def xlog1py(x, y):
x, y = _promote_args_inexact("xlog1py", x, y)
x_ok = x != 0.
safe_x = jnp.where(x_ok, x, 1.)
safe_y = jnp.where(x_ok, y, 1.)
return jnp.where(x_ok, lax.mul(safe_x, lax.log1p(safe_y)),
jnp.zeros_like(x))
def _log_ndtr_asymptotic_series(x, series_order):
"""Calculates the asymptotic series used in log_ndtr."""
dtype = lax.dtype(x).type
if series_order <= 0:
return np.array(1, dtype)
x_2 = lax.square(x)
even_sum = jnp.zeros_like(x)
odd_sum = jnp.zeros_like(x)
x_2n = x_2 # Start with x^{2*1} = x^{2*n} with n = 1.
for n in range(1, series_order + 1):
y = np.array(_double_factorial(2 * n - 1), dtype) / x_2n
if n % 2:
odd_sum += y
else:
even_sum += y
x_2n *= x_2
return dtype(1.) + even_sum - odd_sum
def _slogdet_jvp(primals, tangents):
x, = primals
g, = tangents
sign, ans = slogdet(x)
ans_dot = jnp.trace(solve(x, g), axis1=-1, axis2=-2)
if jnp.issubdtype(jnp._dtype(x), jnp.complexfloating):
sign_dot = (ans_dot - np.real(ans_dot)) * sign
ans_dot = np.real(ans_dot)
else:
sign_dot = jnp.zeros_like(sign)
return (sign, ans), (sign_dot, ans_dot)
def _cofactor_solve(a, b):
"""Equivalent to det(a)*solve(a, b) for nonsingular mat.
Intermediate function used for jvp and vjp of det.
This function borrows heavily from jax.numpy.linalg.solve and
jax.numpy.linalg.slogdet to compute the gradient of the determinant
in a way that is well defined even for low rank matrices.
This function handles two different cases:
* rank(a) == n or n-1
* rank(a) < n-1
For rank n-1 matrices, the gradient of the determinant is a rank 1 matrix.
Rather than computing det(a)*solve(a, b), which would return NaN, we work
directly with the LU decomposition. If a = p @ l @ u, then
det(a)*solve(a, b) =
prod(diag(u)) * u^-1 @ l^-1 @ p^-1 b =
prod(diag(u)) * triangular_solve(u, solve(p @ l, b))
If a is rank n-1, then the lower right corner of u will be zero and the
triangular_solve will fail.
Let x = solve(p @ l, b) and y = det(a)*solve(a, b).
Then y_{n}
x_{n} / u_{nn} * prod_{i=1...n}(u_{ii}) =
x_{n} * prod_{i=1...n-1}(u_{ii})
So by replacing the lower-right corner of u with prod_{i=1...n-1}(u_{ii})^-1
we can avoid the triangular_solve failing.
To correctly compute the rest of y_{i} for i != n, we simply multiply
x_{i} by det(a) for all i != n, which will be zero if rank(a) = n-1.
For the second case, a check is done on the matrix to see if `solve`
returns NaN or Inf, and gives a matrix of zeros as a result, as the
gradient of the determinant of a matrix with rank less than n-1 is 0.
This will still return the correct value for rank n-1 matrices, as the check
is applied *after* the lower right corner of u has been updated.
Args:
a: A square matrix or batch of matrices, possibly singular.
b: A matrix, or batch of matrices of the same dimension as a.
Returns:
det(a) and cofactor(a)^T*b, aka adjugate(a)*b
"""
a = _promote_arg_dtypes(jnp.asarray(a))
b = _promote_arg_dtypes(jnp.asarray(b))
a_shape = jnp.shape(a)
b_shape = jnp.shape(b)
a_ndims = len(a_shape)
if not (a_ndims >= 2 and a_shape[-1] == a_shape[-2]
and b_shape[-2:] == a_shape[-2:]):
msg = ("The arguments to _cofactor_solve must have shapes "
"a=[..., m, m] and b=[..., m, m]; got a={} and b={}")
raise ValueError(msg.format(a_shape, b_shape))
if a_shape[-1] == 1:
return a[..., 0, 0], b
# lu contains u in the upper triangular matrix and l in the strict lower
# triangular matrix.
# The diagonal of l is set to ones without loss of generality.
lu, pivots, permutation = lax_linalg.lu(a)
dtype = lax.dtype(a)
batch_dims = lax.broadcast_shapes(lu.shape[:-2], b.shape[:-2])
x = jnp.broadcast_to(b, batch_dims + b.shape[-2:])
lu = jnp.broadcast_to(lu, batch_dims + lu.shape[-2:])
# Compute (partial) determinant, ignoring last diagonal of LU
diag = jnp.diagonal(lu, axis1=-2, axis2=-1)
parity = jnp.count_nonzero(pivots != jnp.arange(a_shape[-1]), axis=-1)
sign = jnp.asarray(-2 * (parity % 2) + 1, dtype=dtype)
# partial_det[:, -1] contains the full determinant and
# partial_det[:, -2] contains det(u) / u_{nn}.
partial_det = jnp.cumprod(diag, axis=-1) * sign[..., None]
lu = lu.at[..., -1, -1].set(1.0 / partial_det[..., -2])
permutation = jnp.broadcast_to(permutation, batch_dims + (a_shape[-1], ))
iotas = jnp.ix_(*(lax.iota(jnp.int32, b) for b in batch_dims + (1, )))
# filter out any matrices that are not full rank
d = jnp.ones(x.shape[:-1], x.dtype)
d = lax_linalg.triangular_solve(lu, d, left_side=True, lower=False)
d = jnp.any(jnp.logical_or(jnp.isnan(d), jnp.isinf(d)), axis=-1)
d = jnp.tile(d[..., None, None], d.ndim * (1, ) + x.shape[-2:])
x = jnp.where(d, jnp.zeros_like(x), x) # first filter
x = x[iotas[:-1] + (permutation, slice(None))]
x = lax_linalg.triangular_solve(lu,
x,
left_side=True,
lower=True,
unit_diagonal=True)
x = jnp.concatenate(
(x[..., :-1, :] * partial_det[..., -1, None, None], x[..., -1:, :]),
axis=-2)
x = lax_linalg.triangular_solve(lu, x, left_side=True, lower=False)
x = jnp.where(d, jnp.zeros_like(x), x) # second filter
return partial_det[..., -1], x
def _create_polynomial(var, coeffs):
"""Compute n_th order polynomial via Horner's method."""
coeffs = np.array(coeffs, dtype)
if not coeffs.size:
return jnp.zeros_like(var)
return coeffs[0] + _create_polynomial(var, coeffs[1:]) * var
def _spectral_helper(x,
y,
fs=1.0,
window='hann',
nperseg=None,
noverlap=None,
nfft=None,
detrend_type='constant',
return_onesided=True,
scaling='density',
axis=-1,
mode='psd',
boundary=None,
padded=False):
"""LAX-backend implementation of `scipy.signal._spectral_helper`.
Unlike the original helper function, `y` can be None for explicitly
indicating auto-spectral (non cross-spectral) computation. In addition to
this, `detrend` argument is renamed to `detrend_type` for avoiding internal
name overlap.
"""
if mode not in ('psd', 'stft'):
raise ValueError(f"Unknown value for mode {mode}, "
"must be one of: ('psd', 'stft')")
def make_pad(mode, **kwargs):
def pad(x, n, axis=-1):
pad_width = [(0, 0) for unused_n in range(x.ndim)]
pad_width[axis] = (n, n)
return jnp.pad(x, pad_width, mode, **kwargs)
return pad
boundary_funcs = {
'even': make_pad('reflect'),
'odd': odd_ext,
'constant': make_pad('edge'),
'zeros': make_pad('constant', constant_values=0.0),
None: lambda x, *args, **kwargs: x
}
# Check/ normalize inputs
if boundary not in boundary_funcs:
raise ValueError(f"Unknown boundary option '{boundary}', "
f"must be one of: {list(boundary_funcs.keys())}")
axis = jax.core.concrete_or_error(operator.index, axis,
"axis of windowed-FFT")
axis = canonicalize_axis(axis, x.ndim)
if y is None:
_check_arraylike('spectral_helper', x)
x, = _promote_dtypes_inexact(x)
outershape = tuple_delete(x.shape, axis)
else:
if mode != 'psd':
raise ValueError(
"two-argument mode is available only when mode=='psd'")
_check_arraylike('spectral_helper', x, y)
x, y = _promote_dtypes_inexact(x, y)
if x.ndim != y.ndim:
raise ValueError(
"two-arguments must have the same rank ({x.ndim} vs {y.ndim})."
)
# Check if we can broadcast the outer axes together
try:
outershape = jnp.broadcast_shapes(tuple_delete(x.shape, axis),
tuple_delete(y.shape, axis))
except ValueError as err:
raise ValueError('x and y cannot be broadcast together.') from err
result_dtype = dtypes._to_complex_dtype(x.dtype)
freq_dtype = np.finfo(result_dtype).dtype
if nperseg is not None: # if specified by user
nperseg = jax.core.concrete_or_error(int, nperseg,
"nperseg of windowed-FFT")
if nperseg < 1:
raise ValueError('nperseg must be a positive integer')
# parse window; if array like, then set nperseg = win.shape
win, nperseg = signal_helper._triage_segments(window,
nperseg,
input_length=x.shape[axis],
dtype=result_dtype)
if noverlap is None:
noverlap = nperseg // 2
else:
noverlap = jax.core.concrete_or_error(int, noverlap,
"noverlap of windowed-FFT")
if nfft is None:
nfft = nperseg
else:
nfft = jax.core.concrete_or_error(int, nfft, "nfft of windowed-FFT")
# Special cases for size == 0
if y is None:
if x.size == 0:
return jnp.zeros(x.shape, freq_dtype), jnp.zeros(
x.shape, freq_dtype), jnp.zeros(x.shape, result_dtype)
else:
if x.size == 0 or y.size == 0:
shape = tuple_insert(outershape,
min([x.shape[axis], y.shape[axis]]), axis)
return jnp.zeros(shape, freq_dtype), jnp.zeros(
shape, freq_dtype), jnp.zeros(shape, result_dtype)
# Move time-axis to the end
x = jnp.moveaxis(x, axis, -1)
if y is not None and y.ndim > 1:
y = jnp.moveaxis(y, axis, -1)
# Check if x and y are the same length, zero-pad if necessary
if y is not None and x.shape[-1] != y.shape[-1]:
if x.shape[-1] < y.shape[-1]:
pad_shape = list(x.shape)
pad_shape[-1] = y.shape[-1] - x.shape[-1]
x = jnp.concatenate((x, jnp.zeros_like(x, shape=pad_shape)), -1)
else:
pad_shape = list(y.shape)
pad_shape[-1] = x.shape[-1] - y.shape[-1]
y = jnp.concatenate((y, jnp.zeros_like(x, shape=pad_shape)), -1)
if nfft < nperseg:
raise ValueError('nfft must be greater than or equal to nperseg.')
if noverlap >= nperseg:
raise ValueError('noverlap must be less than nperseg.')
nstep = nperseg - noverlap
# Apply paddings
if boundary is not None:
ext_func = boundary_funcs[boundary]
x = ext_func(x, nperseg // 2, axis=-1)
if y is not None:
y = ext_func(y, nperseg // 2, axis=-1)
if padded:
# Pad to integer number of windowed segments
# I.e make x.shape[-1] = nperseg + (nseg-1)*nstep, with integer nseg
nadd = (-(x.shape[-1] - nperseg) % nstep) % nperseg
x = jnp.concatenate(
(x, jnp.zeros_like(x, shape=(*x.shape[:-1], nadd))), axis=-1)
if y is not None:
y = jnp.concatenate(
(y, jnp.zeros_like(x, shape=(*y.shape[:-1], nadd))), axis=-1)
# Handle detrending and window functions
if not detrend_type:
detrend_func = lambda d: d
elif not callable(detrend_type):
detrend_func = partial(detrend, type=detrend_type, axis=-1)
elif axis != -1:
# Wrap this function so that it receives a shape that it could
# reasonably expect to receive.
def detrend_func(d):
d = jnp.moveaxis(d, axis, -1)
d = detrend_type(d)
return jnp.moveaxis(d, -1, axis)
else:
detrend_func = detrend_type
# Determine scale
if scaling == 'density':
scale = 1.0 / (fs * (win * win).sum())
elif scaling == 'spectrum':
scale = 1.0 / win.sum()**2
else:
raise ValueError(f'Unknown scaling: {scaling}')
if mode == 'stft':
scale = jnp.sqrt(scale)
# Determine onesided/ two-sided
if return_onesided:
sides = 'onesided'
if jnp.iscomplexobj(x) or jnp.iscomplexobj(y):
sides = 'twosided'
warnings.warn('Input data is complex, switching to '
'return_onesided=False')
else:
sides = 'twosided'
if sides == 'twosided':
freqs = jax.numpy.fft.fftfreq(nfft, 1 / fs).astype(freq_dtype)
elif sides == 'onesided':
freqs = jax.numpy.fft.rfftfreq(nfft, 1 / fs).astype(freq_dtype)
# Perform the windowed FFTs
result = _fft_helper(x.astype(result_dtype), win, detrend_func, nperseg,
noverlap, nfft, sides)
if y is not None:
# All the same operations on the y data
result_y = _fft_helper(y.astype(result_dtype), win, detrend_func,
nperseg, noverlap, nfft, sides)
result = jnp.conjugate(result) * result_y
elif mode == 'psd':
result = jnp.conjugate(result) * result
result *= scale
if sides == 'onesided' and mode == 'psd':
end = None if nfft % 2 else -1
result = result.at[..., 1:end].mul(2)
time = jnp.arange(nperseg / 2,
x.shape[-1] - nperseg / 2 + 1,
nperseg - noverlap,
dtype=freq_dtype) / fs
if boundary is not None:
time -= (nperseg / 2) / fs
result = result.astype(result_dtype)
# All imaginary parts are zero anyways
if y is None and mode != 'stft':
result = result.real
# Move frequency axis back to axis where the data came from
result = jnp.moveaxis(result, -1, axis)
return freqs, time, result