def wl_is_faster_hamming():
# show hamming error to true state sequence decreases faster with wl
### get samples
wl_samples = get_hdphsmm_wl_poisson_samples(hsmm_data, nruns=100, niter=300, L=10)
da_samples = get_hdphsmm_da_poisson_samples(hsmm_data, nruns=24, niter=150)
### get hamming errors for samples
def f(tup):
return util.stateseq_hamming_error(tup[0], tup[1])
wl_errs = np.array(dv.map_sync(f, zip(wl_samples, [hsmm_labels] * len(wl_samples))))
da_errs = np.array(dv.map_sync(f, zip(da_samples, [hsmm_labels] * len(da_samples))))
### plot
plt.figure()
for errs, samplername, color in zip([wl_errs, da_errs], ["Weak Limit", "Direct Assignment"], ["b", "g"]):
plt.plot(np.median(errs, axis=0), color + "-", label="%s Sampler" % samplername)
plt.plot(util.scoreatpercentile(errs.copy(), per=25, axis=0), color + "--")
plt.plot(util.scoreatpercentile(errs.copy(), per=75, axis=0), color + "--")
plt.legend()
plt.xlabel("iteration")
plt.ylabel("Hamming error")
save("figures/wl_is_faster_hamming.pdf")
return wl_errs, da_errs
def hsmm_vs_stickyhmm():
# show convergence rates in #iter are same
### get samples
hsmm_samples = get_hdphsmm_da_geo_samples(hmm_data, nruns=50, niter=100)
shmm_samples = get_shdphmm_da_samples(hmm_data, nruns=50, niter=100)
### get hamming errors for samples
def f(tup):
return util.stateseq_hamming_error(tup[0], tup[1])
hsmm_errs = np.array(dv.map_sync(f, zip(hsmm_samples, [hmm_labels] * len(hsmm_samples))))
shmm_errs = np.array(dv.map_sync(f, zip(shmm_samples, [hmm_labels] * len(shmm_samples))))
### plot
plt.figure()
for errs, samplername, color in zip([hsmm_errs, shmm_errs], ["Geo-HDP-HSMM DA", "Sticky-HDP-HMM DA"], ["b", "g"]):
plt.plot(np.median(errs, axis=0), color + "-", label="%s Sampler" % samplername)
plt.plot(util.scoreatpercentile(errs.copy(), per=25, axis=0), color + "--")
plt.plot(util.scoreatpercentile(errs.copy(), per=75, axis=0), color + "--")
plt.legend()
plt.xlabel("iteration")
plt.ylabel("Hamming error")
save("figures/hsmm_vs_stickyhmm.pdf")
return hsmm_errs, shmm_errs
