How to monitor convergence of Gensim LDA model?

Question

I can't seem to find it or probably my knowledge on statistics and its terms are the problem here but I want to achieve something similar to the graph found on the bottom page of the LDA lib from PyPI and observe the uniformity/convergence of the lines. How can I achieve this with Gensim LDA?

Answer 1

You are right to wish to plot the convergence of your model fitting. Gensim unfortunately does not seem to make this very straight forward.

1. Run the model in such a way that you will be able to analyze the output of the model fitting function. I like to setup a log file.

   import logging
   logging.basicConfig(filename='gensim.log',
                       format="%(asctime)s:%(levelname)s:%(message)s",
                       level=logging.INFO)
   

2. Set the eval_every parameter in LdaModel. The lower this value is the better resolution your plot will have. However, computing the perplexity can slow down your fit a lot!

   lda_model = 
   LdaModel(corpus=corpus,
            id2word=id2word,
            num_topics=30,
            eval_every=10,
            pass=40,
            iterations=5000)
   

3. Parse the log file and make your plot.

   import re
   import matplotlib.pyplot as plt
   p = re.compile("(-*\d+\.\d+) per-word .* (\d+\.\d+) perplexity")
   matches = [p.findall(l) for l in open('gensim.log')]
   matches = [m for m in matches if len(m) > 0]
   tuples = [t[0] for t in matches]
   perplexity = [float(t[1]) for t in tuples]
   liklihood = [float(t[0]) for t in tuples]
   iter = list(range(0,len(tuples)*10,10))
   plt.plot(iter,liklihood,c="black")
   plt.ylabel("log liklihood")
   plt.xlabel("iteration")
   plt.title("Topic Model Convergence")
   plt.grid()
   plt.savefig("convergence_liklihood.pdf")
   plt.close()