Hi, In CVB0Driver.java, I notice the change rate of perplexity is calculated as follows:
private static double rateOfChange(List<Double> perplexities) {
...
return Math.abs(perplexities.get(sz - 1) - perplexities.get(sz -
2)) / perplexities.get(0);
}
Why do we always use the first perplexity as the denominator?
shouldn't it be better if we use perplexities.get(sz - 2), which will
make us easier to choose the value of convergence delta.
Thanks,
Bocong Liu
