hdplda.py:sampling_k

[ttjs219]

Hello,master:
Thanks for you code which give me more thought. I'am a new scholar in learning hdp-model.when i understand the code in your iir/lda/hdplda.py ,the row of 252 and 253 makes me puzzle.when we sample k,what is the mean of 'log_p_k[i]' and 'log_p_k[K]'.
Otherwise,the code 'k_new = self.sampling_topic(numpy.exp(log_p_k - log_p_k.max()))',should i change 'log_p_k.max()' to 'log_p_k.sum()'.can you tell me the program process or formula about that.thanks!

[shuyo]

Subtracting the maximum is to suppress overflow/underflow.
It gives the same result if the maximum is replaced with any value as long as no overflow/underflow.
Check an implementation of softmax function if you want to know more detail.
Thanks.