def __init__(self, *args, **kwargs):
"""A simple matcher that prefers paths where each matched location is as close as possible to the
observed position.
:param avoid_goingback: Change the transition probability to be lower for the direction the path is coming
from.
:param kwargs: Arguments passed to :class:`BaseMatcher`.
"""
if "matching" not in kwargs:
kwargs['matching'] = SimpleMatching
super().__init__(*args, **kwargs)
self.obs_noise_dist = halfnorm(scale=self.obs_noise)
self.obs_noise_dist_ne = halfnorm(scale=self.obs_noise_ne)
# normalize to max 1 to simulate a prob instead of density
self.obs_noise_logint = math.log(self.obs_noise *
math.sqrt(2 * math.pi) / 2)
self.obs_noise_logint_ne = math.log(self.obs_noise_ne *
math.sqrt(2 * math.pi) / 2)
# Transition probability is divided (in logprob_trans) by this factor if we move back on the
# current edge.
self.avoid_goingback = kwargs.get('avoid_goingback', True)
self.gobackonedge_factor_log = math.log(0.99)
# Transition probability is divided (in logprob_trans) by this factor if the next state is
# also the previous state, thus if we go back
self.gobacktoedge_factor_log = math.log(0.5)
# Transition probability is divided (in logprob_trans) by this factor if a transition is made
# This is to try to stay on the same node unless there is a good reason
self.transition_factor = math.log(0.9)
def minimal_priors():
return [
lambda x: halfnorm(scale=1.).logpdf(np.sqrt(np.exp(x))) + x / 2.0 - np.
log(2.0), lambda x: invgamma(a=5.0, scale=1.).logpdf(np.exp(x)) + x,
lambda x: halfnorm(scale=1.).logpdf(np.sqrt(np.exp(x))
) + x / 2.0 - np.log(2.0)
]
def create_priors(
n_parameters: int,
signal_scale: float = 4.0,
lengthscale_lower_bound: float = 0.1,
lengthscale_upper_bound: float = 0.5,
noise_scale: float = 0.0006,
) -> List[Callable[[float], float]]:
"""Create a list of priors to be used for the hyperparameters of the tuning process.
Parameters
----------
n_parameters : int
Number of parameters to be optimized.
signal_scale : float
Prior scale of the signal (standard deviation) which is used to parametrize a
half-normal distribution.
lengthscale_lower_bound : float
Lower bound of the inverse-gamma lengthscale prior. It marks the point at which
1 % of the cumulative density is reached.
lengthscale_upper_bound : float
Upper bound of the inverse-gamma lengthscale prior. It marks the point at which
99 % of the cumulative density is reached.
noise_scale : float
Prior scale of the noise (standard deviation) which is used to parametrize a
half-normal distribution.
Returns
-------
list of callables
List of priors in the following order:
- signal prior
- lengthscale prior (n_parameters times)
- noise prior
"""
if signal_scale <= 0.0:
raise ValueError(
f"The signal scale needs to be strictly positive. Got {signal_scale}."
)
if noise_scale <= 0.0:
raise ValueError(
f"The noise scale needs to be strictly positive. Got {noise_scale}."
)
signal_prior = halfnorm(scale=signal_scale)
lengthscale_prior = make_invgamma_prior(
lower_bound=lengthscale_lower_bound,
upper_bound=lengthscale_upper_bound)
noise_prior = halfnorm(scale=noise_scale)
priors = [
lambda x: signal_prior.logpdf(np.sqrt(np.exp(x))) + x / 2.0 - np.log(
2.0)
]
for _ in range(n_parameters):
priors.append(lambda x: lengthscale_prior.logpdf(np.exp(x)) + x)
priors.append(lambda x: noise_prior.logpdf(np.sqrt(np.exp(x))) + x / 2.0 -
np.log(2.0))
return priors
def test_multiple_basis_and_training_set_sizes(self):
np.random.seed(seed)
Ns = [
50,
100,
] # triaing set sizes
Ms = [
5,
10,
] # basis set size
epsilon = stats.norm(loc=0, scale=0.01)
tfun = lambda x: np.sin(x) + np.cos(2. * x)
init_beta = distribution_wrapper(stats.halfnorm(scale=1),
size=1,
single=True)
init_alphas = distribution_wrapper(stats.halfnorm(scale=1),
single=False)
for N in Ns:
t_est, t_err = [], []
for M in Ms:
x = np.linspace(0, 1, N)
k = M
trafo = FourierFeatures(k=k)
base_trafo = trafo.fit_transform
model_type = RelevanceVectorMachine
model_kwargs = dict(n_iter=50,
verbose=False,
compute_score=True,
init_beta=init_beta,
init_alphas=init_alphas)
runtimes, coefs = repeated_regression(
x,
base_trafo,
model_type,
t=None,
tfun=tfun,
epsilon=epsilon,
model_kwargs=model_kwargs,
Nruns=Nruns,
return_coefs=True)
#print_run_stats(base_trafo,x,runtimes,coefs,Nruns)
t_est.append(runtimes.mean())
t_err.append(runtimes.std(ddof=1) * 2)
print("\ntime for N = {}:".format(N))
for est, err in zip(t_est, t_err):
print(" estimate = {:.4f}s, 2*std = {:.4f}s".format(
est, 2 * err))
def __init__(
self,
n_particles,
prior={
'logk': norm(loc=np.log(1 / 365), scale=2),
's': halfnorm(loc=1, scale=2),
'α': halfnorm(loc=0, scale=3)
}):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {'ϵ': 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def __init__(
self,
n_particles,
prior={
'a': norm(loc=0.01, scale=0.1),
'b': halfnorm(loc=0.001, scale=3),
'α': halfnorm(loc=0, scale=3)
}):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {'ϵ': 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def __init__(
self,
n_particles,
prior={
'gamma_reward': halfnorm(loc=0, scale=10),
'gamma_delay': halfnorm(loc=0, scale=10),
'k': norm(loc=0, scale=2),
'α': halfnorm(loc=0, scale=3)
}):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {'ϵ': 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def test_semimanual_hyperparameters(self):
np.random.seed(seed)
init_beta = stats.halfnorm(scale=1).rvs(size=1)[0]
init_alphas = stats.halfnorm(scale=1).rvs(size=self.X.shape[1])
model = RelevanceVectorMachine(n_iter=50,
verbose=False,
compute_score=True,
init_beta=init_beta,
init_alphas=init_alphas)
model.fit(self.X, t)
y, yerr = model.predict(self.X, return_std=True)
def __init__(
self,
n_particles,
prior={
"logk": norm(loc=np.log(1 / 365), scale=2),
"s": halfnorm(loc=1, scale=2),
"α": halfnorm(loc=0, scale=3),
},
):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {"ϵ": 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def __init__(
self,
n_particles,
prior={
"a": norm(loc=0.01, scale=0.1),
"b": halfnorm(loc=0.001, scale=3),
"α": halfnorm(loc=0, scale=3),
},
):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {"ϵ": 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def test_random_hyperparameters(self):
np.random.seed(seed)
init_beta = distribution_wrapper(stats.halfnorm(scale=1), single=True)
init_alphas = distribution_wrapper(stats.halfnorm(scale=1),
single=False)
model = RelevanceVectorMachine(n_iter=50,
verbose=False,
compute_score=True,
init_beta=init_beta,
init_alphas=init_alphas)
model.fit(self.X, t)
y, yerr = model.predict(self.X, return_std=True)
def __init__(
self,
n_particles,
prior={
"gamma_reward": halfnorm(loc=0, scale=10),
"gamma_delay": halfnorm(loc=0, scale=10),
"k": norm(loc=0, scale=2),
"α": halfnorm(loc=0, scale=3),
},
):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {"ϵ": 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def _recursive_priors(kernel, prior_list):
if hasattr(kernel, "kernel"): # Unary operations
_recursive_priors(kernel.kernel, prior_list)
elif hasattr(kernel, "k1"): # Binary operations
_recursive_priors(kernel.k1, prior_list)
_recursive_priors(kernel.k2, prior_list)
elif hasattr(kernel, "kernels"): # CompoundKernel
for k in kernel.kernels:
_recursive_priors(k, prior_list)
else:
name = type(kernel).__name__
if name in ["ConstantKernel", "WhiteKernel"]:
# We use a half-normal prior distribution on the signal variance and
# noise. The input x is sampled in log-space, which is why the
# change of variables is necessary.
# This prior assumes that the function values are standardized.
# Note, that we do not know the structure of the kernel, which is
# why this is just only a best guess.
prior_list.append(
lambda x: halfnorm(scale=2.0).logpdf(np.sqrt(np.exp(x))) + x /
2.0 - np.log(2.0), )
elif name in ["Matern", "RBF"]:
# Here we apply an inverse gamma distribution to any lengthscale
# parameter we find. We assume the input variables are normalized
# to lie in [0, 1]. The specific values for a and scale were
# obtained by fitting the 1% and 99% quantile to 0.15 and 0.8.
prior_list.append(
lambda x: invgamma(a=8.286, scale=2.4605).logpdf(np.exp(x)) +
x, )
else:
raise NotImplementedError(
f"Unable to guess priors for this kernel: {kernel}.")
def mcmc_regression():
np.random.seed(123) # Uncomment to reproduce the plot exactly
X, Y = generate_data(100)
alpha_dist = beta1_dist = beta2_dist = norm(0, 10)
sigma_dist = halfnorm(0, 1)
dists = (alpha_dist, beta1_dist, beta2_dist, sigma_dist)
nburnin = 50000
niter = 50000 + nburnin
ncomps = 4
chain = np.zeros(shape=(niter, ncomps))
chain[0, :] = np.abs(np.random.normal(size=ncomps))
for i in tqdm(range(niter - 1)):
v = chain[i]
log_posterior_old = loglik(v, X, Y) + logprior(v, dists)
proposal = proposalfunc(v)
log_posterior_new = loglik(proposal, X, Y) + logprior(proposal, dists)
a = min(0.0, log_posterior_new - log_posterior_old)
if np.random.random() < np.exp(a):
chain[i + 1, :] = proposal
else:
chain[i + 1, :] = v
plot(chain, nburnin)
def mcmc_regression():
np.random.seed(123) # Uncomment to reproduce the plot exactly
X, Y = generate_data(100)
alpha_dist = beta1_dist = beta2_dist = norm(0, 10)
sigma_dist = halfnorm(0, 1)
dists = (alpha_dist, beta1_dist, beta2_dist, sigma_dist)
nburnin = 50000
niter = 50000 + nburnin
ncomps = 4
chain = np.zeros(shape=(niter, ncomps))
chain[0, :] = np.abs(np.random.normal(size=ncomps))
for i in tqdm(range(niter - 1)):
v = chain[i]
log_posterior_old = loglik(v, X, Y) + logprior(v, dists)
proposal = proposalfunc(v)
log_posterior_new = loglik(proposal, X, Y) + logprior(proposal, dists)
a = min(0.0, log_posterior_new - log_posterior_old)
if np.random.random() < np.exp(a):
chain[i + 1, :] = proposal
else:
chain[i + 1, :] = v
plot(chain, nburnin)
def sample(self):
self.std += self.dx
# variance increases as t-> infinity
delta = stats.halfnorm(scale=self.std).rvs()
state = self._clip(self.state + delta)
self.state = state
return state
def __init__(
self,
n_particles,
prior={"indiff": uniform(0, 1), "α": halfnorm(loc=0, scale=0.1)},
):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {"ϵ": 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def __init__(
self,
n_particles,
prior={"logk": norm(loc=-4.5, scale=1), "α": halfnorm(loc=0, scale=2)},
):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {"ϵ": 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def filter_events_array(trace_arr, scale=2):
from scipy import stats
filt = stats.halfnorm(loc=0, scale=scale).pdf(np.arange(20))
filtered_arr = np.empty(trace_arr.shape)
for ind_cell in range(trace_arr.shape[0]):
this_trace = trace_arr[ind_cell, :]
this_trace_filtered = np.convolve(this_trace, filt)[:len(this_trace)]
filtered_arr[ind_cell, :] = this_trace_filtered
return filtered_arr
def __init__(self,
n_particles,
prior={
"γ": beta(1, 1),
"α": halfnorm(loc=0, scale=3)
}):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {"ϵ": 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def testHalfNormalQuantile(self):
batch_size = 50
scale = self._rng.rand(batch_size) + 1.0
p = np.linspace(0., 1.0, batch_size).astype(np.float64)
halfnorm = tfd.HalfNormal(scale=scale)
x = halfnorm.quantile(p)
self._testBatchShapes(halfnorm, x)
expected_x = sp_stats.halfnorm(scale=scale).ppf(p)
self.assertAllClose(expected_x, self.evaluate(x), atol=0)
def sampling1(nsamples,
ndim,
scale=1.0,
to_tensor=True,
device="cpu",
**kwargs):
res = scale * spstats.halfnorm().rvs(
(nsamples, 1)) * uniform_sphere(nsamples, ndim)
if to_tensor:
res = torch.tensor(res.astype(np.float32), device=device)
return res
def __init__(
self,
n_particles,
prior={
'm': norm(loc=-2.43, scale=2),
'c': norm(loc=0, scale=100),
'α': halfnorm(loc=0, scale=3)
}):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {'ϵ': 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def __init__(self, n_particles=1000,
prior={'logk': norm(loc=-4.25, scale=1.5),
'α': halfnorm(loc=0, scale=2)}):
self.n_particles = n_particles
self.prior = prior
self.θ_fixed = {'ϵ': 0.01}
# Annoying, this is why they invented probabilistic programming
true_alpha = np.abs(scipy.random.normal(loc=0., scale=2.))
true_logk = scipy.random.normal(loc=-4.25, scale=1.5)
self.θ_true = pd.DataFrame([{'α': true_alpha, 'logk': true_logk}])
self.choiceFunction = CumulativeNormalChoiceFunc
def __init__(
self,
n_particles,
prior={
"m": norm(loc=-2.43, scale=2),
"c": norm(loc=0, scale=100),
"α": halfnorm(loc=0, scale=3),
},
):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {"ϵ": 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def response_xr_filtered(scale):
filt = stats.halfnorm(loc=0, scale=scale).pdf(np.arange(20))
filtered_arr = np.empty(trace_arr.shape)
for ind_cell in range(trace_arr.shape[0]):
this_trace = trace_arr[ind_cell, :]
this_trace_filtered = np.convolve(this_trace, filt)[:len(this_trace)]
filtered_arr[ind_cell, :] = this_trace_filtered
response_xr_filtered = rp.get_response_xr(dataset, filtered_arr,
ophys_timestamps, events,
event_ids, trace_ids,
response_analysis_params)
return response_xr_filtered
def _recursive_priors(kernel, prior_list):
if hasattr(kernel, "kernel"): # Unary operations
_recursive_priors(kernel.kernel, prior_list)
elif hasattr(kernel, "k1"): # Binary operations
_recursive_priors(kernel.k1, prior_list)
_recursive_priors(kernel.k2, prior_list)
elif hasattr(kernel, "kernels"): # CompoundKernel
# It seems that the skopt kernels are not compatible with the
# CompoundKernel. This is therefore not officially supported.
for k in kernel.kernels:
_recursive_priors(k, prior_list)
else:
name = type(kernel).__name__
if name in ["ConstantKernel", "WhiteKernel"]:
if name == "ConstantKernel" and kernel.constant_value_bounds == "fixed":
return
if name == "WhiteKernel" and kernel.noise_level_bounds == "fixed":
return
# We use a half-normal prior distribution on the signal variance and
# noise. The input x is sampled in log-space, which is why the
# change of variables is necessary.
# This prior assumes that the function values are standardized.
# Note, that we do not know the structure of the kernel, which is
# why this is just only a best guess.
prior_list.append(
lambda x: halfnorm(scale=2.0).logpdf(np.sqrt(np.exp(x))) + x /
2.0 - np.log(2.0), )
elif name in ["Matern", "RBF"]:
# Here we apply a round-flat prior distribution to any lengthscale
# parameter we find. We assume the input variables are normalized
# to lie in [0, 1].
# For common optimization problems, we expect the lengthscales to
# lie in the range [0.1, 0.6]. The round-flat prior allows values
# outside the range, if supported by enough datapoints.
if isinstance(kernel.length_scale,
(collections.Sequence, np.ndarray)):
n_priors = len(kernel.length_scale)
else:
n_priors = 1
roundflat = make_roundflat(
lower_bound=0.1,
upper_bound=0.6,
lower_steepness=2.0,
upper_steepness=8.0,
)
for _ in range(n_priors):
prior_list.append(lambda x: roundflat(np.exp(x)) + x)
else:
raise NotImplementedError(
f"Unable to guess priors for this kernel: {kernel}.")
def testHalfNormalSurvivalFunction(self):
batch_size = 50
scale = self._rng.rand(batch_size) + 1.0
x = np.linspace(-8.0, 8.0, batch_size).astype(np.float64)
halfnorm = tfd.HalfNormal(scale=scale)
sf = halfnorm.survival_function(x)
self._testBatchShapes(halfnorm, sf)
log_sf = halfnorm.log_survival_function(x)
self._testBatchShapes(halfnorm, log_sf)
expected_logsf = sp_stats.halfnorm(scale=scale).logsf(x)
self.assertAllClose(expected_logsf, self.evaluate(log_sf), atol=0)
self.assertAllClose(np.exp(expected_logsf), self.evaluate(sf), atol=0)
def testHalfNormalLogPDFMultidimensional(self):
batch_size = 6
scale = tf.constant([[3.0, 1.0]] * batch_size)
x = np.array([[-2.5, 2.5, 4.0, 0.0, -1.0, 2.0]], dtype=np.float32).T
halfnorm = tfd.HalfNormal(scale=scale, validate_args=False)
log_pdf = halfnorm.log_prob(x)
self._testBatchShapes(halfnorm, log_pdf)
pdf = halfnorm.prob(x)
self._testBatchShapes(halfnorm, pdf)
expected_log_pdf = sp_stats.halfnorm(scale=self.evaluate(scale)).logpdf(x)
self.assertAllClose(expected_log_pdf, self.evaluate(log_pdf))
self.assertAllClose(np.exp(expected_log_pdf), self.evaluate(pdf))
def __init__(
self,
n_particles,
prior={
'β0': norm(loc=0, scale=50),
'β1': norm(loc=0, scale=50),
'β2': norm(loc=0, scale=50),
'β3': norm(loc=0, scale=50),
'β4': norm(loc=0, scale=50),
'α': halfnorm(loc=0, scale=3)
}):
self.n_particles = int(n_particles)
self.prior = prior
self.θ_fixed = {'ϵ': 0.01}
self.choiceFunction = CumulativeNormalChoiceFunc
def testHalfNormalCDF(self):
batch_size = 50
scale = self._rng.rand(batch_size) + 1.0
x = np.linspace(-8.0, 8.0, batch_size).astype(np.float64)
halfnorm = tfd.HalfNormal(scale=scale, validate_args=False)
cdf = halfnorm.cdf(x)
self._testBatchShapes(halfnorm, cdf)
log_cdf = halfnorm.log_cdf(x)
self._testBatchShapes(halfnorm, log_cdf)
expected_logcdf = sp_stats.halfnorm(scale=scale).logcdf(x)
self.assertAllClose(expected_logcdf, self.evaluate(log_cdf), atol=0)
self.assertAllClose(np.exp(expected_logcdf), self.evaluate(cdf), atol=0)
var_mean = dot(pdf_mean, (x-mean_mean)**2)/pdf_mean.sum()
f1 = figure(num=1, figsize=(13.66,7.68), dpi=100)
plot(x, pdf_actual, color='blue', label='Numerical')
plot(x, pdf_peak, color='red', linestyle='--', label='Theoretical')
plot(x, pdf_mean, color='green', linestyle='--', label='Theo mean')
xlabel('Membrane potential (volt)')
title('Theoretical and numerical pre-spike volt distributions')
legend(loc='best')
figname = 'tmp/%s_psvolt.png' % filehead
savefig(figname)
f1.clf()
rms_mean = sqrt(mean((pdf_mean - pdf_actual)**2))
rms_peak = sqrt(mean((pdf_peak - pdf_actual)**2))
hn_mean = halfnorm(loc=v_th-Mt_mean, scale=sqrt(Vt_mean)/dt)
hn_peak = halfnorm(loc=v_th-Mt_peak, scale=sqrt(Vt_peak)/dt)
quantiles = frange(0, 1, 0.01)
hn_mean_qs = hn_mean.ppf(quantiles)
hn_peak_qs = hn_peak.ppf(quantiles)
#from IPython import embed
#embed()
#sys.exit()
#print "RMS mean dist: %s\nRMS peak dist: %f" % (rms_mean, rms_peak)
#print "KS Mean: %f\nKS Peak: %f" % (ks_mean[1], ks_peak[1])
#print "Avgs Actual, Mean, Peak: %10.9f, %10.9f, %10.9f" % (
# mean_actual,