def _sample(self, shape=(), mean=None, sd=None):
"""
Sampling.
Args:
shape (tuple): The shape of the sample. Default: ().
mean (Tensor): The mean of the samples. Default: self._mean_value.
sd (Tensor): The standard deviation of the samples. Default: self._sd_value.
Returns:
Tensor, shape is shape + batch_shape.
"""
shape = self.checktuple(shape, 'shape')
mean, sd = self._check_param_type(mean, sd)
batch_shape = self.shape(mean + sd)
origin_shape = shape + batch_shape
if origin_shape == ():
sample_shape = (1,)
else:
sample_shape = origin_shape
sample_norm = C.normal(sample_shape, mean, sd, self.seed)
value = self.cast(sample_norm, self.dtype)
if origin_shape == ():
value = self.squeeze(value)
return value
def _sample(self, shape=(), mean=None, sd=None):
"""
Sampling.
Args:
shape (tuple): shape of the sample. Default: ().
mean (Tensor): mean of the samples. Default: self._mean_value.
sd (Tensor): standard deviation of the samples. Default: self._sd_value.
Returns:
Tensor, shape is shape + batch_shape.
"""
mean = self.cast(
mean,
self.parameter_type) if mean is not None else self._mean_value
if mean is None:
raise_none_error("mean")
sd = self.cast(
sd, self.parameter_type) if sd is not None else self._sd_value
if sd is None:
raise_none_error("sd")
batch_shape = self.shape(mean + sd)
origin_shape = shape + batch_shape
if origin_shape == ():
sample_shape = (1, )
else:
sample_shape = origin_shape
sample_norm = C.normal(sample_shape, mean, sd, self.seed)
value = self.cast(sample_norm, self.dtype)
if origin_shape == ():
value = self.squeeze(value)
return value
def construct(self, empirical_fraction, norm_bound):
"""
Update value of norm_bound.
Args:
empirical_fraction(Tensor): empirical fraction of samples with the
value at most `target_unclipped_quantile`.
norm_bound(Tensor): Clipping bound for the l2 norm of the gradients.
Returns:
Tensor, generated noise with shape like given gradients.
"""
fraction_noise = normal((1,), self._zero, self._fraction_stddev, self._seed)
empirical_fraction = self._add(empirical_fraction, fraction_noise)
if self._decay_policy == 'Linear':
grad_clip = self._sub(empirical_fraction,
self._target_unclipped_quantile)
next_norm_bound = self._sub(norm_bound,
self._mul(self._learning_rate, grad_clip))
# decay_policy == 'Geometric'
else:
grad_clip = self._sub(empirical_fraction,
self._target_unclipped_quantile)
grad_clip = self._exp(self._mul(-self._learning_rate, grad_clip))
next_norm_bound = self._mul(norm_bound, grad_clip)
return next_norm_bound
def _sample(self, name, shape=(), probs=None):
"""
Sampling.
Args:
name (str): name of the function. Should always be 'sample' when passed in from construct.
shape (tuple): shape of the sample. Default: ().
probs (Tensor): probs1 of the samples. Default: self._probs.
Returns:
Tensor, shape is shape + batch_shape.
"""
if name == 'sample':
probs1 = self._probs if probs is None else probs
batch_shape = self.shape(probs1)
sample_shape = shape + batch_shape
mean_zero = self.const(0.0)
sd_one = self.const(1.0)
sqrt_two = self.sqrt(self.const(2.0))
sample_norm = C.normal(sample_shape, mean_zero, sd_one, self.seed)
sample_uniform = 0.5 * (1 + self.erf(self.realdiv(sample_norm, sqrt_two)))
sample = self.less(sample_uniform, probs1)
sample = self.cast(sample, self._dtype)
return sample
return None
def construct(self, gradients):
"""
Generated Gaussian noise.
Args:
gradients(Tensor): The gradients.
Returns:
Tensor, generated noise with shape like given gradients.
"""
shape = P.Shape()(gradients)
stddev = P.Mul()(self._norm_bound, self._initial_noise_multiplier)
noise = normal(shape, self._mean, stddev, self._seed)
return noise
def _sample(self, shape=(), mean=None, sd=None):
"""
Sampling.
Args:
shape (tuple): shape of the sample. Default: ().
mean (Tensor): mean of the samples. Default: self._mean_value.
sd (Tensor): standard deviation of the samples. Default: self._sd_value.
Returns:
Tensor, shape is shape + batch_shape.
"""
mean = self._mean_value if mean is None else mean
sd = self._sd_value if sd is None else sd
batch_shape = self.shape(self.zeroslike(mean) + self.zeroslike(sd))
sample_shape = shape + batch_shape
sample_norm = C.normal(sample_shape, mean, sd, self.seed)
return sample_norm
def _sample(self, name, shape=(), mean=None, sd=None):
"""
Sampling.
Args:
name (str): name of the function. Should always be 'sample' when passed in from construct.
shape (tuple): shape of the sample. Default: ().
mean (Tensor): mean of the samples. Default: self._mean_value.
sd (Tensor): standard deviation of the samples. Default: self._sd_value.
Returns:
Tensor, shape is shape + batch_shape.
"""
if name == 'sample':
mean = self._mean_value if mean is None else mean
sd = self._sd_value if sd is None else sd
batch_shape = self.shape(self.add(self.zeroslike(mean), self.zeroslike(sd)))
sample_shape = shape + batch_shape
mean_zero = self.const(0.0)
sd_one = self.const(1.0)
sample_norm = C.normal(sample_shape, mean_zero, sd_one, self.seed)
sample = self.add(mean, self.mul(sample_norm, sd))
return sample
return None
def construct(self, mean, stddev):
return C.normal(self.shape, mean, stddev, self.seed)
def construct(self, mean, stddev):
s1 = C.normal(self.shape, mean, stddev, self.seed)
s2 = C.normal(self.shape, mean, stddev, self.seed)
s3 = C.normal(self.shape, mean, stddev, self.seed)
return s1, s2, s3
def construct(self, mean, stddev):
C.set_seed(20)
return C.normal(self.shape, mean, stddev, self.seed)
