def estimate_loglikelihood_for_dataset(self,
x_dataset,
minibatch_size=1000):
"""
Estimate the average LOG LIKELIHOOD for the dataset.
NOTE: Because this is NADE and not a real distribution,
these likelihoods CAN and often ARE NEGATIVE!
Basically returns ld.sum() / n
"""
loglikelihood = 0.0
loglikelihood_sq = 0.0
n = 0
iterator = x_dataset.iterator(batch_size=minibatch_size,
get_smaller_final_batch=True)
for x in iterator:
x = x.T # VxB
n += x.shape[1]
# get negative log density for array
ld = self.logdensity(x)
# add together all negative log densities in array
loglikelihood += ld.sum()
loglikelihood_sq += (ld**2).sum()
e = Estimation.sample_mean_from_sum_and_sum_sq(loglikelihood,
loglikelihood_sq, n)
return e
def estimate_loglikelihood_for_dataset(self, x_dataset, minibatch_size=1000):
loglikelihood = 0.0
loglikelihood_sq = 0.0
n = 0
iterator = x_dataset.iterator(batch_size=minibatch_size, get_smaller_final_batch=True)
for x in iterator:
x = x.T # VxB
n += x.shape[1]
ld = self.logdensity(x)
loglikelihood += ld.sum()
loglikelihood_sq += (ld ** 2).sum()
return Estimation.sample_mean_from_sum_and_sum_sq(loglikelihood, loglikelihood_sq, n)
def estimate_loss_for_dataset(dataset, loss_f, minibatch_size=1000):
loss_sum = 0.0
loss_sq_sum = 0.0
n = 0
iterator = dataset.iterator(batch_size=minibatch_size, get_smaller_final_batch=True, shuffle=False)
for x in iterator:
if not isinstance(x, tuple):
x = [x]
n += x[0].shape[0]
x = [e.T for e in x]
loss = loss_f(*x)
loss_sum += loss.sum()
loss_sq_sum += (loss ** 2).sum()
return Estimation.sample_mean_from_sum_and_sum_sq(loss_sum, loss_sq_sum, n)
def estimate_loglikelihood_for_dataset(self,
x_dataset,
minibatch_size=1000):
loglikelihood = 0.0
loglikelihood_sq = 0.0
n = 0
iterator = x_dataset.iterator(batch_size=minibatch_size,
get_smaller_final_batch=True)
for x in iterator:
x = x.T # VxB
n += x.shape[1]
ld = self.logdensity(x)
loglikelihood += ld.sum()
loglikelihood_sq += (ld**2).sum()
return Estimation.sample_mean_from_sum_and_sum_sq(
loglikelihood, loglikelihood_sq, n)
def estimate_average_loglikelihood_for_dataset_using_masks(self, x_dataset, masks_dataset, minibatch_size=20000, loops=1):
loglikelihood = 0.0
loglikelihood_sq = 0.0
n = 0
x_iterator = x_dataset.iterator(batch_size=minibatch_size, get_smaller_final_batch=True)
m_iterator = masks_dataset.iterator(batch_size=minibatch_size)
for _ in xrange(loops):
for x, m in izip(x_iterator, m_iterator):
x = x.T # VxB
batch_size = x.shape[1]
m = m.T[:, :batch_size]
n += batch_size
lls = self.compiled_mask_logdensity_estimator(x, m)
loglikelihood += np.sum(lls)
loglikelihood_sq += np.sum(lls ** 2)
return Estimation.sample_mean_from_sum_and_sum_sq(loglikelihood, loglikelihood_sq, n)
def estimate_loss_for_dataset(dataset, loss_f, minibatch_size=1000):
loss_sum = 0.0
loss_sq_sum = 0.0
n = 0
iterator = dataset.iterator(batch_size=minibatch_size,
get_smaller_final_batch=True,
shuffle=False)
for x in iterator:
if not isinstance(x, tuple):
x = [x]
n += x[0].shape[0]
x = [e.T for e in x]
loss = loss_f(*x)
loss_sum += loss.sum()
loss_sq_sum += (loss**2).sum()
return Estimation.sample_mean_from_sum_and_sum_sq(loss_sum, loss_sq_sum, n)
def estimate_average_loglikelihood_for_dataset_using_masks(
self, x_dataset, masks_dataset, minibatch_size=20000, loops=1):
loglikelihood = 0.0
loglikelihood_sq = 0.0
n = 0
x_iterator = x_dataset.iterator(batch_size=minibatch_size,
get_smaller_final_batch=True)
m_iterator = masks_dataset.iterator(batch_size=minibatch_size)
for _ in xrange(loops):
for x, m in izip(x_iterator, m_iterator):
x = x.T # VxB
batch_size = x.shape[1]
m = m.T[:, :batch_size]
n += batch_size
lls = self.compiled_mask_logdensity_estimator(x, m)
loglikelihood += np.sum(lls)
loglikelihood_sq += np.sum(lls**2)
return Estimation.sample_mean_from_sum_and_sum_sq(
loglikelihood, loglikelihood_sq, n)
