def new_sample(self, bytes, independent_sites):
# Posterior sample given previous peak/call counts (or zero
# counts, if it's the initial sample.)
peak_posterior = self.posterior()
# New peak/call counts
counts = self.zero_counts()
# New call counts
call_counts = scipy.zeros(10)
site_dists = []
total_observations = dict((s, 0) for s in '+-')
for byteidx, cbytes in enumerate(bytes):
# Get the posterior distributions given each observed peak
# at this site.
peaks, dists = zip(*[peak_posterior[s].call_dist(b,s)
for b, s in cbytes])
# Sample a posterior call
site_dist = self.gdist*scipy.product(dists, axis=0)
assert sum(site_dist)
site_dist = site_dist/sum(site_dist)
site_dists.append(site_dist)
call = SNPPrior.sample_call(site_dist)
strandcall = {'+': call, '-': dict_revcalls[call]}
# Count the peak/call pairs.
for peak, (byte, strand) in zip(peaks, cbytes):
# Undo the reverse complement, if this observation
# was on the reverse strand.
ccall = strandcall[strand]
ccounts = counts[strand]
# Flat peak/call count
ccounts[ccall][SNPPrior.peakidxs[peak.peaks]] += 1
# Call->(peak,score) count
ccounts['%s-%s' % (ccall, peak.peaks)][peak.bin]+=1
total_observations[strand] += 1
# Count of imputed calls
call_counts[callidxs[call]] += 1
# Blur the conditional distribution, to speed convergence.
for strand, ccounts in counts.items():
for val in ccounts.values():
val /= total_observations[strand]
val *= min(total_observations[strand], 100)
gdist = hyperprior.posterior(call_counts).sample()
return Sample(gdist, counts, site_dists)
def zero_counts(cls):
return dict((strand, SNPPrior.counts()) for strand in '+-')