def propose(self):
if self._dist == 'RoundedNormal':
self.parameter.value = int(
round(rnormal(self.parameter.value, self.proposal_sig)))
# Default to normal random-walk proposal
else:
self.parameter.value = rnormal(self.parameter.value,
self.proposal_sig)
def propose(self):
# Propose new values using normal distribution
if self.proposal_distribution == "Normal":
# New normal deviate, centred on current value
new_val = rnormal(self.stochastic.value, self.adaptive_scale_factor * self.proposal_sd)
# Round before setting proposed value
self.stochastic.value = round_array(new_val)
elif self.proposal_distribution == "Poisson":
k = shape(self.stochastic.value)
# Add or subtract (equal probability) Poisson sample
new_val = self.stochastic.value + rpoisson(self.adaptive_scale_factor * self.proposal_sd) * (-ones(k))**(random(k)>0.5)
if self._positive:
# Enforce positive values
self.stochastic.value = abs(new_val)
else:
self.stochastic.value = new_val
elif self.proposal_distribution == "Prior":
self.stochastic.random()
def propose(self):
# Propose new values using normal distribution
if self.proposal_distribution == "Normal":
# New normal deviate, centred on current value
new_val = rnormal(self.stochastic.value, self.adaptive_scale_factor * self.proposal_sd)
# Round before setting proposed value
self.stochastic.value = round_array(new_val)
elif self.proposal_distribution == "Poisson":
k = shape(self.stochastic.value)
# Add or subtract (equal probability) Poisson sample
new_val = self.stochastic.value + rpoisson(self.adaptive_scale_factor * self.proposal_sd) * (-ones(k))**(random(k)>0.5)
if self._positive:
# Enforce positive values
self.stochastic.value = abs(new_val)
else:
self.stochastic.value = new_val
elif self.proposal_distribution == "Prior":
self.stochastic.random()
def hop(self):
"""Hop proposal kernel"""
if self.verbose>1:
print '\t' + self._id + ' Running Hop proposal kernel'
# Mask for values to move
phi = self.phi
if self._prime:
xp, x = self.values
else:
x, xp = self.values
if self.verbose>1:
print '\t' + 'Current value of x = ' + str(x)
sigma = max(phi*abs(xp - x))/3.0
x = (xp + sigma*rnormal())*phi + x*(phi==False)
if self.verbose>1:
print '\t' + 'Proposed value = ' + str(x)
self.hastings_factor = self._g(x, xp, sigma) - self._g(self.stochastic.value, xp, sigma)
self.stochastic.value = x
def blow(self):
"""Blow proposal kernel"""
if self.verbose>1:
print '\t' + self._id + ' Running Blow proposal kernel'
# Mask for values to move
phi = self.phi
if self._prime:
xp, x = self.values
else:
x, xp = self.values
if self.verbose>1:
print '\t' + 'Current value ' + str(x)
sigma = max(phi*abs(xp - x))
x = x + phi*sigma*rnormal()
if self.verbose>1:
print '\t' + 'Proposed value = ' + str(x)
self.hastings_factor = self._g(x, xp, sigma) - self._g(self.stochastic.value, xp, sigma)
self.stochastic.value = x
def propose(self):
"""
This method is called by step() to generate proposed values
if self.proposal_distribution is "Normal" (i.e. no proposal specified).
"""
if self.proposal_distribution == "Normal":
self.stochastic.value = rnormal(self.stochastic.value, self.adaptive_scale_factor * self.proposal_sd)
elif self.proposal_distribution == "Prior":
self.stochastic.random()
def propose(self):
"""
This method is called by step() to generate proposed values
if self.proposal_distribution is "Normal" (i.e. no proposal specified).
"""
if self.proposal_distribution == "Normal":
self.stochastic.value = rnormal(self.stochastic.value, self.adaptive_scale_factor * self.proposal_sd)
elif self.proposal_distribution == "Prior":
self.stochastic.random()
def random(self, size=None):
mu = self.mu
lam = self.lam
alpha = self.alpha
v = rnormal(loc=0., scale=1., size=size)**2
x = mu + (mu**2) * v / (2. * lam) - mu / (
2. * lam) * sqrt(4. * mu * lam * v + (mu * v)**2)
z = runiform(low=0., high=1., size=size)
# i = z > mu / (mu + x)
# x[i] = (mu**2) / x[i]
i = floor(z - mu / (mu + x)) * 2 + 1
x = (x**-i) * (mu**(i + 1))
return x + alpha
def propose(self):
"""
Proposals for positive definite matrix using random walk deviations on the Cholesky
factor of the current value.
"""
# Locally store size of matrix
dims = self.stochastic.value.shape
# Add normal deviate to value and symmetrize
dev = rnormal(0, self.adaptive_scale_factor * self.proposal_sd, size=dims)
symmetrize(dev)
# Replace
self.stochastic.value = dev + self.stochastic.value
def propose(self):
"""
Proposals for positive definite matrix using random walk deviations on the Cholesky
factor of the current value.
"""
# Locally store size of matrix
dims = self.stochastic.value.shape
# Add normal deviate to value and symmetrize
dev = rnormal(0, self.adaptive_scale_factor * self.proposal_sd, size=dims)
symmetrize(dev)
# Replace
self.stochastic.value = dev + self.stochastic.value
def random(self, size=None):
u = runiform(low=0., high=1., size=size)
n = rnormal(self.mu, self.sigma, size=size)
return n - self.nu * log(u)
def cross_render_sample_jitter(
target_sample,
source_sample,
grain_duration=0.01,
grain_dur_jitter=0.0,
grain_interval_jitter=0.0,
grain_jitter=0.0, # proportional offset grain centers from each other
density=4.0,
verbose=0,
code_size=4,
hop_length=1024,
frame_length=4096,
delay_step_size=8,
n_delays=128,
source_anchor='left', # or 'center', not actually implemented
dest_anchor='center', # or 'left'
window='hann',
n_workers=1, # parallelism
chunk_size=None, # paralleism sync
seamless=False,
seed=None,
random_flip=True,
**kwargs):
"""
random everything; but all grains from a single fit share the same
target location.
When anchor is 'center', grain locations are centres.
"""
rseed(seed)
if chunk_size is None:
chunk_size = n_workers * 4
output_sample = sample.zero_sample(duration=target_sample.duration(),
sr=target_sample.sr,
stem=source_sample.stem,
parent_path=target_sample.parent_path)
output_sample.append_stem('_x_')
output_sample.append_stem(target_sample.stem)
kwarg_fn = timed_stream.DictStream(kwargs)
inter_onset_period = grain_duration / density
dest_high = target_sample.duration()
approx_last_step = dest_high / inter_onset_period
if verbose >= 10:
print("gp", inter_onset_period, density, grain_duration, code_size)
print('dest_high', dest_high)
if verbose >= 2:
print('approx_last_step', approx_last_step)
param_list = [
(step, target_t, kwarg_fn(target_t)) for step, target_t in enumerate(
pattern_iter.gamma_renewal_proc_scale(interval=inter_onset_period,
var=inter_onset_period *
grain_interval_jitter**2,
high=dest_high,
verbose=verbose))
]
# I chunk this for intermediate results:
# https://stackoverflow.com/a/43085080
with Parallel(n_jobs=n_workers, verbose=verbose * 3) as pool:
step = 0
target_t_prev = 0.0
try:
for chunk in chunker(iter(param_list), chunk_size):
for local_step, match in enumerate(
pool([
delayed(match_from_sample)(
target_sample=target_sample,
source_sample=source_sample,
target_t=target_t,
verbose=verbose,
hop_length=hop_length,
frame_length=frame_length,
delay_step_size=delay_step_size,
n_delays=n_delays,
anchor=source_anchor,
window=window,
code_size=code_size,
**kw) for step, target_t, kw in chunk
])):
grains = match.grains()
ac_gains = match.gains()
rates = match.rates()
# if verbose >= 2:
# elapsed = time() - start_time()
# print(
# "step", step+1,
# "/", approx_last_step, ",",
# elapsed, "/",
# elapsed * (step+1)/approx_last_step)
target_t = match.target_t
target_t_inc = target_t - target_t_prev
target_t_prev = target_t
if verbose >= 3:
print(
"step", step, local_step,
't {:.5f}+{:.5f}'.format(target_t, target_t_inc),
'loss {:.5f}*{:.5f}'.format(
match.loss, match.scale))
if verbose >= 16:
print("match ", match)
for grain, ac_gain, rate in zip(grains, ac_gains, rates):
if ac_gain == 0.0:
continue
gain = sqrt(abs(ac_gain * rate))
if random_flip:
gain = gain * (2 * randint(2) - 1)
dest_t = match.target_t + np.asscalar(
rnormal(loc=0.0, scale=grain_jitter, size=1), )
if seamless:
base_duration = target_t_inc
else:
base_duration = inter_onset_period
this_grain_duration = min(
gamma_v(base_duration * density, grain_dur_jitter *
base_duration * density), 5.0)
if verbose >= 6:
print(
"grain", 'dest {:.5f}'.format(dest_t),
'duration {:.5f}*{:.5f}'.format(
this_grain_duration,
this_grain_duration / target_t_inc))
if verbose >= 19:
print("is", grain, gain, rate)
sample.overdub_t(
dest_sample=output_sample,
source_sample=grain.sample,
dest_t=dest_t,
source_t=grain.t,
duration=this_grain_duration,
rate=rate,
mul=gain,
window=window,
)
step += 1
except StopIteration as e:
print('some iterator broke {}'.format(e))
traceback.print_tb(e.__traceback__)
except KeyboardInterrupt:
print('stopping early')
return output_sample
def cross_render_sample_adaptive(
target_sample,
source_sample,
density=2.0,
verbose=0,
code_size=4,
hop_length=1024,
frame_length=4096,
delay_step_size=8,
grain_jitter=0.0, # proportional offset grain centers from each other
n_delays=128,
source_anchor='center', # or 'left', maybe not actually implemented?
dest_anchor='center', # or 'left'
analysis_window='cosine',
synthesis_window='hann',
seed=None,
random_flip=True,
match_rtol=0.01, # search early stopping parameter; not implemented
progress=False,
**kwargs):
"""
Fit one new match at a time.
"""
rseed(seed)
sr = target_sample.sr
output_sample = sample.zero_sample(duration=target_sample.duration(),
sr=sr,
stem=source_sample.stem,
parent_path=target_sample.parent_path)
output_sample.append_stem('_x_')
output_sample.append_stem(target_sample.stem)
kwarg_fn = timed_stream.DictStream(kwargs)
grain_duration = frame_length / sr
inter_onset_period = grain_duration / density
dest_high = target_sample.duration()
approx_last_step = dest_high / inter_onset_period
if verbose >= 10:
print("gp", inter_onset_period, density, grain_duration, code_size)
print('dest_high', dest_high)
if verbose >= 2:
print('approx_last_step', approx_last_step)
step = 0
target_t_prev = 0.0
target_t = 0.0
target_i_prev = 0
target_i = 0
early_stop = False
while target_i < output_sample.end:
try:
target_i += hop_length
target_t = target_i / sr
kw = kwarg_fn(target_t)
match = match_from_sample(target_sample=target_sample,
source_sample=source_sample,
target_t=target_t,
hop_length=hop_length,
frame_length=frame_length,
delay_step_size=delay_step_size,
n_delays=n_delays,
anchor=source_anchor,
window=analysis_window,
code_size=code_size,
match_rtol=match_rtol,
verbose=verbose,
**kw)
grains = match.grains()
ac_gains = match.gains()
rates = match.rates()
# if verbose >= 2:
# elapsed = time() - start_time()
# print(
# "step", step+1,
# "/", approx_last_step, ",",
# elapsed, "/",
# elapsed * (step+1)/approx_last_step)
target_t_inc = target_t - target_t_prev
target_t_prev = target_t
target_i_prev = target_i
if verbose >= 3:
print("step", step,
't {:.5f}+{:.5f}'.format(target_t, target_t_inc),
'loss {:.5f}*{:.5f}'.format(match.loss, match.scale))
if verbose >= 16:
print("match ", match)
for g_i, grain, ac_gain, rate in zip(range(len(grains)), grains,
ac_gains, rates):
if ac_gain == 0.0:
continue
gain = sqrt(abs(ac_gain * rate))
if random_flip:
gain = gain * (2 * randint(2) - 1)
dest_t = target_t + np.asscalar(
rnormal(
loc=0.0, scale=grain_duration * grain_jitter, size=1))
this_grain_duration = target_t_inc * density
if verbose >= 8:
print(
"subgrain", g_i, 'dest {:.5f}'.format(dest_t),
'gain {:.5f}'.format(gain), 'rate {:.5f}'.format(rate),
'duration {:.5f}*{:.5f}'.format(
this_grain_duration,
this_grain_duration / target_t_inc))
if verbose >= 19:
print("is", grain, gain, rate)
landscape = np.copy(
output_sample.get_audio_data_t(
dest_t, duration=this_grain_duration, anchor="center"))
sample.overdub_t(
dest_sample=output_sample,
source_sample=grain.sample,
dest_t=dest_t,
source_t=grain.t,
duration=this_grain_duration,
rate=rate,
mul=gain,
window=synthesis_window,
source_anchor=source_anchor,
dest_anchor=dest_anchor,
verbose=verbose,
)
if verbose >= 6:
y_so_far = output_sample.get_audio_data_t(
0, target_t + this_grain_duration)
print(
"glitch",
np.percentile(np.abs(np.gradient(y_so_far)),
[50, 95, 99, 99.5]))
step += 1
if progress:
print(step, '/', approx_last_step)
except StopIteration as e:
print('some iterator broke {}'.format(e))
traceback.print_tb(e.__traceback__)
except KeyboardInterrupt:
print('stopping early at {} -- {}'.format(step, target_t))
early_stop = True
break
return output_sample
from numpy.random import normal as rnormal from numpy import array, arange, sin, cos, pi, identity, matmul, zeros, cov from numpy.linalg import inv from mpl_toolkits.mplot3d import Axes3D import matplotlib.pyplot as plt del_t = 0.1 t = arange(0, 1000, del_t) x = .002*t + rnormal(0,.5,10000) y = .0000003*(t-500)**3 + rnormal(0,.5,10000) z = 4*sin(pi*t/1000) + rnormal(0,.5, 10000) xt = .002*t yt = .0000003*(t-500)**3 zt = 4*sin(pi*t/1000) v_x = array([(x[i] - x[i-1])/del_t for i in xrange(1,len(x))]+[(x[-1]-x[-2])/del_t]) v_y = array([(y[i] - y[i-1])/del_t for i in xrange(1,len(y))]+[(y[-1]-y[-2])/del_t]) v_z = array([(z[i] - z[i-1])/del_t for i in xrange(1,len(z))]+[(z[-1]-z[-2])/del_t]) X = array([x,v_x]) Y = array([y,v_y]) Z = array([z,v_z]) F = array([[1,del_t],[0,1]]) G = array([[del_t**2/2], [del_t]]) Q = matmul(G, G.transpose())*.5**2 Px = identity(2)*5
