def bp_update_params_new(args):
lmds, pis = args.lmds, args.pis
vert_parent, vert_children = args.vert_parent, args.vert_children
#print pis[0].shape
I, T, L = args.X.shape
K = args.gamma.shape[0]
theta, alpha, beta, gamma, emit_probs, X = (args.theta, args.alpha, args.beta, args.gamma, args.emit_probs,
args.X)
evidence = args.evidence #evid_allchild(lmds, vert_children)
emit_probs_mat = sp.exp(args.log_obs_mat)
#emit_probs_mat /= emit_probs_mat.max(axis=2).reshape(I, T, 1)
gamma_p = copy.copy(gamma)
alpha_p = copy.copy(alpha)
beta_p = copy.copy(beta)
theta_p = copy.copy(theta)
emit_probs_p = copy.copy(emit_probs)
theta[:] = args.pseudocount
alpha[:] = args.pseudocount
beta[:] = args.pseudocount
gamma[:] = args.pseudocount
emit_probs[:] = args.pseudocount
#evidence = evid_allchild(lmds, args.vert_children)
##support = casual_support(pis)
emit_sum = sp.zeros(K)
for i in xrange(I):
vp = vert_parent[i]
for t in xrange(T):
if i==0 and t==0:
gamma += emit_probs_mat[i, t, :]*evidence[i,t,:]
Q = emit_probs_mat[i, t, :]*evidence[i,t,:]
else:
if i == 0:
tmp1, tmp2 = sp.ix_(pis[i][-1,t-1,:], emit_probs_mat[i, t, :]*evidence[i,t,:])
tmp = alpha_p * (tmp1*tmp2)
#tmp /= tmp.sum()
Q = tmp.sum(axis=0)
alpha += tmp/tmp.sum()
elif t == 0:
tmp1, tmp2 = sp.ix_(pis[vp][i-1,t,:], emit_probs_mat[i, t, :]*evidence[i,t,:])
tmp = beta_p *(tmp1*tmp2)
Q = tmp.sum(axis=0)
beta += tmp/tmp.sum()
else:
tmp1, tmp2, tmp3 = sp.ix_(pis[vp][i-1,t,:], pis[i][-1, t-1,:], emit_probs_mat[i, t, :]*evidence[i,t,:])
tmp = theta_p *(tmp1*tmp2 *tmp3)
Q = (tmp.sum(axis=0)).sum(axis=0)
theta += tmp/tmp.sum()
Q /= Q.sum()
for l in xrange(L):
if X[i,t,l]:
emit_probs[:, l] += Q
emit_sum += Q
normalize_trans(theta, alpha, beta, gamma)
emit_probs[:] = sp.dot(sp.diag(1./emit_sum), emit_probs)
args.emit_sum = emit_sum
make_log_obs_matrix(args)
def params_from_clique_marginals(theta, alpha, beta, gamma, emit_probs, clq_Q,
clq_Q_pairs, X, args):
"""Recompute parameters using marginal probabilities"""
I, T, L = X.shape
K = gamma.shape[0]
vp = 0
pc = 1e-10 # pseudocount
theta[:] = pc
alpha[:] = pc
beta[:] = pc
gamma[:] = pc
emit_probs[:] = pc
#theta[:] = 0; alpha[:] = 0; beta[:] = 0; gamma[:] = 0; emit_probs[:] = 0
global combinations
combinations = list(enumerate(itertools.product(range(K), repeat=I)))
#e_sum = sp.ones(emit_probs.shape[0]) * pc
e_sum = sp.ones_like(emit_probs) * pc
#e_sum = sp.ones(emit_probs.shape[1]) * 0
t = 0
for k_to, to_val in combinations:
for i in xrange(I):
for l in xrange(L):
if X[i, t, l]:
emit_probs[to_val[i], l] += clq_Q[t, k_to]
#e_sum[to_val[i]] += clq_Q[t, k_to]
e_sum[to_val[i], l] += clq_Q[t, k_to]
gamma[to_val[vp]] += clq_Q[t, k_to]
for i in xrange(1, I):
beta[to_val[vp], to_val[i]] += clq_Q[t, k_to]
#import ipdb; ipdb.set_trace()
for t in xrange(1, T):
for k_to, to_val in combinations:
for i in xrange(I):
for l in xrange(L):
if X[i, t, l]:
#import ipdb; ipdb.set_trace()
emit_probs[to_val[i], l] += clq_Q[t, k_to]
#e_sum[to_val[i]] += clq_Q[t, k_to]
e_sum[to_val[i], l] += clq_Q[t, k_to]
for k_from, from_val in combinations:
alpha[from_val[vp], to_val[vp]] += clq_Q_pairs[t, k_from, k_to]
for i in xrange(1, I):
theta[to_val[vp], from_val[i],
to_val[i]] += clq_Q_pairs[t, k_from, k_to]
#import ipdb; ipdb.set_trace()
#theta, alpha, beta, gamma = theta+pc*theta.max(), alpha+pc*alpha.max(),beta+pc*beta.max(),gamma+pc*gamma.max()
normalize_trans(theta, alpha, beta, gamma)
#emit_probs[:] = sp.dot(sp.diag(1./e_sum), emit_probs)
#emit_probs[:] = sp.dot(emit_probs, sp.diag(1./e_sum * L))
emit_probs[:] = emit_probs / e_sum
#emit_probs[:] = sp.dot(emit_probs + pc*emit_probs.max(), sp.diag(1./(e_sum + pc*emit_probs.max())))
args.emit_sum = e_sum
def params_from_clique_marginals(theta, alpha, beta, gamma, emit_probs, clq_Q, clq_Q_pairs, X, args):
"""Recompute parameters using marginal probabilities"""
I,T,L = X.shape
K = gamma.shape[0]
vp = 0
pc = 1e-10 # pseudocount
theta[:] = pc; alpha[:] = pc; beta[:] = pc; gamma[:] = pc; emit_probs[:] = pc
#theta[:] = 0; alpha[:] = 0; beta[:] = 0; gamma[:] = 0; emit_probs[:] = 0
global combinations
combinations = list(enumerate(itertools.product(range(K), repeat=I)))
#e_sum = sp.ones(emit_probs.shape[0]) * pc
e_sum = sp.ones_like(emit_probs) * pc
#e_sum = sp.ones(emit_probs.shape[1]) * 0
t = 0
for k_to, to_val in combinations:
for i in xrange(I):
for l in xrange(L):
if X[i,t,l]:
emit_probs[to_val[i], l] += clq_Q[t, k_to]
#e_sum[to_val[i]] += clq_Q[t, k_to]
e_sum[to_val[i], l] += clq_Q[t, k_to]
gamma[to_val[vp]] += clq_Q[t,k_to]
for i in xrange(1, I):
beta[to_val[vp], to_val[i]] += clq_Q[t,k_to]
#import ipdb; ipdb.set_trace()
for t in xrange(1, T):
for k_to, to_val in combinations:
for i in xrange(I):
for l in xrange(L):
if X[i,t,l]:
#import ipdb; ipdb.set_trace()
emit_probs[to_val[i], l] += clq_Q[t, k_to]
#e_sum[to_val[i]] += clq_Q[t, k_to]
e_sum[to_val[i], l] += clq_Q[t, k_to]
for k_from, from_val in combinations:
alpha[from_val[vp], to_val[vp]] += clq_Q_pairs[t,k_from, k_to]
for i in xrange(1, I):
theta[to_val[vp], from_val[i], to_val[i]] += clq_Q_pairs[t,k_from, k_to]
#import ipdb; ipdb.set_trace()
#theta, alpha, beta, gamma = theta+pc*theta.max(), alpha+pc*alpha.max(),beta+pc*beta.max(),gamma+pc*gamma.max()
normalize_trans(theta, alpha, beta, gamma)
#emit_probs[:] = sp.dot(sp.diag(1./e_sum), emit_probs)
#emit_probs[:] = sp.dot(emit_probs, sp.diag(1./e_sum * L))
emit_probs[:] = emit_probs / e_sum
#emit_probs[:] = sp.dot(emit_probs + pc*emit_probs.max(), sp.diag(1./(e_sum + pc*emit_probs.max())))
args.emit_sum = e_sum
def do_parallel_inference(args):
"""Perform inference in parallel on several observations matrices with
joint parameters
"""
from histone_tree_hmm import random_params, do_inference, plot_params, plot_energy, load_params
from vb_mf import normalize_trans
args.I, _, args.L = sp.load(args.observe_matrix[0]).shape
K = args.K
L = args.L
args.T = 'all'
args.free_energy = []
args.observe = 'all.npy'
args.last_free_energy = 0
args.emit_sum = 0
args.out_dir = args.out_dir.format(timestamp=time.strftime('%x_%X').replace('/','-'), **args.__dict__)
try:
print 'making', args.out_dir
os.makedirs(args.out_dir)
except OSError:
pass
if args.warm_start:
#args.last_free_energy, args.theta, args.alpha, args.beta, args.gamma, args.emit_probs, args.emit_sum = load_params(args)
#args.warm_start = False
print '# loading previous params for warm start from %s' % args.warm_start
tmpargs = copy.deepcopy(args)
tmpargs.out_dir = args.warm_start
tmpargs.observe = 'all.npy'
args.free_energy, args.theta, args.alpha, args.beta, args.gamma, args.emit_probs, args.emit_sum = load_params(tmpargs)
try:
args.free_energy = list(args.free_energy)
except TypeError: # no previous free energy
args.free_energy = []
print 'done'
args.warm_start = False
else:
(args.theta, args.alpha, args.beta, args.gamma, args.emit_probs) = \
random_params(args.K,args.L)
for p in ['free_energy', 'theta', 'alpha', 'beta', 'gamma', 'emit_probs', 'last_free_energy', 'emit_sum']:
sp.save(os.path.join(args.out_dir, args.out_params.format(param=p, **args.__dict__)),
args.__dict__[p])
print '# setting up job arguments'
# set up new versions of args for other jobs
job_args = [copy.copy(args) for i in range(len(args.observe_matrix))]
for j, a in enumerate(job_args):
a.observe_matrix = args.observe_matrix[j]
a.observe = os.path.split(args.observe_matrix[j])[1]
a.subtask = True
a.func = None
a.iteration = 0
a.max_iterations = 1
a.quiet_mode = True
if j % 1000 == 0:
print j
if args.run_local:
pool = multiprocessing.Pool()
else:
pool = sge.SGEPool()
job_handle = pool.imap_unordered(do_inference, job_args)
converged = False
for args.iteration in range(args.max_iterations):
#import ipdb; ipdb.set_trace()
# fresh parameters-- to be aggregated after jobs are run
print 'iteration', args.iteration
total_free = 0
args.theta = sp.zeros((K,K,K), dtype=sp.float64)
args.alpha = sp.zeros((K,K), dtype=sp.float64)
args.beta = sp.zeros((K,K), dtype=sp.float64)
args.gamma = sp.zeros((K), dtype=sp.float64)
args.emit_probs = sp.zeros((K,L), dtype=sp.float64)
args.emit_sum = sp.zeros(K, sp.float64)
if args.run_local:
iterator = pool.imap_unordered(do_inference, job_args)
# wait for jobs to finish
for result in iterator:
pass
else:
jobs_handle = pool.map_async(do_inference, job_args, chunksize=100)
# wait for all jobs to finish
for j in jobs_handle:
j.wait()
# sum free energies and parameters from jobs
for a in job_args:
#print '# loading from %s' % a.observe
free_energy, theta, alpha, beta, gamma, emit_probs, emit_sum = load_params(a)
if len(free_energy) > 0:
last_free_energy = free_energy[-1]
else:
last_free_energy = 0
total_free += last_free_energy
args.theta += theta
args.alpha += alpha
args.beta += beta
args.gamma += gamma
args.emit_probs += emit_probs
args.emit_sum += emit_sum
# renormalize and plot
print 'normalize aggregation... total free energy is:', total_free
args.free_energy.append(total_free)
if len(args.free_energy) > 1 and args.free_energy[-1] != 0 and args.free_energy[-2] != 0 \
and abs(args.free_energy[-2] - args.free_energy[-1]) / args.free_energy[-2] < args.epsilon:
converged = True
normalize_trans(args.theta, args.alpha, args.beta, args.gamma)
args.emit_probs[:] = sp.dot(sp.diag(1./args.emit_sum), args.emit_probs)
for a in job_args:
a.theta, a.alpha, a.beta, a.gamma, a.emit_probs = args.theta, args.alpha, args.beta, args.gamma, args.emit_probs
for p in ['free_energy', 'theta', 'alpha', 'beta', 'gamma', 'emit_probs']:
sp.save(os.path.join(args.out_dir, args.out_params.format(param=p, **args.__dict__)),
args.__dict__[p])
plot_params(args)
plot_energy(args)
if args.save_Q >= 3:
print '# reconstructing chromosomes from *chunk*',
in_order = {}
# Q_chr16_all.trimmed.chunk*.npy => Q_chr16_all.trimmed.npy
all_chunks = glob.glob(os.path.join(args.out_dir, '*_Q_*chunk*.npy'))
for chunk in all_chunks:
print chunk
chunk_num = int(re.search(r'chunk(\d+)', chunk).groups()[0])
chrom_out = re.sub('chunk(\d+)\.', '', chunk)
if chrom_out not in in_order:
in_order[chrom_out] = {}
in_order[chrom_out][chunk_num] = sp.load(chunk)
for chrom_out in in_order:
print 'reconstructing chromosomes from', in_order[chrom_out]
if len(in_order[chrom_out]) > 1:
final_array = sp.concatenate((in_order[chrom_out][0], in_order[chrom_out][1]), axis=1)
for i in range(2, max(in_order[chrom_out])):
final_array = sp.concatenate((final_array, in_order[chrom_out][i]), axis=1)
else:
final_array = in_order[chrom_out][0]
sp.save(chrom_out, final_array)
if converged:
break
def independent_update_params(args, renormalize=True):
X = args.X
Q, Q_pairs, theta, alpha, beta, gamma, vert_parent, vert_children, log_obs_mat, pseudocount = (
args.Q, args.Q_pairs, args.theta,
args.alpha, args.beta,
args.gamma, args.vert_parent, args.vert_children, args.log_obs_mat, args.pseudocount)
I, T, K = Q.shape
L = X.shape[2]
if args.continuous_observations:
new_means = np.zeros_like(args.means)
new_variances = np.zeros_like(args.variances)
total_q = np.zeros_like(args.variances)
else:
emit_probs = args.emit_probs
emit_probs[:] = pseudocount
theta[:] = pseudocount
alpha[:] = pseudocount
beta[:] = pseudocount
gamma[:] = pseudocount
for i in xrange(I):
vp = vert_parent[i]
for t in xrange(T):
for k in xrange(K):
if i==0 and t==0:
gamma[k] += Q[i, t, k]
else:
for v in xrange(K):
if t == 0:
beta[v,k] += Q[vp,t,v] * Q[i,t,k]
else:
alpha[v,k] += Q_pairs[i,t,v,k]
if not args.continuous_observations:
for l in xrange(L):
if X[i,t,l]:
emit_probs[k, l] += Q[i, t, k]
if args.continuous_observations:
for i in xrange(I):
for t in xrange(T):
for k in xrange(K):
for l in xrange(L):
new_means[k,l] += Q[i, t, k] * X[i,t,l] # expectation of X wrt Q
total_q[k,l] += Q[i,t,k]
args.means[:] = new_means = new_means / total_q + 1e-50
np.seterr(under='ignore')
for i in xrange(I):
for t in xrange(T):
for k in xrange(K):
for l in xrange(L):
new_variances[k,l] += Q[i, t, k] * (X[i,t,l] - new_means[k,l]) * (X[i,t,l] - new_means[k,l])
np.seterr(under='print')
args.variances[:] = new_variances / total_q # 1 / N_k
args.variances += pseudocount
else:
normalize_emit(Q, emit_probs, pseudocount, args, renormalize)
if renormalize:
theta += theta.max() * (pseudocount * 1e-20)
alpha += alpha.max() * (pseudocount * 1e-20)
beta += beta.max() * (pseudocount * 1e-20)
gamma += gamma.max() * (pseudocount * 1e-20)
normalize_trans(theta, alpha, beta, gamma)
if args.continuous_observations:
make_log_obs_matrix_gaussian(args)
else:
make_log_obs_matrix(args)
def independent_update_params(args, renormalize=True):
X = args.X
Q, Q_pairs, theta, alpha, beta, gamma, vert_parent, vert_children, log_obs_mat, pseudocount = (
args.Q, args.Q_pairs, args.theta, args.alpha, args.beta, args.gamma,
args.vert_parent, args.vert_children, args.log_obs_mat,
args.pseudocount)
I, T, K = Q.shape
L = X.shape[2]
if args.continuous_observations:
new_means = np.zeros_like(args.means)
new_variances = np.zeros_like(args.variances)
total_q = np.zeros_like(args.variances)
else:
emit_probs = args.emit_probs
emit_probs[:] = pseudocount
theta[:] = pseudocount
alpha[:] = pseudocount
beta[:] = pseudocount
gamma[:] = pseudocount
for i in xrange(I):
vp = vert_parent[i]
for t in xrange(T):
for k in xrange(K):
if i == 0 and t == 0:
gamma[k] += Q[i, t, k]
else:
for v in xrange(K):
if t == 0:
beta[v, k] += Q[vp, t, v] * Q[i, t, k]
else:
alpha[v, k] += Q_pairs[i, t, v, k]
if not args.continuous_observations:
for l in xrange(L):
if X[i, t, l]:
emit_probs[k, l] += Q[i, t, k]
if args.continuous_observations:
for i in xrange(I):
for t in xrange(T):
for k in xrange(K):
for l in xrange(L):
new_means[k,
l] += Q[i, t,
k] * X[i, t,
l] # expectation of X wrt Q
total_q[k, l] += Q[i, t, k]
args.means[:] = new_means = new_means / total_q + 1e-50
np.seterr(under='ignore')
for i in xrange(I):
for t in xrange(T):
for k in xrange(K):
for l in xrange(L):
new_variances[k, l] += Q[i, t, k] * (
X[i, t, l] - new_means[k, l]) * (X[i, t, l] -
new_means[k, l])
np.seterr(under='print')
args.variances[:] = new_variances / total_q # 1 / N_k
args.variances += pseudocount
else:
normalize_emit(Q, emit_probs, pseudocount, args, renormalize)
if renormalize:
theta += theta.max() * (pseudocount * 1e-20)
alpha += alpha.max() * (pseudocount * 1e-20)
beta += beta.max() * (pseudocount * 1e-20)
gamma += gamma.max() * (pseudocount * 1e-20)
normalize_trans(theta, alpha, beta, gamma)
if args.continuous_observations:
make_log_obs_matrix_gaussian(args)
else:
make_log_obs_matrix(args)
def bp_update_params_new(args, renormalize=True):
lmds, pis = args.lmds, args.pis
vert_parent, vert_children = args.vert_parent, args.vert_children
#print pis[0].shape
I, T, L = args.X.shape
K = args.gamma.shape[0]
theta, alpha, beta, gamma, emit_probs, X = (args.theta, args.alpha,
args.beta, args.gamma,
args.emit_probs, args.X)
evidence = args.evidence #evid_allchild(lmds, vert_children)
emit_probs_mat = sp.exp(args.log_obs_mat)
#emit_probs_mat /= emit_probs_mat.max(axis=2).reshape(I, T, 1)
gamma_p = copy.copy(gamma)
alpha_p = copy.copy(alpha)
beta_p = copy.copy(beta)
theta_p = copy.copy(theta)
# emit_probs_p = copy.copy(emit_probs)
theta[:] = args.pseudocount
alpha[:] = args.pseudocount
beta[:] = args.pseudocount
gamma[:] = args.pseudocount
emit_probs[:] = args.pseudocount
#evidence = evid_allchild(lmds, args.vert_children)
##support = casual_support(pis)
emit_sum = sp.zeros((K, L))
for i in xrange(I):
vp = vert_parent[i]
if i != 0:
idx_i = vert_children[vp].tolist().index(i)
for t in xrange(T):
if i == 0 and t == 0:
gamma += emit_probs_mat[i, t, :] * evidence[i, t, :]
Q = emit_probs_mat[i, t, :] * evidence[i, t, :]
elif i == 0:
tmp1, tmp2 = sp.ix_(
pis[i][-1, t - 1, :],
emit_probs_mat[i, t, :] * evidence[i, t, :])
tmp = alpha_p * (tmp1 * tmp2) # belief
#tmp /= tmp.sum()
Q = tmp.sum(axis=0)
alpha += tmp / tmp.sum()
elif t == 0:
tmp1, tmp2 = sp.ix_(pis[vp][idx_i,
t, :], emit_probs_mat[i, t, :] *
evidence[i, t, :]) # i-1->0
tmp = beta_p * (tmp1 * tmp2) # belief
Q = tmp.sum(axis=0)
beta += tmp / tmp.sum()
else:
tmp1, tmp2, tmp3 = sp.ix_(
pis[vp][idx_i, t, :], pis[i][-1, t - 1, :],
emit_probs_mat[i, t, :] * evidence[i, t, :])
tmp = theta_p * (tmp1 * tmp2 * tmp3)
Q = (tmp.sum(axis=0)).sum(axis=0)
theta += tmp / tmp.sum()
Q /= Q.sum()
for l in xrange(L):
if args.mark_avail[i, l] and X[i, t, l]:
emit_probs[:, l] += Q
emit_sum[:, l] += Q
if renormalize:
normalize_trans(theta, alpha, beta, gamma)
emit_probs[:] = sp.dot(sp.diag(1. / emit_sum), emit_probs)
args.emit_sum = emit_sum
make_log_obs_matrix(args)