Incorrect documentation for APMeter

[arunmallya]

https://github.com/pytorch/tnt/blob/master/torchnet/meter/apmeter.py#L36-L37 states The probabilities should sum to one over all classes, which is not a requirement. If multiple classes are present in an image, we would want our method to produce high confidences for each and each confidence is in the [0, 1] range. I don't think normalizing scores across classes would make much sense.

The rest of the implementation doesn't seem to be making use of this requirement.

[erobic]

The documentation is misleading, but APMeter is actually good for multi-label classification right?

It says:

when the model is more convinced that the example should be positively labeled, and smaller when the model believes the example should be negatively labeled (for instance, the output of a sigmoid function)