def _compute_alpha(self, particleWeights, finalIndex):
# Remove the weight of the chosen xi from the list instead of
# trying to subtract in logspace to prevent catastrophic
# cancellation (for the same reason as
# PGibbsOperator._compute_alpha)
particleWeightsNoXi = copy.copy(particleWeights)
particleWeightsNoXi.pop(finalIndex)
weightMinusXi = logaddexp(particleWeightsNoXi)
weightMinusRho = logaddexp(particleWeights[0:-1])
alpha = weightMinusRho - weightMinusXi
return alpha
def _compute_alpha(self, rhoWeight, xiWeights, finalIndex):
# Remove the weight of the chosen xi from the list instead of
# trying to subtract in logspace to prevent catastrophic
# cancellation (as would happen if the chosen xi weight were
# significantly larger than all the other xi weights and the rho
# weight).
otherXiWeightsWithRho = copy.copy(xiWeights)
otherXiWeightsWithRho.pop(finalIndex)
otherXiWeightsWithRho.append(rhoWeight)
weightMinusXi = logaddexp(otherXiWeightsWithRho)
weightMinusRho = logaddexp(xiWeights)
alpha = weightMinusRho - weightMinusXi
return alpha
def dp_inside_batch(batch_size,sentence_len,tags_dim,weights):
inside_table = torch.DoubleTensor(batch_size, sentence_len * sentence_len * 8, tags_dim, tags_dim)
inside_table.fill_(-np.inf)
if torch.cuda.is_available():
inside_table = inside_table.cuda()
m = sentence_len
seed_spans, base_left_spans, base_right_spans, left_spans, right_spans, ijss, ikss, kjss, id_span_map, span_id_map = test_constituent_indexes(
m, False)
for ii in seed_spans:
inside_table[:, ii, :, :] = 0.0
for ii in base_right_spans:
(l, r, c) = id_span_map[ii]
swap_weights = weights.permute(0, 1, 4, 3, 2)
inside_table[:, ii, :, :] = swap_weights[:, r, :, l, :]
for ii in base_left_spans:
(l, r, c) = id_span_map[ii]
inside_table[:, ii, :, :] = weights[:, l, :, r, :]
for ij in ijss:
(l, r, c) = id_span_map[ij]
if ij in left_spans:
ids = span_id_map.get((l, r, get_state_code(0, 0, 0)), -1)
prob = inside_table[:, ids, :, :] + weights[:, l, :, r, :]
inside_table[:, ij, :, :] = utils.logaddexp(inside_table[:, ij, :, :], prob)
elif ij in right_spans:
ids = span_id_map.get((l, r, get_state_code(0, 0, 0)), -1)
swap_weights = weights.permute(0, 1, 4, 3, 2)
prob = inside_table[:, ids, :, :] + swap_weights[:, r, :, l, :]
inside_table[:, ij, :, :] = utils.logaddexp(inside_table[:, ij, :, :], prob)
else:
num_k = len(ikss[ij])
beta_ik, beta_kj = inside_table[:, ikss[ij], :, :], inside_table[:, kjss[ij], :, :]
probs = beta_ik.contiguous().view(batch_size, num_k, tags_dim, tags_dim, 1) +\
beta_kj.contiguous().view(batch_size, num_k, 1, tags_dim, tags_dim)
probs = utils.logsumexp(probs, axis=(1, 3))
inside_table[:, ij, :, :] = utils.logaddexp(inside_table[:, ij, :, :], probs)
id1 = span_id_map.get((0, m - 1, get_state_code(0, 1, 0)), -1)
id2 = span_id_map.get((0, m - 1, get_state_code(0, 1, 1)), -1)
score1 = inside_table[:, id1, 0, :].contiguous().view(batch_size, 1, tags_dim)
score2 = inside_table[:, id2, 0, :].contiguous().view(batch_size, 1, tags_dim)
ll = utils.logaddexp(utils.logsumexp(score1, axis=2), utils.logsumexp(score2, axis=2))
return inside_table, ll
def propose(self, trace, scaffold):
program = trace.proposal_programs[self.program_name]
pnodes = [node for tar in program.tlist for node in trace.getNodesInBlock(tar[0],tar[1])]
cnodes = [node for cond in program.clist for node in trace.scopes[cond[0]][cond[1]]]
# drawScaffoldKernel(trace,scaffold,pnodes,cnodes,program.tlist,program.clist)
try:
assert len(pnodes) == program.n_target
except:
raise(Exception('Expect to have one and only one random node in a block. Check if 1) multiple nodes are defined in a target block; 2) some target nodes are observed (and thus not random).'))
old_target = [node.value.number for node in pnodes]
conditioned = [node.value.number for node in cnodes]
# print conditioned
registerDeterministicLKernels(trace,scaffold,pnodes,map(lambda x:VentureNumber(x),old_target))
rhoWeight = self.prepare(trace, scaffold)
# print "CustomMHOperator rhoWeight: ", rhoWeight, "old_target: ", old_target
# print rhoWeight
# TODO: get conditioned values
# TODO: invoke program, get target values and qratio from proposal
# print 'cond:', conditioned
# print 'old tar:', old_target
# (new_target, qratio) = program.propose(conditioned,old_target)
conditioned_labels, latent_labels, target_labels = [], [], []
conditioned_src = program.gen_conditioned_src(conditioned)
if conditioned_src:
(conditioned_labels, conditioned_strings, _) = execute_and_record(program.ripl, conditioned_src, "conditioned")
latent_src = program.gen_latent_src(conditioned)
if latent_src:
(latent_labels, latent_strings, _) = execute_and_record(program.ripl, latent_src, "latent")
target_src = program.gen_target_src(conditioned)
if target_src:
(target_labels, target_strings, new_target) = execute_and_record(program.ripl, target_src, "target")
else:
raise(Exception("No target_src found. Procedure 'gen_target_src' must be specified."))
for label in reversed(target_labels):
program.ripl.forget(label)
if self.method == 'assumed_gibbs':
qratio = 1e10
else:
old_logscores, new_logscores = [], []
if not latent_src: self.mc_samples = 1
for i in range(self.mc_samples):
predict_to_observe(program.ripl, target_strings, new_target)
new_logscores.append(sum_directive_logscore(program.ripl, target_labels))
for label in reversed(target_labels):
program.ripl.forget(label)
predict_to_observe(program.ripl, target_strings, old_target)
old_logscores.append(sum_directive_logscore(program.ripl, target_labels))
for label in reversed(target_labels):
program.ripl.forget(label)
for label in reversed(latent_labels):
program.ripl.forget(label)
execute_and_record(program.ripl, latent_src, "latent")
old_logscore = logaddexp(old_logscores)
new_logscore = logaddexp(new_logscores)
qratio = old_logscore - new_logscore
# print "CustomMHOperator old_logscores: ", old_logscores, "CustomMHOperator new_logscores: ", new_logscores
for label in reversed(latent_labels):
program.ripl.forget(label)
for label in reversed(conditioned_labels):
program.ripl.forget(label)
# print map(lambda x:VentureNumber(x),new_target)
# TODO: use DeterministicLKernel to fill in new target values
registerDeterministicLKernels(trace,scaffold,pnodes,map(lambda x:VentureNumber(x),new_target))
# TODO: get xiWeight
# print scaffold.border[0]
xiWeight = regenAndAttach(trace,scaffold.border[0],scaffold,False,OmegaDB(),{})
# print "CustomMHOperator xiWeight: ", xiWeight, "new_target: ", new_target
# print "CustomMHOperator qratio: ", qratio
# TODO: return (new_trace, xiWeight + qratio - rhoWeight)
# print '====================='
# print 'new_target', new_target
# print 'xiWeight:', xiWeight
# print 'qratio:', qratio
# print 'rhoWeight:', rhoWeight
return (trace, xiWeight + qratio - rhoWeight)
def batch_outside(self, inside_table, crf_score):
inside_complete_table = inside_table[0]
inside_incomplete_table = inside_table[1]
outside_complete_table = torch.DoubleTensor(
self.batch_size, self.sentence_length * self.sentence_length * 2,
self.tag_num)
outside_incomplete_table = torch.DoubleTensor(
self.batch_size, self.sentence_length * self.sentence_length * 2,
self.tag_num, self.tag_num)
if torch.cuda.is_available():
outside_complete_table = outside_complete_table.cuda()
outside_incomplete_table = outside_incomplete_table.cuda()
span_2_id, id_2_span, ijss, ikcs, ikis, kjcs, kjis, basic_span = utils.constituent_index(
self.sentence_length, False)
outside_complete_table.fill_(LOGZERO)
outside_incomplete_table.fill_(LOGZERO)
root_id = span_2_id.get((0, self.sentence_length - 1, 1))
outside_complete_table[:, root_id, 0] = 0.0
complete_span_used = set()
incomplete_span_used = set()
complete_span_used.add(root_id)
for ij in reversed(ijss):
(l, r, dir) = id_2_span[ij]
# complete span consists of one incomplete span and one complete span
num_kc = len(ikcs[ij])
if dir == 0:
outside_ij_cc = outside_complete_table[:, ij, :].contiguous(
).view(self.batch_size, 1, 1, self.tag_num)
inside_kj_ic = inside_incomplete_table[:, kjcs[
ij], :, :].contiguous().view(self.batch_size, num_kc,
self.tag_num, self.tag_num)
inside_ik_cc = inside_complete_table[:,
ikcs[ij], :].contiguous(
).view(
self.batch_size,
num_kc, self.tag_num,
1)
outside_ik_cc = (outside_ij_cc + inside_kj_ic).permute(
0, 1, 3, 2)
# swap left-right position since right tags are to be indexed
outside_kj_ic = outside_ij_cc + inside_ik_cc
for i in range(num_kc):
ik = ikcs[ij][i]
kj = kjcs[ij][i]
outside_ik_cc_i = utils.logsumexp(outside_ik_cc[:,
i, :, :],
axis=1)
if ik in complete_span_used:
outside_complete_table[:, ik, :] = utils.logaddexp(
outside_complete_table[:, ik, :], outside_ik_cc_i)
else:
outside_complete_table[:,
ik, :] = outside_ik_cc_i.clone(
)
complete_span_used.add(ik)
if kj in incomplete_span_used:
outside_incomplete_table[:, kj, :, :] = utils.logaddexp(
outside_incomplete_table[:, kj, :, :],
outside_kj_ic[:, i, :, :])
else:
outside_incomplete_table[:,
kj, :, :] = outside_kj_ic[:,
i, :, :]
incomplete_span_used.add(kj)
else:
outside_ij_cc = outside_complete_table[:, ij, :].contiguous(
).view(self.batch_size, 1, self.tag_num, 1)
inside_ik_ic = inside_incomplete_table[:, ikcs[
ij], :, :].contiguous().view(self.batch_size, num_kc,
self.tag_num, self.tag_num)
inside_kj_cc = inside_complete_table[:,
kjcs[ij], :].contiguous(
).view(
self.batch_size,
num_kc, 1,
self.tag_num)
outside_kj_cc = outside_ij_cc + inside_ik_ic
outside_ik_ic = outside_ij_cc + inside_kj_cc
for i in range(num_kc):
kj = kjcs[ij][i]
ik = ikcs[ij][i]
outside_kj_cc_i = utils.logsumexp(outside_kj_cc[:,
i, :, :],
axis=1)
if kj in complete_span_used:
outside_complete_table[:, kj, :] = utils.logaddexp(
outside_complete_table[:, kj, :], outside_kj_cc_i)
else:
outside_complete_table[:,
kj, :] = outside_kj_cc_i.clone(
)
complete_span_used.add(kj)
if ik in incomplete_span_used:
outside_incomplete_table[:, ik, :, :] = utils.logaddexp(
outside_incomplete_table[:, ik, :, :],
outside_ik_ic[:, i, :, :])
else:
outside_incomplete_table[:,
ik, :, :] = outside_ik_ic[:,
i, :, :]
incomplete_span_used.add(ik)
# incomplete span consists of two complete spans
num_ki = len(ikis[ij])
outside_ij_ii = outside_incomplete_table[:, ij, :, :].contiguous(
).view(self.batch_size, 1, self.tag_num, self.tag_num)
inside_ik_ci = inside_complete_table[:, ikis[ij], :].contiguous(
).view(self.batch_size, num_ki, self.tag_num, 1)
inside_kj_ci = inside_complete_table[:, kjis[ij], :].contiguous(
).view(self.batch_size, num_ki, 1, self.tag_num)
if dir == 0:
outside_ik_ci = outside_ij_ii + inside_kj_ci + crf_score[:, r, l, :, :]. \
permute(0, 2, 1).contiguous().view(self.batch_size, 1, self.tag_num, self.tag_num)
outside_kj_ci = outside_ij_ii + inside_ik_ci + crf_score[:, r, l, :, :]. \
permute(0, 2, 1).contiguous().view(self.batch_size, 1, self.tag_num, self.tag_num)
else:
outside_ik_ci = outside_ij_ii + inside_kj_ci + crf_score[:, l, r, :, :].contiguous(
).view(self.batch_size, 1, self.tag_num, self.tag_num)
outside_kj_ci = outside_ij_ii + inside_ik_ci + crf_score[:, l, r, :, :].contiguous(
).view(self.batch_size, 1, self.tag_num, self.tag_num)
for i in range(num_ki):
ik = ikis[ij][i]
kj = kjis[ij][i]
outside_ik_ci_i = utils.logsumexp(outside_ik_ci[:, i, :, :],
axis=2)
outside_kj_ci_i = utils.logsumexp(outside_kj_ci[:, i, :, :],
axis=1)
if ik in complete_span_used:
outside_complete_table[:, ik, :] = utils.logaddexp(
outside_complete_table[:, ik, :], outside_ik_ci_i)
else:
outside_complete_table[:, ik, :] = outside_ik_ci_i.clone()
complete_span_used.add(ik)
if kj in complete_span_used:
outside_complete_table[:, kj, :] = utils.logaddexp(
outside_complete_table[:, kj, :], outside_kj_ci_i)
else:
outside_complete_table[:, kj, :] = outside_kj_ci_i.clone()
complete_span_used.add(kj)
return (outside_complete_table, outside_incomplete_table)
def dp_outside_batch(self, inside_table, weights):
outside_table = torch.DoubleTensor(self.batch_size, self.sentence_len * self.sentence_len * 8, self.tags_dim, self.tags_dim)
outside_table.fill_(LOGZERO)
if torch.cuda.is_available():
outside_table = outside_table.cuda()
m = self.sentence_len
seed_spans, base_left_spans, base_right_spans, left_spans, right_spans, ijss, ikss, kjss, id_span_map, span_id_map = utils.constituent_indexes(
m, self.is_multi_root)
id1 = span_id_map.get((0, m - 1, utils.get_state_code(0, 1, 0)), -1)
id2 = span_id_map.get((0, m - 1, utils.get_state_code(0, 1, 1)), -1)
outside_table[:, id1, :, :] = 0.0
outside_table[:, id2, :, :] = 0.0
for ij in reversed(ijss):
(l, r, c) = id_span_map[ij]
if ij in left_spans:
assert c == utils.get_state_code(0, 1, 1)
prob = outside_table[:, ij, :, :] + weights[:, l, :, r, :]
ids = span_id_map.get((l, r, utils.get_state_code(0, 0, 0)), -1)
outside_table[:, ids, :, :] = utils.logaddexp(outside_table[:, ids, :, :], prob)
elif ij in right_spans:
assert c == utils.get_state_code(1, 0, 1)
swap_weights = weights.permute(0, 1, 4, 3, 2)
prob = outside_table[:, ij, :, :] + swap_weights[:, r, :, l, :]
ids = span_id_map.get((l, r, utils.get_state_code(0, 0, 0)), -1)
outside_table[:, ids, :, :] = utils.logaddexp(outside_table[:, ids, :, :], prob)
else:
num_k = len(ikss[ij])
if l == 0:
# ROOT's value can only be 0
alpha_ij = outside_table[:, ij, 0, :].contiguous().view(self.batch_size, 1, 1, 1, self.tags_dim)
beta_left = inside_table[:, ikss[ij], [0], :].contiguous().view(self.batch_size, num_k, 1, self.tags_dim, 1)
beta_right = inside_table[:, kjss[ij], :, :].contiguous().view(self.batch_size, num_k, 1, self.tags_dim, self.tags_dim)
new_left = alpha_ij + beta_right
new_left = utils.logsumexp(new_left, axis=4).contiguous().view(self.batch_size, num_k, 1, self.tags_dim, 1)
new_right = alpha_ij + beta_left
if len(list(set(ikss[ij]))) == num_k:
outside_table[:, ikss[ij], [0], :] = utils.logaddexp(
outside_table[:, ikss[ij], [0], :].contiguous().view(self.batch_size, num_k, 1, self.tags_dim, 1), new_left).contiguous().view(self.batch_size, num_k, self.tags_dim)
outside_table[:, kjss[ij], :, :] = utils.logaddexp(
outside_table[:, kjss[ij], :, :].contiguous().view(self.batch_size, num_k, 1, self.tags_dim, self.tags_dim),
new_right).contiguous().view(self.batch_size, num_k, self.tags_dim, self.tags_dim)
else:
alpha_ij = outside_table[:, ij, :, :].contiguous().view(self.batch_size, 1, self.tags_dim, 1, self.tags_dim)
beta_left = inside_table[:, ikss[ij], :, :].contiguous().view(self.batch_size, num_k, self.tags_dim, self.tags_dim, 1)
beta_right = inside_table[:, kjss[ij], :, :].contiguous().view(self.batch_size, num_k, 1, self.tags_dim, self.tags_dim)
new_left = alpha_ij + beta_right
new_left = utils.logsumexp(new_left, axis=4).contiguous().view(self.batch_size, num_k, self.tags_dim, self.tags_dim, 1)
new_right = alpha_ij + beta_left
new_right = utils.logsumexp(new_right, axis=2).contiguous().view(self.batch_size, num_k, 1, self.tags_dim, self.tags_dim)
if len(list(set(ikss[ij]))) == num_k:
outside_table[:, ikss[ij], :, :] = utils.logaddexp(
outside_table[:, ikss[ij], :, :].contiguous().view(self.batch_size, num_k, self.tags_dim, self.tags_dim, 1),
new_left).contiguous().view(self.batch_size, num_k, self.tags_dim, self.tags_dim)
outside_table[:, kjss[ij], :, :] = utils.logaddexp(outside_table[:, kjss[ij], :, :].contiguous().view(self.batch_size, num_k, 1, self.tags_dim, self.tags_dim), new_right).contiguous().view(self.batch_size, num_k, self.tags_dim, self.tags_dim)
if len(list(set(ikss[ij]))) == num_k:
# Already done in the above
pass
else:
# TODO make this vectorized, the problem is the id in ikss is not unique
for i in range(num_k):
ik = ikss[ij][i]
kj = kjss[ij][i]
if l == 0:
alpha_ij = outside_table[:, ij, 0, :].contiguous().view(self.batch_size, 1, 1, 1, self.tags_dim)
beta_right = inside_table[:, kj, :, :].contiguous().view(self.batch_size, 1, 1, self.tags_dim, self.tags_dim)
new_left = alpha_ij + beta_right
new_left = utils.logsumexp(new_left, axis=4).contiguous().view(self.batch_size, 1, 1, self.tags_dim, 1)
outside_table[:, ik, 0, :] = utils.logaddexp(
outside_table[:, ik, 0, :].contiguous().view(self.batch_size, 1, 1, self.tags_dim, 1), new_left).contiguous().view(self.batch_size, self.tags_dim, )
else:
alpha_ij = outside_table[:, ij, :, :].contiguous().view(self.batch_size, 1, self.tags_dim, 1, self.tags_dim)
beta_right = inside_table[:, kj, :, :].contiguous().view(self.batch_size, 1, 1, self.tags_dim, self.tags_dim)
new_left = alpha_ij + beta_right
new_left = utils.logsumexp(new_left, axis=4).contiguous().view(self.batch_size, 1, self.tags_dim, self.tags_dim, 1)
outside_table[:, ik, :, :] = utils.logaddexp(outside_table[:, ik, :, :].contiguous().view(self.batch_size, 1, self.tags_dim, self.tags_dim, 1), new_left).contiguous().view(self.batch_size, self.tags_dim, self.tags_dim)
for ij in base_left_spans:
(l, r, c) = id_span_map[ij]
prob = outside_table[:, ij, :, :] + weights[:, l, :, r, :]
ids = span_id_map.get((l, r, utils.get_state_code(0, 0, 1)), -1)
outside_table[:, ids, :, :] = utils.logaddexp(outside_table[:, ids, :, :], prob)
for ij in base_right_spans:
(l, r, c) = id_span_map[ij]
swap_weights = weights.permute(0, 1, 4, 3, 2)
prob = outside_table[:, ij, :, :] + swap_weights[:, r, :, l, :]
ids = span_id_map.get((l, r, utils.get_state_code(0, 0, 1)), -1)
outside_table[:, ids, :, :] = utils.logaddexp(outside_table[:, ids, :, :], prob)
return outside_table
