本文整理匯總了Python中utils.rms方法的典型用法代碼示例。如果您正苦於以下問題：Python utils.rms方法的具體用法？Python utils.rms怎麽用？Python utils.rms使用的例子？那麽, 這裏精選的方法代碼示例或許可以為您提供幫助。您也可以進一步了解該方法所在類utils的用法示例。
在下文中一共展示了utils.rms方法的5個代碼示例，這些例子默認根據受歡迎程度排序。您可以為喜歡或者感覺有用的代碼點讚，您的評價將有助於係統推薦出更棒的Python代碼示例。
示例1: _create_loss
# 需要導入模塊: import utils [as 別名]
# 或者: from utils import rms [as 別名]
def _create_loss(self):
# Hard loss
logQHard, samples = self._recognition_network()
reinforce_learning_signal, reinforce_model_grad = self._generator_network(samples, logQHard)
logQHard = tf.add_n(logQHard)
# REINFORCE
learning_signal = tf.stop_gradient(U.center(reinforce_learning_signal))
self.optimizerLoss = -(learning_signal*logQHard +
reinforce_model_grad)
self.lHat = map(tf.reduce_mean, [
reinforce_learning_signal,
U.rms(learning_signal),
])
return reinforce_learning_signal 示例2: get_nvil_gradient
# 需要導入模塊: import utils [as 別名]
# 或者: from utils import rms [as 別名]
def get_nvil_gradient(self):
"""Compute the NVIL gradient."""
# Hard loss
logQHard, samples = self._recognition_network()
ELBO, reinforce_model_grad = self._generator_network(samples, logQHard)
logQHard = tf.add_n(logQHard)
# Add baselines (no variance normalization)
learning_signal = tf.stop_gradient(ELBO) - self._create_baseline()
# Set up losses
self.baseline_loss.append(tf.square(learning_signal))
optimizerLoss = -(tf.stop_gradient(learning_signal)*logQHard +
reinforce_model_grad)
optimizerLoss = tf.reduce_mean(optimizerLoss)
nvil_gradient = self.optimizer_class.compute_gradients(optimizerLoss)
debug = {
'ELBO': ELBO,
'RMS of centered learning signal': U.rms(learning_signal),
}
return nvil_gradient, debug 示例3: _create_loss
# 需要導入模塊: import utils [as 別名]
# 或者: from utils import rms [as 別名]
def _create_loss(self):
# Hard loss
logQHard, samples = self._recognition_network()
reinforce_learning_signal, reinforce_model_grad = self._generator_network(samples, logQHard)
logQHard = tf.add_n(logQHard)
# REINFORCE
learning_signal = tf.stop_gradient(center(reinforce_learning_signal))
self.optimizerLoss = -(learning_signal*logQHard +
reinforce_model_grad)
self.lHat = map(tf.reduce_mean, [
reinforce_learning_signal,
U.rms(learning_signal),
])
return reinforce_learning_signal 示例4: get_simple_muprop_gradient
# 需要導入模塊: import utils [as 別名]
# 或者: from utils import rms [as 別名]
def get_simple_muprop_gradient(self):
""" Computes the simple muprop gradient.
This muprop control variate does not include the linear term.
"""
# Hard loss
logQHard, hardSamples = self._recognition_network()
hardELBO, reinforce_model_grad = self._generator_network(hardSamples, logQHard)
# Soft loss
logQ, muSamples = self._recognition_network(sampler=self._mean_sample)
muELBO, _ = self._generator_network(muSamples, logQ)
scaling_baseline = self._create_eta(collection='BASELINE')
learning_signal = (hardELBO
- scaling_baseline * muELBO
- self._create_baseline())
self.baseline_loss.append(tf.square(learning_signal))
optimizerLoss = -(tf.stop_gradient(learning_signal) * tf.add_n(logQHard)
+ reinforce_model_grad)
optimizerLoss = tf.reduce_mean(optimizerLoss)
simple_muprop_gradient = (self.optimizer_class.
compute_gradients(optimizerLoss))
debug = {
'ELBO': hardELBO,
'muELBO': muELBO,
'RMS': U.rms(learning_signal),
}
return simple_muprop_gradient, debug 示例5: get_muprop_gradient
# 需要導入模塊: import utils [as 別名]
# 或者: from utils import rms [as 別名]
def get_muprop_gradient(self):
"""
random sample function that actually returns mean
new forward pass that returns logQ as a list
can get x_i from samples
"""
# Hard loss
logQHard, hardSamples = self._recognition_network()
hardELBO, reinforce_model_grad = self._generator_network(hardSamples, logQHard)
# Soft loss
logQ, muSamples = self._recognition_network(sampler=self._mean_sample)
muELBO, _ = self._generator_network(muSamples, logQ)
# Compute gradients
muELBOGrads = tf.gradients(tf.reduce_sum(muELBO),
[ muSamples[i]['activation'] for
i in xrange(self.hparams.n_layer) ])
# Compute MuProp gradient estimates
learning_signal = hardELBO
optimizerLoss = 0.0
learning_signals = []
for i in xrange(self.hparams.n_layer):
dfDiff = tf.reduce_sum(
muELBOGrads[i] * (hardSamples[i]['activation'] -
muSamples[i]['activation']),
axis=1)
dfMu = tf.reduce_sum(
tf.stop_gradient(muELBOGrads[i]) *
tf.nn.sigmoid(hardSamples[i]['log_param']),
axis=1)
scaling_baseline_0 = self._create_eta(collection='BASELINE')
scaling_baseline_1 = self._create_eta(collection='BASELINE')
learning_signals.append(learning_signal - scaling_baseline_0 * muELBO - scaling_baseline_1 * dfDiff - self._create_baseline())
self.baseline_loss.append(tf.square(learning_signals[i]))
optimizerLoss += (
logQHard[i] * tf.stop_gradient(learning_signals[i]) +
tf.stop_gradient(scaling_baseline_1) * dfMu)
optimizerLoss += reinforce_model_grad
optimizerLoss *= -1
optimizerLoss = tf.reduce_mean(optimizerLoss)
muprop_gradient = self.optimizer_class.compute_gradients(optimizerLoss)
debug = {
'ELBO': hardELBO,
'muELBO': muELBO,
}
debug.update(dict([
('RMS learning signal layer %d' % i, U.rms(learning_signal))
for (i, learning_signal) in enumerate(learning_signals)]))
return muprop_gradient, debug
# REBAR gradient helper functions