Layouts
Our system implements 5 source layout types:
Layout Type 0
<src_filename> specifies a large file that contains all the input field values as rows of data.
There is no row end marker and the row dimensions are determined by the input node count specification in the model.
The last set of columns represents the 1-of-k label value as bitmap, in 64 bit bitmaps or n-of-k values as a bitmap.
The last set of columns may also represent a single value or a range of values or multiple ranges or values, each stored with the same precision as the data.
This layout does not need any label file
This layout is the default input layout for our data processing implementation.
For all other data layouts, we offer a complimentary and editable Java module that can convert your other layout types to the one listed above. The source code uses a standardized library and may be edited freely by you to suit your layout.
Layout Type 1
This input layout will contain two files: one containing rows of input data.
The second file for labels that contains the label values, where the label data represents the 1-of-k label value as a bitmap or n-of-k values as a bitmap.
The label data can also be represented as a single value, or a range of values, or multiple ranges of values, each stored with the same precision as the data.
The label file must contain a column <record_no> that matches the label to the corresponding index of the <source_file>.
This column is used ti match the rows in the two files.
This program will process the input file to convert it to <destination_data_precision> and will process the label file according to the destination specifications, and produce an output of layout type 0.
Layout Type 2
This layout requires that all the images an additional text file for the labels are stored under <src_dir>.
The text file for labels should contain the label values, where the label data represents the 1-of-k label value as a bitmap or n-of-k values as a bitmap.
The label values can also be represented as a single value, or a range of values or multiple ranges of values, each stored with the same precision as the data.
The code as offered handles jpeg files. However the code can also be modified to handle non-image data files, or images of other storage formats.
Layout Type 3
This layout requires that all the images are stored under <src_dir>. Each label is a sub-folder, and the sub-folder in turn contains the image files.
The label value is implied as a single-value classification through the sub-folder name string. The program converts the strings into sequential index value from 1… <count-of-distinct-sub-folders> and saves the value as the label column.
Layout Type 4
Similar to layout type 3, however, each sub-folder contains a label file, that contains the following record structure: <file_name>, <label_value>
The labels in this case may be value, range, or 1-of-k or n-of-k bitmaps or multi-valued ranges, as desired.
The sub-folder name is not used for any label-related purposes.
Label Structures
The field <label_type> determines the single-valued or multi-valued classification.
The count of total sets must be transferred to the model specifications.
The field <label_form> determines the label format.
Implemented types are:
“bitmap” represents 1-of-k or n-of-k
“range” for range values
“value” represents a single-valued label. This will default to the index number of the sub-folder, or the file. It can also be read from a file for non-sequential values.
