onyx.am.testhmm2 – Various test drivers for Hmm class; these use the log domain

>>> True
True

onyx.am.testhmm2.logreftest()

onyx.am.testhmm2.make_gmm(dimension, num_mixtures)

onyx.am.testhmm2.make_gmm_diag(dimension, num_mixtures)

onyx.am.testhmm2.obs_generator(generators, target_durations)

onyx.am.testhmm2.test0(num_obs, num_passes)

onyx.am.testhmm2.test1(num_obs, num_passes)

onyx.am.testhmm2.test2(num_obs, num_passes)

onyx.am.testhmm2.test2a(num_obs, num_passes)

onyx.am.testhmm2.test3()

onyx.am.testhmm2.test3_helper(dataset_idx, num_passes)

This tests mimics a run ChrisW did with HTK. The models are 2-D single-mode Gaussians embedded in a 1-state Hmm. Each data point is taken as a sequence of length 1.

onyx.am.testhmm2.test4()

onyx.am.testhmm2.test5(num_obs)

onyx.am.testhmm2.test5_helper(num_obs, num_passes)

This tests mimics a run ChrisW did with HTK. The models are 2-D single-mode Gaussians embedded in a 3-state Hmm. Each observation is a sequence of length 11, taken by sampling 2, 3, and 6 times, respectively, from three target distributions.

onyx.am.testhmm2.test6()

This test builds an Hmm with dummy models which always give a score of 1, but with a somewhat unusual topology in which there are 6 actual states chained together with 2 virtual inputs and 3 virtual outputs. The point is to make sure we can handle this asymetric case correctly.

onyx.am.testhmm2.test7()

This test builds an Hmm with dummy models which always give a score of 1, but with a somewhat unusual topology in which there are 6 actual states chained together with 2 virtual inputs and 3 virtual outputs. The point is to make sure we can handle this asymetric case correctly. This is the same as test6 except that now we’ll use the network adaptation interface instead.

onyx.am.testhmm2.test8(num_obs)

onyx.am.testhmm2.test8_helper(num_obs, num_passes)

This tests mimics a run ChrisW did with HTK. The models are 2-D single-mode Gaussians embedded in a 3-state Hmm. Each observation is a sequence of length 11, taken by sampling 2, 3, and 6 times, respectively, from three target distributions. This is identical to test5 except that here I have built the Hmm with only one Gmm, which is shared by all three states.