def mu_law_encode(x: Tensor, mu: int = 256, quantized: bool = True) -> Tensor:
"""Mu-law encoding.
Compute the mu-law decoding given an input code.
When quantized is True, the result will be converted to
integer in range [0,mu-1]. Otherwise, the resulting signal
is in range [-1,1]
Parameters:
x(Tensor): the input tensor of arbitrary shape to be encoded.
mu(int): the maximum value (depth) of encoded signal. The signal will be
clip to be in range [0,mu-1].
quantized(bool): indicate whether the signal will quantized to integers.
Examples:
.. code-block:: python
import paddle
import paddleaudio.functional as F
F.mu_law_encode(paddle.randn((2, 8)))
>> Tensor(shape=[2, 8], dtype=int32, place=CUDAPlace(0), stop_gradient=True,
[[0, 5, 30, 255, 255, 255, 12, 13],
[0, 241, 8, 243, 7, 35, 84, 228]])
Reference:
https://en.wikipedia.org/wiki/%CE%9C-law_algorithm
"""
mu = mu - 1
y = paddle.sign(x) * paddle.log1p(mu * paddle.abs(x)) / math.log1p(mu)
if quantized:
y = (y + 1) / 2 * mu + 0.5 # convert to [0 , mu-1]
y = paddle.clip(y, min=0, max=mu).astype('int32')
return y
def _probs_to_logits(self, probs, is_binary=False):
r"""
Converts probabilities into logits. For the binary, probs denotes the
probability of occurrence of the event indexed by `1`. For the
multi-dimensional, values of last axis denote the probabilities of
occurrence of each of the events.
"""
return (paddle.log(probs) - paddle.log1p(-probs)) \
if is_binary else paddle.log(probs)
def _binomial_logpmf(self, count, value):
logits = self._probs_to_logits(self.probs, is_binary=True)
factor_n = paddle.lgamma(count + 1)
factor_k = paddle.lgamma(value + 1)
factor_nmk = paddle.lgamma(count - value + 1)
norm = (count * _clip_by_zero(logits) +
count * paddle.log1p(paddle.exp(-paddle.abs(logits))) -
factor_n)
return value * logits - factor_k - factor_nmk - norm
def __init__(self, range_max, n_sample):
with paddle.no_grad():
self.range_max = range_max
log_indices = paddle.log(
paddle.arange(1., range_max + 2., 1., dtype=global_dtype))
self.dist = (log_indices[1:] - log_indices[:-1]) / log_indices[-1]
self.log_q = paddle.cast(paddle.log(
paddle.exp(-(
paddle.log1p(-paddle.cast(self.dist, dtype=global_dtype)) *
2 * n_sample)) - 1),
dtype=global_dtype)
self.n_sample = n_sample
def gwd_loss(self,
pred,
target,
fun='log',
tau=1.0,
alpha=1.0,
normalize=False):
xy_p, R_p, S_p = self.xywhr2xyrs(pred)
xy_t, R_t, S_t = self.xywhr2xyrs(target)
xy_distance = (xy_p - xy_t).square().sum(axis=-1)
Sigma_p = R_p.matmul(S_p.square()).matmul(R_p.transpose([0, 2, 1]))
Sigma_t = R_t.matmul(S_t.square()).matmul(R_t.transpose([0, 2, 1]))
whr_distance = paddle.diagonal(
S_p, axis1=-2, axis2=-1).square().sum(axis=-1)
whr_distance = whr_distance + paddle.diagonal(
S_t, axis1=-2, axis2=-1).square().sum(axis=-1)
_t = Sigma_p.matmul(Sigma_t)
_t_tr = paddle.diagonal(_t, axis1=-2, axis2=-1).sum(axis=-1)
_t_det_sqrt = paddle.diagonal(S_p, axis1=-2, axis2=-1).prod(axis=-1)
_t_det_sqrt = _t_det_sqrt * paddle.diagonal(
S_t, axis1=-2, axis2=-1).prod(axis=-1)
whr_distance = whr_distance + (-2) * (
(_t_tr + 2 * _t_det_sqrt).clip(0).sqrt())
distance = (xy_distance + alpha * alpha * whr_distance).clip(0)
if normalize:
wh_p = pred[..., 2:4].clip(min=1e-7, max=1e7)
wh_t = target[..., 2:4].clip(min=1e-7, max=1e7)
scale = ((wh_p.log() + wh_t.log()).sum(dim=-1) / 4).exp()
distance = distance / scale
if fun == 'log':
distance = paddle.log1p(distance)
if tau >= 1.0:
return 1 - 1 / (tau + distance)
return distance
def __init__(self,
num_classes,
n_sample,
unique=True,
remove_accidental_hits=True,
subtract_log_q=True,
num_true=1,
batch_size=None):
super(Mind_SampledSoftmaxLoss_Layer, self).__init__()
self.range_max = num_classes
self.n_sample = n_sample
self.unique = unique
self.remove_accidental_hits = remove_accidental_hits
self.subtract_log_q = subtract_log_q
self.num_true = num_true
self.prob = np.array([0.0] * self.range_max)
self.batch_size = batch_size
for i in range(1, self.range_max):
self.prob[i] = (np.log(i+2) - np.log(i+1)) / \
np.log(self.range_max + 1)
self.new_prob = paddle.assign(self.prob.astype("float32"))
self.log_q = paddle.log(-(paddle.exp(
(-paddle.log1p(self.new_prob) * 2 * n_sample)) - 1.0))
def _sigmoid_cross_entropy_with_logits(logits, labels): # to be compatible with tensorflow, we don't use ignore_idx loss = paddle.clip(logits, min=0) - logits * labels.astype(logits.dtype) loss += paddle.log1p(paddle.exp(-paddle.abs(logits))) return loss
def logit_transform(image, lam=1e-6):
image = lam + (1 - 2 * lam) * image
return paddle.log(image) - paddle.log1p(-image)
def forward(self, inputs):
"""
forward
"""
x = paddle.log1p(inputs)
return x
