PathMultiLabelGenerator: interface to infer the label(s) of a file directly from the URI/path
Similar to
PathLabelGenerator, with 2 main differences:
(a) Can be used for multi-label, multi-class classification (i.e., return *multiple* NDArray writables, for use in
networks with multiple output layers)
(b) Does
not support inferring label classes
Regarding (b) above, this means that the implementations of PathMultiLabelGenerator typically need to (for classification
use cases) do one of two things (either will work, though down-stream usage of these arrays can vary slightly):
(a) Perform label to integer index assignment (i.e., return an IntWritable(0) for A, if you have 3 classes {A,B,C})
(b) Create a one-hot NDArrayWritable. For 3 classes {A,B,C} you should return a [1,0,0], [0,1,0] or [0,0,1] NDArrayWritable
Comparatively, PathLabelGenerator can return a Text writable with the label (i.e., "class_3" or "cat") for classification.
More generally, PathMultiLabelGenerator must return Writables of one of the following types:
DoubleWritable,
FloatWritable,
IntWritable,
LongWritable or
NDArrayWritable.
NDArrayWritable is used for classification (via one-hot NDArrayWritable) or multi-output regression (where all values
are grouped together into a single array/writable) - whereas the others (double/float/int/long writables) are
typically used for single output regression cases, or (IntWritable) for classification where downstream classes (notably
DL4J's RecordReader(Multi)DataSetIterator) will convert the integer index (IntWritable) to a one-hot array ready for
training.
In principle, you can also return time series (3d - shape [1,size,seqLength]) or images (4d - shape
[1,channels,height,width]) as a "label" for a given input image.