aakhundov / tf-example-models Goto Github PK

Hi,
Which article does the code implementation of guassian mixture model refer to? I mean the process of evaluating the maximum likelihood? There are two items of likelihood, i.e., tf.reduce_sum(tf.log(exp_log_shifted_sum)) and tf.reduce_sum(log_shift). From the results I run the code, the first item (i.e., tf.reduce_sum(tf.log(exp_log_shifted_sum))) always give a nearly zero value. And the second item (i.e., tf.reduce_sum(log_shift)) is really useful. I'm not sure that I obtain the right results. I also can not understand the implementation of this part. So would you tell me which article of the method refer to? Or would you explain the process briefly?

Hi there,

I am confused about the part of applying prior to the computed variances. Would you by any chance explain it? Thanks ;)
Link

# applying prior to the computed variances
variances_ *= tf.expand_dims(gamma_sum, 1)
variances_ += (2.0 * beta)
variances_ /= tf.expand_dims(gamma_sum + (2.0 * (alpha + 1.0)), 1)

Also, does the log_shit means the normalization here?

log_weighted = log_components + tf.expand_dims(tf.log(weights), 1)
log_shift = tf.expand_dims(tf.reduce_max(log_weighted, 0), 0)
exp_log_shifted = tf.exp(log_weighted - log_shift)

In addition, at line 82:

log_likelihood = tf.reduce_sum(tf.log(exp_log_shifted_sum)) + tf.reduce_sum(log_shift)

Why added above mentioned maximum value of each component log_shift?

How do you actually get a result shown in gmm-best.png. I mean, among several algorithms, which one did you use to get the best result. By just running your codes (e.g. tf_gmm_diag.py, tf_gmm_full.py, tf_gmm_grad, tf_gmm.py), they mostly give incomplete result as below.

Is gmm sensitive to initial mean assignment?