def _global_threshold(self):
estimator = self.session.estimator
gamma_name = 'NetworkSlimmer.gamma'
threshold_name = 'NetworkSlimmer.threshold'
if estimator.max_len(gamma_name) == 0:
try:
return estimator.get_value(threshold_name)
except KeyError:
raise RuntimeError(
'Train for a while before running update to collect '
'gamma values.')
# extract all gammas globally
gammas = []
for overrider, gamma in estimator.get_values(gamma_name).items():
if not overrider.should_update:
continue
if self.incremental:
mask = self.session.run(overrider.mask)
gamma = gamma[util.nonzero(mask)]
gammas += gamma.tolist()
threshold = self._threshold(gammas)
log.debug(
'Extracted a global threshold for all gammas: {}.'
.format(threshold))
estimator.flush_all(gamma_name)
estimator.add(threshold, threshold_name)
return threshold
def _quantize(self, value, point, width, compute_overflow_rate=False):
# decompose
sign = util.cast(value > 0, float) - util.cast(value < 0, float)
value = util.log(util.abs(value), 2.0)
# quantize
value = self.quantizer.apply(
value, compute_overflow_rate=compute_overflow_rate)
if compute_overflow_rate:
return value
# represent
return util.where(util.nonzero(sign), sign * (2**value), 0)
def _updated_mask(self, var, mask):
mask, gamma = self.session.run([mask, self.gamma])
if self.global_threshold:
threshold = self._global_threshold()
else:
if self.incremental:
gammas = gamma[util.nonzero(self.mask)]
threshold = self._threshold(gammas)
new_mask = gamma > threshold
if self.incremental:
return util.logical_and(mask, new_mask)
return new_mask