Makes a corresponding set of IHead format image files depicting the weights points or gradients; these image files can then be displayed using dpyimage. Rwpics can be used as a sanity check on whether optrws is working properly, and to help decide when the time has come that it is reasonable to stop an optrws run because further significant change of the weights seems unlikely.
The pcasys "matrix" format file(s) to be depicted. Either they should each be a basepoint of the optrws optimization (point in regional weights space), or they should each be an estimated gradient, also produced by optrws. In either case, they must be matrices of size 14 x 15, corresponding to the weights, each of which is associated with one 2 x 2 vector block of the 28 x 30 pattern of orientation vectors. (Usually the output of optrws.)
A code telling rwpics what kind of pictures to make:
If rws ("regional weights"), the program makes a grayscale picture that is reasonable if the input file represents a set of regional weights, e.g. one of the bspt files produced by an optrws run. To do so, it linearly maps absolute values of input values to gray tones, setting the mapping so that 0. maps to black and the maximum absolute value across all components of all input matrices, to white. Absolute values are the reasonable thing to depict when examining a point in regional-weights space, since the sign of a regional weights has no effect on the PNN classifier. (Optimization may sometimes cause some unimportant outer weights to be slightly negative.) The mapping is adapted to the maximum absolute value across all input files, rather than being separately adapted for each input file; this is done so that the several resulting pictures can be examined side by side with the knowledge that all gray tones are on the same scale.
If eg ("estimated gradient"), the program makes a grayscale picture that is reasonable if the input file represents an estimated gradient of the error function, e.g. one of the egrad files produced by an optrws run. To do so, it affinely maps input values to gray tones, setting the mapping so that the minimum input value across all input files is mapped to white and the maximum input value, to black.
If seg ("sign of estimated gradient"), the program makes a grayscale-binary (ie. 0 and 255 pixel values) picture that is reasonable if the input file represents an estimated gradient of the error function. To do so, it maps negative values to white (255) and nonnegative values to black (0). This is interesting because if the component of the estimated gradient (i.e., the estimated partial derivative) associated with a region is negative, that shows that the weight for this region should be increased (and will be increased by optrws).
The directory in which the program should produce its output files, which will be raster images in the NIST IHead format. (The directory must already exist, i.e. rwpics does not produce it.) Each output file's name is produced by taking the last component of the corresponding input file and appending an underscore, then the rws|eg|seg code, then .pct (the standard IHEAD file suffix). The output image files may be examined using the dpyimage command.
From test/pcasys/execs/rwpics/rwpics.src:
% rwpics ../optrws/optrws.bin rws rwpics Produces an image of the optimized regional weights, which can be converted to JPEG format using the cjpegb command.
% rwpics ../optrws/optdir/egrad_1.bin eg rwpics Produces an image of the estimated gradient, which can be converted to JPEG format using the cjpegb command.
% rwpics ../optrws/optdir/egrad_1.bin seg rwpics Produces an image of the sign of estimated gradient, which can be converted to JPEG format using the cjpegb command.