def sample(self, eta, count=1):
exp_x0 = safe_exp(eta[0])
prob_x0 = exp_x0 / (exp_x0 + safe_exp(self.base_model.log_partition(eta[1:])))
results = np.zeros(count)
nonzero = np.random.random(size=count) > prob_x0
results[nonzero] = self.base_model.sample(eta[1:], count=nonzero.sum())
return results
def grad_log_partition(self, eta):
exp_base_log_partition = safe_exp(self.base_model.log_partition(eta[1:]))
exp_x0 = safe_exp(eta[0])
denominator = exp_base_log_partition + exp_x0
w = (exp_base_log_partition / denominator)
return np.concatenate(((exp_x0 / denominator)[:,np.newaxis].T,
self.base_model.grad_log_partition(eta[1:]) * w), axis=0)
def sample(self, eta, count=1):
exp_x0 = safe_exp(eta[0])
prob_x0 = exp_x0 / (exp_x0 +
safe_exp(self.base_model.log_partition(eta[1:])))
results = np.zeros(count)
nonzero = np.random.random(size=count) > prob_x0
results[nonzero] = self.base_model.sample(eta[1:], count=nonzero.sum())
return results
def grad_log_partition(self, eta):
exp_base_log_partition = safe_exp(
self.base_model.log_partition(eta[1:]))
exp_x0 = safe_exp(eta[0])
denominator = exp_base_log_partition + exp_x0
w = (exp_base_log_partition / denominator)
return np.concatenate(
((exp_x0 / denominator)[:, np.newaxis].T,
self.base_model.grad_log_partition(eta[1:]) * w),
axis=0)
def diagonal_hessian_log_partition(self, eta):
base_log_partition = self.base_model.log_partition(eta[1:])
exp_base_log_partition = safe_exp(base_log_partition)
exp_x0 = safe_exp(eta[0])
exp_sum = safe_exp(eta[0] + base_log_partition)
sum_exp = exp_base_log_partition + exp_x0
sq_sum_exp = safe_sq(sum_exp)
diag_hess_base = self.base_model.diagonal_hessian_log_partition(eta[1:])
sq_grad_base = safe_sq(self.base_model.grad_log_partition(eta[1:]))
numerator = np.zeros(diag_hess_base.shape)
numerator[:,sq_sum_exp != np.inf] = (sum_exp[sq_sum_exp != np.inf] * diag_hess_base[:, sq_sum_exp != np.inf] + exp_x0[sq_sum_exp != np.inf] * sq_grad_base[:, sq_sum_exp != np.inf]) / sq_sum_exp[sq_sum_exp != np.inf]
return np.concatenate(((exp_sum / sq_sum_exp)[:,np.newaxis].T,
exp_base_log_partition * numerator), axis=0)
def test_safe_exp(self):
""" Tests safe_exp. """
a = safe_exp(0)
self.assertEqual(int(a), 1)
max_float = sys.float_info.max
log_max = np.log(max_float)
a = safe_exp(log_max)
self.assertLess(a, float("inf"))
tmp_io = io.StringIO()
with redirect_stdout(tmp_io):
a = safe_exp(log_max * 10)
assert ('RuntimeWarning' not in tmp_io.getvalue())
def compute_perplexity(model, sess, name):
"""Compute perplexity of the output of the model.
Args:
model: model for compute perplexity.
sess: tensorflow session to use.
name: name of the batch.
Returns:
The perplexity of the eval outputs.
"""
total_loss = 0
total_predict_count = 0
start_time = time.time()
step = 0
while True:
try:
loss, predict_count, batch_size = model.eval(sess)
total_loss += loss * batch_size
total_predict_count += predict_count
step += 1
if step % 500 == 0:
ls = total_loss / total_predict_count
ppl = misc.safe_exp(ls)
print_out(" ## After %d steps, loss %.2f - ppl %.3f" %
(step, ls, ppl))
except tf.errors.OutOfRangeError:
break
perplexity = safe_exp(total_loss / total_predict_count)
print_time(" eval %s: perplexity %.2f" % (name, perplexity), start_time)
return perplexity
def compute_perplexity(model, sess, name):
"""Compute perplexity of the output of the model.
Args:
model: model for compute perplexity.
sess: tensorflow session to use.
name: name of the batch.
Returns:
The perplexity of the eval outputs.
"""
total_loss = 0
total_predict_count = 0
start_time = time.time()
while True:
try:
output_tuple = model.eval(sess)
total_loss += output_tuple.eval_loss * output_tuple.batch_size
total_predict_count += output_tuple.predict_count
except tf.errors.OutOfRangeError:
break
perplexity = utils.safe_exp(total_loss / total_predict_count)
utils.print_time(" eval %s: perplexity %.2f" % (name, perplexity),
start_time)
return perplexity
def rvs(self, n=None):
""" Returns independent observations of this random variable.
"""
mu = self.__mu
S = self.__S
if n is None:
normal_values = np.array(mnormal.rvs(mu, S), ndmin=1)
lognormal_values = [safe_exp(v) for v in normal_values]
return np.array(lognormal_values)
else:
all_lognormal_values = []
for _ in range(n):
normal_values = mnormal.rvs(mu, S)
lognormal_values = [safe_exp(v) for v in normal_values]
all_lognormal_values.append(lognormal_values)
return np.array(all_lognormal_values)
def diagonal_hessian_log_partition(self, eta):
base_log_partition = self.base_model.log_partition(eta[1:])
exp_base_log_partition = safe_exp(base_log_partition)
exp_x0 = safe_exp(eta[0])
exp_sum = safe_exp(eta[0] + base_log_partition)
sum_exp = exp_base_log_partition + exp_x0
sq_sum_exp = safe_sq(sum_exp)
diag_hess_base = self.base_model.diagonal_hessian_log_partition(
eta[1:])
sq_grad_base = safe_sq(self.base_model.grad_log_partition(eta[1:]))
numerator = np.zeros(diag_hess_base.shape)
numerator[:, sq_sum_exp != np.inf] = (
sum_exp[sq_sum_exp != np.inf] *
diag_hess_base[:, sq_sum_exp != np.inf] +
exp_x0[sq_sum_exp != np.inf] *
sq_grad_base[:, sq_sum_exp != np.inf]
) / sq_sum_exp[sq_sum_exp != np.inf]
return np.concatenate(((exp_sum / sq_sum_exp)[:, np.newaxis].T,
exp_base_log_partition * numerator),
axis=0)
def process_stats(hparams, stats, info, global_step, steps_per_stats, log_f):
"""Update info and check for overflow."""
# Per-step info
info["avg_step_time"] = stats["step_time"] / steps_per_stats
info["avg_grad_norm"] = stats["grad_norm"] / steps_per_stats
info["avg_sequence_count"] = stats["sequence_count"] / steps_per_stats
info["speed"] = stats["word_count"] / (1000 * stats["step_time"])
# Check for overflow
is_overflow = False
if hparams.task == "joint":
# Per-predict info
info["train_ppl"] = (utils.safe_exp(stats["slot_train_loss"] /
stats["predict_count"]))
train_ppl = info["train_ppl"]
if math.isnan(train_ppl) or math.isinf(train_ppl) or train_ppl > 1e20:
utils.print_out(" step %d overflow, stop early" % global_step,
log_f)
is_overflow = True
return is_overflow
def log_partition(self, eta):
return np.log(1 + safe_exp(eta))
def log_partition(self, eta):
return np.log(
safe_exp(eta[0]) +
safe_exp(self.base_model.log_partition(eta[1:])))
def sample(self, eta, count=1):
exp_eta = safe_exp(eta)
p = exp_eta / (1 + exp_eta)
return np.random.random(size=count) < p
def hessian_log_partition(self, eta):
exp_eta = safe_exp(eta)
return -exp_eta / safe_sq(exp_eta + 1)
def grad_log_partition(self, eta):
exp_eta = safe_exp(eta)
return exp_eta / (exp_eta + 1.0)
def log_partition(self, eta):
return np.log(safe_exp(eta[0]) + safe_exp(self.base_model.log_partition(eta[1:])))
def log_partition(self, eta):
return np.log(1 + safe_exp(eta))
def hessian_log_partition(self, eta):
exp_eta = safe_exp(eta)
return -exp_eta / safe_sq(exp_eta + 1)
def grad_log_partition(self, eta):
exp_eta = safe_exp(eta)
return exp_eta / (exp_eta + 1.0)
def pdf(self, x):
""" Returns an approximate value of the probability density
function of this random variable on point x. """
return safe_exp(self.log_pdf(x))
def sample(self, eta, count=1):
exp_eta = safe_exp(eta)
p = exp_eta / (1 + exp_eta)
return np.random.random(size=count) < p
