def find_tail_point(w, min_height, ratio):
"""Find best fit area,
that minimize std of gaussian approximation from waveform.
find when waveform fall below area * ratio,
and at least 2 sigma away from mean time.
"""
if np.max(w) < min_height:
return []
# Quick fit to the waveform and locate tail position
ix = np.arange(len(w))
# Use center population to estimate mean and sigma
cnt = w > 0.02*np.max(w)
mean = np.sum(ix[cnt]*w[cnt])/np.sum(w[cnt])
sigma = (np.sum((ix[cnt]-mean)**2*w[cnt])/np.sum(w[cnt]))**0.5
# Use estimated mean and sigma to find amplitude
amp = np.sum(w[cnt]*norm.pdf(ix[cnt], mean, sigma))/np.sum(norm.pdf(ix[cnt], mean, sigma)**2)
# Define tail by waveform height drop below certain ratio of amplitude
tail = np.where(w < ratio*amp)[0]
tail = tail[(tail > mean+2*sigma) & (tail != len(w))]
if len(tail) > 0:
return tail[:1]
else:
return []
def theta(cflag,s,k,v,r,t):
d1=(np.log(s/k)+(r+v*v/2)*t)/(v*np.sqrt(t))
d2=d1-v*np.sqrt(t)
if cflag: # call
return -s*norm.pdf(d1)*v/(2*np.sqrt(t)) - r*k*np.exp(-r*t)*norm.pdf(d2)
else: # put
return -s*norm.pdf(d1)*v/(2*np.sqrt(t)) + r*k*np.exp(-r*t)*norm.pdf(-d2)
def dual_gaussian(x,
amp1=1.0, mean1=0.0, std1=1.0,
amp2=1.0, mean2=0.0, std2=1.0):
"""Sum of two Gaussians.
Parameters
----------
x : array
Function argument
amp1: float
Amplitude parameter of the first Gaussian
mean1: float
Mean parameter of the first Gaussian
std1: float
Standard deviation parameter of the first Gaussian
amp2: float
Amplitude parameter of the second Gaussian
mean2: float
Mean parameter of the second Gaussian
std2: float
Standard deviation parameter of the second Gaussian
"""
from scipy.stats import norm
if std1 <= 0 or std2 <= 0:
return np.nan
return (amp1 * norm.pdf(x, mean1, std1)) + (amp2 * norm.pdf(x, mean2, std2))
def test_mlexplorer(self):
x = np.arange(0,600,1.0)
nb_samples = 100 # number of samples in our dataset
# true partial spectra
S_1 = norm.pdf(x,loc=200.,scale=130.)
S_2 = norm.pdf(x,loc=400,scale=70)
S_true = np.vstack((S_1,S_2))
#60 samples with random concentrations between 0 and 1
C_ = np.random.rand(nb_samples)
C_true = np.vstack((C_,(1-C_))).T
# We make some observations with random noise
Obs = np.dot(C_true,S_true) + np.random.randn(nb_samples,len(x))*1e-4
# new observations
C_new_ = np.random.rand(10) #10 samples with random concentrations between 0 and 1
C_new_true = np.vstack((C_new_,(1-C_new_))).T
noise_new = np.random.randn(len(x))*1e-4
Obs_new = np.dot(C_new_true,S_true) + noise_new
explo = rp.mlexplorer(Obs)
# we just test that it runs for now
explo.algorithm = 'NMF'
explo.nb_compo = 2
explo.test_size = 0.3
explo.scaler = "MinMax"
explo.fit()
explo.refit()
def testGaussian2D(self):
gauss2d = Gauss2D(0.5/np.pi, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0)
self.assertEqual(gauss2d.integral(), 1.0)
self.assertEqual(gauss2d.evaluate(0.0, 0.0), 0.5/np.pi)
gauss2d.update_params(mu0=-78.9)
self.assertEqual(gauss2d.evaluate(-78.9, 0.0), 0.5/np.pi)
self.assertGreater(gauss2d.evaluate(-78.9, 0.0), gauss2d.evaluate(-78.9, 1.0))
gauss2d.update_params(offset=1234.5)
self.assertEqual(gauss2d.integral(), 1.0)
# Test the actual fitting.
mu0 = 40.0
mu1 = 60.0
sigma0 = 10.0
sigma1 = 7.0
x = np.arange(100)
X0, X1 = np.meshgrid(x, x, indexing="ij")
Z = norm.pdf((X0 - mu0) / sigma0) * norm.pdf((X1 - mu1) / sigma1)
gauss2d.initial_guess(Z)
gauss2d.fit(Z)
self.assertAlmostEqual(gauss2d.param_dict["mu0"], mu0)
self.assertAlmostEqual(gauss2d.param_dict["mu1"], mu1)
self.assertAlmostEqual(gauss2d.param_dict["sigma0"], sigma0)
self.assertAlmostEqual(gauss2d.param_dict["sigma1"], sigma1)
def analyze_pedestrian_queue(file_path):
"""Analyze the pedestrians created by the simulation."""
destinations, speeds = np.loadtxt(file_path, delimiter=',', unpack=True)
# Show Histograms of destination selections
plt.hist(destinations, bins=len(np.unique(destinations)), normed=True)
plt.xlabel("Selected Destinations (IDs)")
plt.ylabel("Normalized Frequency (PDF)")
plt.grid(True)
# Intended pedestrian speed distribution
mu = 1.340
sigma = 0.265
pdf_range = np.arange(0, 2.5, 0.001)
# Plot theoretical distribution
plt.figure()
plt.plot(pdf_range, norm.pdf(pdf_range, mu, sigma))
plt.grid(True)
plt.xlabel("Pedestrian Speeds (m/s)")
plt.ylabel("PDF")
# Plot the histogram of speeds with the theoretical distribution overlaid
plt.figure()
plt.hist(speeds, bins=int((2*len(speeds))**(1./3)), normed=True)
plt.grid(True)
plt.xlabel("Pedestrian Speeds (m/s)")
plt.ylabel("Normalized Frequency (PDF)")
plt.plot(pdf_range, norm.pdf(pdf_range, mu, sigma), linewidth=2)
plt.legend(["Theory", "Actual"])
plt.show() # Show the plots
def VOIfunc(self, n, pointNew, grad):
xNew = pointNew
nTraining = self._GP._numberTraining
tempN = nTraining + n
# n=n-1
vec = np.zeros(tempN)
X = self._PointsHist
for i in xrange(tempN):
vec[i] = self._GP.muN(X[i, :], n)
maxObs = np.max(vec)
std = np.sqrt(self._GP.varN(xNew, n))
muNew, gradMu = self._GP.muN(xNew, n, grad=True)
Z = (muNew - maxObs) / std
temp1 = (muNew - maxObs) * norm.cdf(Z) + std * norm.pdf(Z)
if grad == False:
return temp1
var, gradVar = self._GP.varN(xNew, n, grad=True)
gradstd = 0.5 * gradVar / std
gradZ = ((std * gradMu) - (muNew - maxObs) * gradstd) / var
temp10 = (
gradMu * norm.cdf(Z)
+ (muNew - maxObs) * norm.pdf(Z) * gradZ
+ norm.pdf(Z) * gradstd
+ std * (norm.pdf(Z) * Z * (-1.0)) * gradZ
)
return temp1, temp10
def plot_mcmc_samples(data, samples, true_clusters, x_min= -15, x_max=15, stepsize=0.001):
# Plot the data
plt.title('Observed Data Points\n{0} total'.format(len(data)))
plt.hist(data, 20)
plt.savefig('points.pdf')
plt.clf()
# Plot the mean and bands of size 2 stdevs for the samples.
# Also plot the last sample from the MCMC chain.
xvals = np.arange(x_min, x_max, stepsize)
true_pdf = sum(norm.pdf(xvals, mean, stdev) for mean, stdev in true_clusters) / float(len(true_clusters))
no_bands = np.zeros(len(xvals))
sample_pdfs = np.array([sum([norm.pdf(xvals, mean, stdev) for mean, stdev in sample]) / float(len(sample)) for sample in samples])
sample_means = sample_pdfs.mean(axis=0)
sample_bands = sample_pdfs.std(axis=0)*2.
last = sample_pdfs[-1]
means = np.array([true_pdf, sample_means, last])
bands = np.array([no_bands, sample_bands, no_bands])
names = ['True PDF', 'Bayes\nEstimate\n(+/- 2 stdevs)', 'Last MCMC\nsample']
mcmc_params = '{0} points, {1} iterations, {2} burn-in, {3} thin, {4} samples'.format(NUM_POINTS, ITERATIONS, BURN_IN, THIN, len(samples))
dp_params = 'alpha={0}, mu0={1}, nu0={2}, a0={3}, b0={4}'.format(ALPHA, MU_0, NU_0, A_0, B_0)
plot_noisy_means('Dirichlet Process Mixture Results', means, bands, names, xvals=xvals, xlabel='X', ylabel='Probability', subtitle='{0}\n{1}'.format(mcmc_params, dp_params))
# Plot the distribution of clusters in the samples
plot_cluster_distribution(samples)
def get_sent_similarity(user_data):
scores=[]
#=====[ Creates counts for each sentiment score in 21 buckets of width 0.1 from -1 to 1 ]=====
for data in user_data:
user_score = [0]*21
for tweet in data:
score = int(float("%.1f" % tweet['score'])*10+10)
user_score[score] += 1
scores.append(user_score)
#=====[ Forms normalized probability distributions for each users sentiments ]=====
x = np.linspace(-1, 1, 100)
mu, std = norm.fit(scores[0])
p = norm.pdf(x, mu, std)
mu, std = norm.fit(scores[1])
p2 = norm.pdf(x,mu,std)
#=====[ Takes Kullback-Leibler Divergence between probability distributions ]=====
similarity = float("%.5f" % scipy.stats.entropy(p,p2))
#=====[ Converts similarity score to a percentage from 10 - 90 to display on compatability spectrum ]=====
if similarity < 0.003:
return 90
elif similarity > 0.07:
return 10
else:
return int(10 + ((similarity*100)-1)/6.7*80)
return int(similarity)
def dist2weights_gauss(dist, max_r, max_w=1, min_w=1e-3, S=None, rescale=True):
"""Gaussian distance weighting.
Parameters
----------
dist: float or np.ndarray
the distances to be transformed into weights.
max_r: float
maximum radius of the neighbourhood considered.
max_w: int (default=1)
maximum weight to be considered.
min_w: float (default=1e-8)
minimum weight to be considered.
S: float or None (default=None)
the scale magnitude.
rescale: boolean (default=True)
if re-scale the magnitude.
Returns
-------
weights: np.ndarray, array_like, shape (num. of retrievable candidates)
values of the weight of the neighs inferred from the distances.
"""
if S is None:
S = set_scale_gauss(max_r, max_w, min_w)
if rescale:
A = max_w/(norm.pdf(0, scale=S)-norm.pdf(max_r, scale=S))
weights = A*norm.pdf(dist, scale=S)
else:
A = max_w/norm.pdf(0, scale=S)
weights = A*norm.pdf(dist, scale=S)
return weights
def show_cond():
results = load_same_loc()
tt = 8
# Graph the histogram of activity at the center of
# of the room from cell 0
acts = [x[0] for x in results]
plt.subplot(2,1,1)
count, bins, _ = plt.hist(acts,bins=40,normed=Normed)
plt.title('Original dist. Pts: %i'%(len(acts),))
xlm = plt.xlim()
xs = np.linspace(xlm[0],xlm[1],1000)
mn, std = gaussian(acts)
plt.plot(xs,norm.pdf(xs,mn,std))
# Graph the histogram of activity at the center of
# of the room from cell 1
plt.subplot(2,1,2)
nb = [x[1] for x in results]
plt.hist(nb,bins=bins,normed=Normed)
mn, std = gaussian(nb)
plt.plot(xs,norm.pdf(xs,mn,std))
plt.title('Side dist. Pts: %i'%(len(nb),))
plt.xlim(xlm)
mx_i = np.argmax(count)
plt.figure()
plt.subplot(2,1,1)
nb = [x[1] for x in results]
plt.hist(nb,bins=bins,normed=Normed)
plt.title('Side dist. Pts: %i'%(len(nb),))
mn, std = gaussian(nb)
plt.plot(xs,norm.pdf(xs,mn,std))
plt.xlim(xlm)
plt.subplot(2,1,2)
iss = get_iss(mx_i+1, bins, acts)
acts_neighbor = [x[1] for x in results[iss]]
plt.hist(acts_neighbor,bins=bins,normed=Normed)
plt.title('Pts: %i'%(len(acts_neighbor)))
nb = acts_neighbor
mn, std = gaussian(nb)
plt.plot(xs,norm.pdf(xs,mn,std))
plt.xlim(xlm)
plt.figure()
for delt in range(-tt,tt):
iss = get_iss(mx_i+delt, bins, acts)
acts_neighbor = [x[1] for x in results[iss]]
plt.subplot(2,tt,delt+tt+0)
plt.hist(acts_neighbor,bins=bins,normed=Normed)
nb = acts_neighbor
mn, std = gaussian(nb)
plt.plot(xs,norm.pdf(xs,mn,std))
plt.title('Pts: %i'%(len(acts_neighbor)))
plt.xlim(xlm)
plt.show()
def mcmc(N=1000, k={"t1":100, "t2":100, "t3":5}, x=[], v=[]):
chute = {"t1":[10],"t2":[10],"t3":[0.01]}
M = chute
hiper = {"t1":[0,100],"t2":[0,100],"t3":[0.1,0.1]} #VALORES DOS HIPERPARAMETROS
for i in range(N-1):
for j in M.keys():
if j == "t1" or j == "t2":
M[j].append( np.random.normal(loc = M[j][-1]-k[j]/100, scale = k[j], size = 1) )
lista = [ [ M[l][-1] for l in M.keys()] , [ M[l][-1] if l!=j else M[l][-2] for l in M.keys() ] ]
t1 = norm.pdf(M[j][-1], loc = hiper[j][0], scale = hiper[j][1]) * L(x, v, lista[0]) * norm.pdf(M[j][-2], loc = M[j][-1]-k[j]/100, scale = k[j])
t2 = norm.pdf(M[j][-2], loc = hiper[j][0], scale = hiper[j][1]) * L(x, v, lista[1]) * norm.pdf(M[j][-1], loc = M[j][-2]-k[j]/100, scale = k[j])
teste = (t1/t2)
else:
M[j].append( np.random.gamma(shape = M[j][-1]*k[j], scale = k[j], size = 1) )
lista = [ [ M[l][-1] for l in M.keys()] , [ M[l][-1] if l!=j else M[l][-2] for l in M.keys() ] ]
t1 = gamma.pdf(M[j][-1], a = hiper[j][0], scale = hiper[j][1]) * L(x, v, lista[0]) * gamma.pdf(M[j][-2], a = M[j][-1]*k[j], scale = k[j])
t2 = gamma.pdf(M[j][-2], a = hiper[j][0], scale = hiper[j][1]) * L(x, v, lista[1]) * gamma.pdf(M[j][-1], a = M[j][-2]*k[j], scale = k[j])
teste = (t1/t2)
if (min(1 , teste) < np.random.uniform(low = 0, high = 1, size = 1) ) or (np.isinf(teste)) or (np.isnan(teste)) :
M[j][-1] = M[j][-2]
return(M)
def generate_random_traffic(size=200, date_start="2016/01/01", output_path= "simu.csv"):
x = np.linspace(1, size, size)
index = pd.date_range(date_start, periods=size, freq="1H")
rs = np.random.RandomState()
g = pd.Series(0, index=index)
# gaussian mixture
for i in rs.choice(x, size=size / 3, replace=False):
spread = rs.uniform(1, 100)
multiplicator = 5 * spread
g += pd.Series(norm.pdf(x, i, spread) * multiplicator, index=index)
# gaussian 24h seasonality
seasonx = x = np.linspace(1, 24, 24)
season_index = pd.date_range('01/01/2016', periods=24, freq="1H")
season = pd.Series(0, index=season_index)
for i in rs.choice(x[6:-6], size=4, replace=True):
spread = rs.uniform(1, 20)
multiplicator = g.mean()
s = pd.Series(norm.pdf(seasonx, i, spread) * multiplicator, index=season_index)
season += s
for i in range(0, len(g) / len(season)):
g = g.add(season.shift(24 * i, freq="1H") * 3, fill_value=0)
g = g.add(pd.Series(0, index=index).cumsum())
g.to_csv(output_path)
def simulate_stupidDPM(iter_num, M):
# Generate mixture sample
N = 1000
mu = [0.0, 10.0, 3.0]
components = np.random.choice(range(3), size = N, replace = True, p = [0.3, 0.5, 0.2])
samples = [norm.rvs(size = 1, loc = mu[components[i]], scale = 1)[0] for i in range(N)]
## Sample G from DP(M, G0)
v = beta.rvs(a = 1.0, b = M, size = N)
prob_vector = np.append(np.array(v[0]), v[1:] * np.cumprod(1.0 - v[:-1]))
thetas = norm.rvs(size = N, loc = 1.0, scale = 1.0)
### Initialize thetas
thetas = np.random.choice(thetas, size = N, replace = True, p = prob_vector)
### Start MCMC chain
for i in xrange(iter_num):
for j in xrange(N):
theta_temp = np.append(thetas[:j], thetas[j+1:])
p = np.append(norm.pdf(samples[j], loc = theta_temp, scale = 1.0), M * norm.pdf(samples[j], loc = 1.0, scale = np.sqrt(2.0)))
p = p / sum(p)
temp = np.random.choice(np.append(theta_temp, N), size = 1, replace = True, p = p)
if (temp == N):
thetas[j] = norm.rvs(size = 1, loc = 0.5 * (samples[j] + 1), scale = np.sqrt(0.5))
else:
thetas[j] = temp
print(thetas)
return {"thetas": thetas, "y": samples}
def __init__(self, image=None, mesh_factor=14, density_distribution=None):
super(CoordinateSolver, self).__init__()
self.image = image
self.height, self.width, _ = self.image.shape
self.mesh_factor = mesh_factor
self.height /= self.mesh_factor
self.width /= self.mesh_factor
self.image = self.image[:self.mesh_factor*self.height, :self.mesh_factor*self.width]
if type(density_distribution) == np.ndarray:
restricted_density = density_distribution[:self.mesh_factor*self.height, :self.mesh_factor*self.width]
target_areas = restricted_density
target_areas = target_areas[:-1, :-1]
else:
target_areas = np.indices((self.width-1, self.height-1)).T.astype('float32')
target_areas = norm.pdf(target_areas[:, :, 0], self.width/2, self.width/5)\
*norm.pdf(target_areas[:, :, 1], self.height/2, self.height/5)
target_areas /= sum(sum(target_areas))
normalisation_factor = (self.height-1)*(self.width-1)
target_areas_normalised = target_areas*normalisation_factor
self.padded_targets = np.zeros([self.height+1, self.width+1])
self.padded_targets[1:-1, 1:-1] = target_areas_normalised
self.coordinates = np.indices((self.width, self.height)).T.astype('float32')
self.total_error = (self.height-1)*(self.width-1)
self.min_coords = self.coordinates.copy()
self.areas = calculate_areas(self.coordinates)
self.errors = np.zeros(self.padded_targets.shape)
self.x_weights = np.ones([self.height*self.width, self.height + 1, self.width + 1])
self.y_weights = np.ones([self.height*self.width, self.height + 1, self.width + 1])
self.make_weights()
def get_beam_loc(ra,dec,utc,beam_a,beam_b,sn_a,sn_b):
sn_a = float(sn_a)
sn_b = float(sn_b)
boresight = ephem.Equatorial(ra,dec)
ra_a,dec_a,info = coords.radec_of_offset_fanbeam(boresight.ra,boresight.dec,beam_a,utc)
ns_a,ew_a = info['fanbeam_nsew']
lst = info['mol_curr'].sidereal_time()
ra_b,dec_b,info = coords.radec_of_offset_fanbeam(boresight.ra,boresight.dec,beam_b,utc)
ns_b,ew_b =info['fanbeam_nsew']
ns = np.linspace(np.radians(-3.0),np.radians(3.0),size)
ew = np.linspace(np.radians(-100/3600.0),np.radians(100/3600.0),size)
ns,ew = np.meshgrid(ns,ew)
beam_a_map = fanbeam_model(ns_a,ew_a,ns+ns_a,ew+ew_a)
beam_b_map = fanbeam_model(ns_b,ew_b,ns+ns_a,ew+ew_a)
ratio = sn_a/sn_b
ratio_sigma = ratio * np.sqrt(1/sn_a**2 + 1/sn_b**2)
ratio_map = beam_a_map/beam_b_map
ns_fwhm = np.radians(2.0)
prob_map = norm.pdf(ratio_map,ratio,ratio_sigma) * norm.pdf(ns,0.0,ns_fwhm/2.355)
prob_map/=prob_map.max()
ns = ns+ns_a
ew = ew+ew_a
best_ns = ns[(prob_map.max(axis=1).argmax(),prob_map.max(axis=0).argmax())]
best_ew = ew[(prob_map.max(axis=1).argmax(),prob_map.max(axis=0).argmax())]
ha,dec = coords.nsew_to_hadec(best_ns,best_ew)
ra,dec = coords.radec_to_J2000(lst-ha,dec,utc)
return ra,dec,ns,ew,prob_map,lst
def camera_waveforms():
camera = CameraGeometry.from_name("CHEC")
n_pixels = camera.n_pixels
n_samples = 96
mid = n_samples // 2
pulse_sigma = 6
r_hi = np.random.RandomState(1)
r_lo = np.random.RandomState(2)
x = np.arange(n_samples)
# Randomize times
t_pulse_hi = r_hi.uniform(mid - 10, mid + 10, n_pixels)[:, np.newaxis]
t_pulse_lo = r_lo.uniform(mid + 10, mid + 20, n_pixels)[:, np.newaxis]
# Create pulses
y_hi = norm.pdf(x, t_pulse_hi, pulse_sigma)
y_lo = norm.pdf(x, t_pulse_lo, pulse_sigma)
# Randomize amplitudes
y_hi *= r_hi.uniform(100, 1000, n_pixels)[:, np.newaxis]
y_lo *= r_lo.uniform(100, 1000, n_pixels)[:, np.newaxis]
y = np.stack([y_hi, y_lo])
return y, camera
def model_mixture(separation, prevalence_of_1st):
x_arr = np.linspace(-7, 7, 100)
plt.title('%s std delta - %s%% - %s%%'% (separation, prevalence_of_1st*100, (1-prevalence_of_1st)*100))
plt.plot(x_arr, prevalence_of_1st*norm.pdf(x_arr, loc=-separation/2.), 'r')
plt.plot(x_arr, (1-prevalence_of_1st)*norm.pdf(x_arr, loc=separation/2.), 'b')
plt.plot(x_arr, prevalence_of_1st*norm.pdf(x_arr, loc=-separation/2.)
+ (1-prevalence_of_1st)*norm.pdf(x_arr, loc=separation/2.), 'k')
def predictive_recursion_fdr(z, sweeporder, grid_x, theta_guess, mu0 = 0.,
sig0 = 1.0, nullprob = 1.0, decay = -0.67):
gridsize = grid_x.shape[0]
theta_subdens = deepcopy(theta_guess)
pi0 = nullprob
joint1 = np.zeros(gridsize)
ftheta1 = np.zeros(gridsize)
# Begin sweep through the data
for i, k in enumerate(sweeporder):
cc = (3. + i)**decay
joint1 = norm.pdf(grid_x, loc=z[k] - mu0, scale=sig0) * theta_subdens
m0 = pi0 * norm.pdf(z[k] - mu0, 0., sig0)
m1 = trapezoid(grid_x, joint1)
mmix = m0 + m1
pi0 = (1. - cc) * pi0 + cc * (m0 / mmix)
ftheta1 = joint1 / mmix
theta_subdens = (1. - cc) * theta_subdens + cc * ftheta1
# Now calculate marginal distribution along the grid points
y_mix = np.zeros(gridsize)
y_signal = np.zeros(gridsize)
for i, x in enumerate(grid_x):
joint1 = norm.pdf(grid_x, x - mu0, sig0) * theta_subdens
m0 = pi0 * norm.pdf(x, mu0, sig0)
m1 = trapezoid(grid_x, joint1)
y_mix[i] = m0 + m1;
y_signal[i] = m1 / (1. - pi0)
return {'grid_x': grid_x,
'sweeporder': sweeporder,
'theta_subdens': theta_subdens,
'pi0': pi0,
'y_mix': y_mix,
'y_signal': y_signal}
def forward_scaled(self, data): ''' Alpha is the joint probability of observing all data and z alpha(z[n]) = p(x1...xn, z[n]) Use scale parameter to prevent underflow ''' data = data_raw[1] length = data.shape[0] # Convert data to standard deviations and find probabilities data_norm = (data[:,None] - emission_means) / np.sqrt(emission_variances) emission_pdf = norm.pdf(data_norm).T emission_pdf /= emission_pdf.sum(0) alpha = np.zeros([n_states, length], np.float) scale = np.zeros(length, np.float) # Initialize alpha # datum_norm = (data[0] - emission_means) / np.sqrt(emission_variances) alpha[:,0] = prior * norm.pdf(data_norm[0]) scale[0] = 1./alpha[:,0].sum() alpha[:,0] *= scale[0] # Recursively update alpha # for i,d in enumerate(data[1:]): for t in xrange(1,length): for k in xrange(n_states): # # alpha[k] = np.sum( p(X|Z) * p(Z'|Z) * alpha[k]) alpha[k,t] = np.sum( alpha[:,t-1] * transition_matrix[:,k]) * emission_pdf[k,t] scale[t] = 1./alpha[:,t].sum() alpha[:,t] *= scale[t]
def starloglik(para):
res = para[0] * norm.pdf(dtemp[:, 17] + para[1], scale=1) + (1 - para[0]) * norm.pdf(
dtemp[:, 17], loc=para[2], scale=para[3]
)
res[res == 0] = 10 ** (-200)
res = -np.sum(np.log(res))
return res
def BlackScholes():
data = {}
S = float(request.args.get('price'))
K = float(request.args.get('strike'))
T = float(request.args.get('time'))
R = float(request.args.get('rate'))
V = float(request.args.get('vol'))
d1 = (log(float(S)/K)+(R+V*V/2.)*T)/(V*sqrt(T))
d2 = d1-V*sqrt(T)
data['cPrice'] = S*norm.cdf(d1)-K*exp(-R*T)*norm.cdf(d2)
data['pPrice'] = K*exp(-R*T)-S+data['cPrice']
data['cDelta'] = norm.cdf(d1)
data['cGamma'] = norm.pdf(d1)/(S*V*sqrt(T))
data['cTheta'] = (-(S*V*norm.pdf(d1))/(2*sqrt(T))-R*K*exp(-R*T)*norm.cdf(d2))/365
data['cVega'] = S*sqrt(T)*norm.pdf(d1)/100
data['cRho'] = K*T*exp(-R*T)*norm.cdf(d2)/100
data['pDelta'] = data['cDelta']-1
data['pGamma'] = data['cGamma']
data['pTheta'] = (-(S*V*norm.pdf(d1))/(2*sqrt(T))+R*K*exp(-R*T)*norm.cdf(-d2))/365
data['pVega'] = data['cVega']
data['pRho'] = -K*T*exp(-R*T)*norm.cdf(-d2)/100
return json.dumps(data)
def obj_func1(params, price_history, vol, price_durations, y_bar): """ Objective function for optimization """ num_trades = len(price_history) - 1 P_last = price_history[:-1] pdelta = np.array(price_history[1:]) - P_last K = len(price_durations) + 1 informed_prices = params[-K:] P_informed = abmtools.create_price_vector(informed_prices, price_durations, num_trades) x = params[:-K] Sigma_0 = x[4] lfunc1 = lambda x: likelihood1(pdelta, vol, x[0], x[1], x[2], x[3], P_informed, P_last, Sigma_0, y_bar) \ + np.log(norm.pdf(x[0], 0.5, 0.1)) \ + np.log(norm.pdf(x[1], 5e-2, 1e-1)) \ + np.log(norm.pdf(x[2], 100., 15.)) \ + np.log(norm.pdf(x[3], 0.02**2, 0.01)) \ + np.log(norm.pdf(x[4], 10, 5)) return -lfunc1(x)
def log_posterior_1(params0, price_history, vol, price_durations, y_bar): """ Threshold function for metropolis hastings. It is the the log of posterior distribution """ num_trades = len(price_history) - 1 P_last = price_history[:-1] pdelta = np.array(price_history[1:]) - P_last K = len(price_durations) + 1 informed_prices = params0[-K:] P_informed = abmtools.create_price_vector(informed_prices, price_durations, num_trades) x = params0[:-K] Sigma_0 = x[4] lfunc1 = lambda x: likelihood1(pdelta, vol, x[0], x[1], x[2], x[3], P_informed, P_last, Sigma_0, y_bar) \ + np.log(norm.pdf(x[0], 0.3, 0.1)) \ + np.log(norm.pdf(x[1], 5e-2, 1e-1)) \ + np.log(norm.pdf(x[2], 80., 15.)) \ + np.log(norm.pdf(x[3], 0.02**2, 0.01)) \ + np.log(norm.pdf(x[4], 10, 5)) return lfunc1(x)
def lnprob(self, theta):
lp = 0
# Covariance amplitude
lp += self.ln_prior.lnprob(theta[0])
# Lengthscales
lp += self.tophat.lnprob(theta[1:self.n_ls + 1])
# Prior for the Bayesian regression kernel
pos = (self.n_ls + 1)
end = (self.n_ls + self.n_lr + 1)
lp += -np.sum((theta[pos:end]) ** 2 / 10.)
# Env Noise
#lp += self.ln_prior_env_noise.lnprob(theta[end])
#lp += self.ln_prior_env_noise.lnprob(theta[end + 1])
# alpha
lp += norm.pdf(theta[end], loc=-7, scale=1)
# beta
lp += norm.pdf(theta[end + 1], loc=0.5, scale=1)
# Noise
lp += self.horseshoe.lnprob(theta[-1])
return lp
def windowScores3(seq, length): diffs = seq[1:] - seq[:-1] absDiffs = np.abs(diffs) diffs_2 = diffs[1:] - diffs[:-1] absDiffs_2 = np.abs(diffs_2) # variance_diff = np.var(diffs) # expectedDiff = np.mean(absDiffs) expectedDiff = np.std(diffs) expectedDiffs_2 = absDiffs[:-1] + expectedDiff scores = np.zeros(seq.shape) actualDiffs = diffs[:-1] actualDiffs_2 = absDiffs_2 # want (actual diff / expected diff) * (expected diff_2 / actual diff_2) firstProbs = norm.pdf(actualDiffs / expectedDiff) secondProbs = norm.pdf((actualDiffs_2 - expectedDiffs_2) / expectedDiff) # scores[1:-1] = (actualDiffs / expectedDiff) * (expectedDiffs_2 / actualDiffs_2) scores[1:-1] = firstProbs * secondProbs # scores[1:-1] = firstProbs # scores[1:-1] = secondProbs print describe(scores, 'scores') return scores
def theta(self, S, t, vol, r):
if t > self.T:
return np.zeros(len(S))
if self.op_type == 'c':
return np.subtract(
np.true_divide(
np.multiply(-vol,
np.multiply(S, norm.pdf(self.d1(S, t, vol, r)))),
np.multiply(2,
np.power(
np.subtract(self.T, t), .5))),
np.multiply(r,
np.multiply(self.K,
np.multiply(
np.exp(
np.multiply(-r,
np.subtract(self.T, t))),
norm.cdf(self.d2(S, t, vol, r))))))
else:
return np.add(
np.true_divide(
np.multiply(-vol,
np.multiply(S, norm.pdf(self.d1(S, t, vol, r)))),
np.multiply(2,
np.power(
np.subtract(self.T, t), .5))),
np.multiply(r,
np.multiply(self.K,
np.multiply(
np.exp(
np.multiply(-r,
np.subtract(self.T, t))),
norm.cdf(
np.multiply(-1, self.d2(S, t, vol, r)))))))
def plotEstimateVsN():
"""
Estimate the mean of a Gaussian distribution, plotting the resulting distributions
as the number of data points increase. The true variance is known in advance.
"""
# Generate points from a Gaussian distribution
mu = -0.8 # Mean
var = 0.1 # Variance
# Use mean of 0 for prior, and the true variance for both prior and likelihoods
data = np.random.normal(mu, var, 10)
# Plot the prior distribution with mean 0 and true variance
x = np.linspace(-1, 1, 100)
plt.plot(x, norm.pdf(x, 0, np.sqrt(var)), label="N = 0")
# Plot distribution as N gets larger
for i in [1, 2, 10]:
# Estimate the mean and variance from i data points
mu, v = estimateBayes(data[:i], 0, var, var)
# Plot the normal distribution curve
plt.plot(x, norm.pdf(x, mu, np.sqrt(v)), label="N = {0}".format(i))
plt.legend()
plt.show()
def P_number_true_obs_fast(args): E_true_links = 0 p_at_least_one_bp_at_given_position = 1- P_breakpoints_in_interval(1, args.bp_ratio, 0) k = 2 * args.readlen / float(args.cov) for i in range((args.insertion_size + args.readlen - args.soft)+1, args.mean + 4*args.stddev): # When internal breakpoint occur within mean + 4*sigma E_true_links += 2*(1/k) * (v(i,args.insertion_size, args.readlen, args.soft) - 1 ) * norm.pdf(i, args.mean,args.stddev)*poisson.pmf(0,args.bp_ratio*i)*p_at_least_one_bp_at_given_position**2 # when no breakpoint occurs on one side E_true_links += 2*(1/k) * 1 * norm.pdf(i, args.mean,args.stddev)*poisson.pmf(0,args.bp_ratio*i)*p_at_least_one_bp_at_given_position # when no breakpoint occurs on both sides E_true_links += (1/k) * 1 * norm.pdf(i, args.mean,args.stddev)*poisson.pmf(0,args.bp_ratio*i) #print v(i,args.insertion_size, args.readlen, args.soft) # when no breakpoint occurs on one side i = args.mean + 4*args.stddev E_true_links += 2*(1/k)*v(i,args.insertion_size, args.readlen, args.soft)*norm.pdf(i, args.mean,args.stddev)*poisson.pmf(0,args.bp_ratio*i)*p_at_least_one_bp_at_given_position # when no breakpoint occurs on both sides i = args.mean + 4*args.stddev print v(i,args.insertion_size, args.readlen, args.soft) print 1/k E_true_links += (1/k)*v(i,args.insertion_size, args.readlen, args.soft)*norm.pdf(i, args.mean,args.stddev)*poisson.pmf(0,args.bp_ratio*i) return E_true_links
def naive_bayes(data,classifier,sample):
from scipy.stats import norm
idx_male = array([i for i in range(len(classifier)) if classifier[i]==0])
idx_female = array([i for i in range(len(classifier)) if classifier[i]==1])
mean_male = mean(data[idx_male,:],0)
std_male = std(data[idx_male,:],0)
mean_female = mean(data[idx_female,:],0)
std_female = std(data[idx_female,:],0)
probs_female = []
for i in range(len(mean_female)):
probs_female.append( norm.pdf(sample[i],mean_female[i],std_female[i]))
probs_male = []
for i in range(len(mean_male)):
probs_male.append( norm.pdf(sample[i],mean_male[i],std_male[i]))
p_male = cumprod(probs_male)[-1] * 0.5
p_female = cumprod(probs_female)[-1] * 0.5
if p_male > p_female:
print 'The person is MALE', 'says: Naive Bayes'
return 0
else:
print 'The person is FEMALE','says: : Naive Bayes'
return 1
def tsmaker(m, s, j):
t = np.arange(0.0, 1.0, 0.01)
v = norm.pdf(t, m, s) + j * np.random.randn(100)
return ts.TimeSeries(v, t)
def gfit(inp, sec_name):
#Gaussian fit at a given energy
sum_edep = inp[sec_name]
nbins = 60
plt.style.use("dark_background")
col = "lime"
#col = "black"
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
set_axes_color(ax, col)
#data plot
hx = plt.hist(sum_edep,
bins=nbins,
color="lime",
density=True,
label="edep")
#bin centers for the fit
centers = (0.5 * (hx[1][1:] + hx[1][:-1]))
#pass1, fit over the full range
fit_data = DataFrame({"E": centers, "density": hx[0]})
pars, cov = curve_fit(lambda x, mu, sig: norm.pdf(x, loc=mu, scale=sig),
fit_data["E"], fit_data["density"])
#print "pass1:", pars[0], pars[1]
#pass2, fit in +/- 2*sigma range
fitran = [pars[0] - 2. * pars[1],
pars[0] + 2. * pars[1]] # fit range at 2*sigma
fit_data = fit_data[fit_data["E"].between(
fitran[0], fitran[1], inclusive=False)] # select the data to the range
pars, cov = curve_fit(lambda x, mu, sig: norm.pdf(x, loc=mu, scale=sig),
fit_data["E"], fit_data["density"])
#print "pass2:", pars[0], pars[1]
#fit function
x = np.linspace(fitran[0], fitran[1], 300)
y = norm.pdf(x, pars[0], pars[1])
plt.plot(x, y, "k-", label="norm", color="red")
ax.set_xlabel("Calorimeter signal (GeV)")
ax.set_ylabel("Counts / {0:.3f} GeV".format(
(plt.xlim()[1] - plt.xlim()[0]) / nbins))
#ax.set_title(r"$\text{"+infile+"}$")
#print infile
set_grid(plt, col)
#mean_str = "{0:.4f} \pm {1:.4f}".format(pars[0], np.sqrt(cov[0,0]))
#sigma_str = "{0:.4f} \pm {1:.4f}".format(pars[1], np.sqrt(cov[1,1]))
#res = pars[1]/pars[0]
#res_str = "{0:.4f}".format(res)
#fit_param = r"\begin{align*}\mu &= " + mean_str + r"\\ \sigma &= " + sigma_str + r"\\"
#fit_param += r"\sigma/\mu &= " + res_str + r"\end{align*}"
#plot legend
#leg_items = [Line2D([0], [0], lw=2, color="red"), Line2D([0], [0], lw=0)]
leg_items = [Line2D([0], [0], lw=0)]
#plt.rc("text", usetex = True)
#plt.rc('text.latex', preamble='\usepackage{amsmath}')
#ax.legend(leg_items, ["Gaussian fit", fit_param])
ax.legend(leg_items, [
"{0:.4f}, {1:.4f}, {2:.4f}".format(pars[0], pars[1], pars[1] / pars[0])
])
#output log
#out = open("out.txt", "w")
#out.write( "{0:.4f} +/- {1:.4f} | ".format(pars[0], np.sqrt(cov[0,0])) )
#out.write( "{0:.4f} +/- {1:.4f} | ".format(pars[1], np.sqrt(cov[1,1])) )
#out.write( "{0:.4f}".format(pars[1]/pars[0]) )
#out.close()
fig.savefig("01fig.pdf", bbox_inches="tight")
plt.close()
return pars[0], pars[1]
def likelihood(x, vol):
sd = np.exp(vol)
prob = 0
for i in norm.rvs(loc=mean_mu, scale=mean_sd, size=100):
prob = prob + norm.pdf(x, i, sd) / 100
return prob
def pOfXpdf(x, feature):
sigma = standardDeviation(feature)
mean = sampleMean(feature)
prob = norm.pdf(x, loc=mean, scale=sigma)
return prob
summ = sum(diff)
# print ("SUM - ",summ)
summ /= (n * h * 1.0)
y.append(summ)
# print ("Prazen " - y)
return y
#Q-7-a
# getting alphas
X = np.arange(-5, 10, 0.1)
alpha = X
pr_dis = get_data()
# True distribution
Y0 = 0.25 * (norm.pdf(alpha, 0, 1) + norm.pdf(alpha, 3, 1) +
norm.pdf(alpha, 6, 1) + norm.pdf(alpha, 9, 1))
# get kernel density estimator
Y = get_prazen(alpha, 0.1, pr_dis)
Y2 = get_prazen(alpha, 1, pr_dis)
Y3 = get_prazen(alpha, 7, pr_dis)
# plotting the graphs
plt.figure('Kernel Density Estimation', figsize=(15, 10))
plt.plot(X, Y0, label='True pdf')
plt.plot(X, Y, label='h = 0.1')
plt.plot(X, Y2, label='h = 1')
plt.plot(X, Y3, label='h = 7')
plt.xlabel('x', fontsize=18)
plt.ylabel('Pr[X<=x]', fontsize=18)
def bayesian_calculation_update(self, sensor_inputs, time_frame, ITER):
error_val_list = []
mse_list = []
self.create_error_lists(time_frame)
err_names = []
probs = []
errors = []
error_names = []
self.search_helper({}, {}, errors, error_names)
# need to make sure all the lists do not have conflicts
errors, error_names = self.check_conflicts(errors, error_names)
for i in range(len(errors)):
err_name = tuple(
set([
item for sublist in error_names[i].values()
for item in sublist
]))
err_val = errors[i]
full_err_name = error_names[i]
if self.fft is True:
for sensor in self.sensor_list:
err_val[sensor] = scipy.ifft(err_val[sensor])
for key, value in self.sirens.items():
if full_err_name[key] in value.keys():
data_size = self.stored_data[key][
full_err_name[key]][1].shape[0]
val = self.error_calculation(error_list=full_err_name[key],
sensor=key,
x=time_frame)
err_val[key] = (1 - 1 / np.log(data_size)) * val + (
1 / np.log(data_size)) * err_val[key]
#if err_name == ('Error2',):
#plt.clf()
#plt.plot(time_frame, err_val[key], color= "red", label = "Estimated Line")
#plt.plot(time_frame, sensor_inputs[key], color = "green", label = "Actual Line")
#plt.xlabel("Timestamp")
#plt.ylabel("Line Val")
#plt.title(str(ITER))
#plt.legend()
#plt.grid()
#plt.show()
error_val_list.append(err_val)
prob = 0.0
mses = {}
for j in self.sensor_list:
for i in range(len(sensor_inputs[j])):
gaussian_err = norm.pdf(sensor_inputs[j][i],
loc=err_val[j][i],
scale=self.sigmas[j])
prob += np.log(gaussian_err)
mse = ((sensor_inputs[j] - err_val[j])**2).mean()
mses[j] = mse
err_names.append(err_name)
probs.append(prob)
mse_list.append(mses)
ind, error, pro, mse = self.find_most_probable_error(
err_names, probs, mse_list)
return error, pro, mse, error_val_list[ind]
def energy_landscape_example():
fig, [ax] = panel(1, 1, l_p=0, r_p=0, b_p=0, t_p=0, dpi=100)
rectangle(-1000, 1000, -1000, 1000, fc='g', ec='m', lw=4)
plt.axis('off')
xlim([-1000, 1000])
ylim([-1000, 1000])
from scipy.stats import norm
np.random.seed(584111)
x = np.arange(-1000, 1000, 1)
y = 50 * np.sin(0.02 * x) + 50 * np.sin(0.05 * x) - 300 * np.sin(
0.007 * x) + 500
for i in range(1, 11):
y -= np.random.random() * 2000 * norm.pdf((-1000 + 300 * i + x) / 20)
y[0], y[-1] = -1000, -1000
ax.fill(x, y, c='m', alpha=0.2)
ax.plot(x, y, c='m', alpha=0.5)
pos = [-670]
for i in pos:
xn = np.argmax(i < x)
ax.arrow(x[xn] + 30 + 100,
y[xn] + 30,
-50,
0,
width=15,
fc='grey',
ec='grey')
ax.arrow(x[xn] - 30 - 100,
y[xn] + 30,
50,
0,
width=15,
fc='grey',
ec='grey')
ellipse([x[xn], y[xn] + 30], 20, alpha=1, fc='b')
pos = [320]
for i in pos:
xn = np.argmax(i < x)
ax.arrow(x[xn] + 30, y[xn] + 30, 50, 0, width=15, fc='grey', ec='grey')
ax.arrow(x[xn] - 30,
y[xn] + 30,
-50,
0,
width=15,
fc='grey',
ec='grey')
ellipse([x[xn], y[xn] + 30], 20, alpha=1, fc='b')
text(400, y[1400] - 100, r'\textbf{Global minima}', ha='center')
text(250 + 100, y[1250] - 300, r'\textbf{Local minima}', ha='right')
ax.arrow(
225,
y[1225] - 200,
0,
130,
width=15,
fc='k',
)
arrow(
pos=[[-300, y[1250] - 300], [x[330], y[330]]],
curve=-0.5,
)
sample_std = np.std(x_sample, ddof=1)
#------------------------------------------------------------
# Plot the sampled data
fig, ax = plt.subplots(figsize=(5, 3.75))
ax.hist(x_sample, 20, histtype='stepfilled', normed=True, fc='#CCCCCC')
x = np.linspace(-2.1, 4.1, 1000)
factor1 = ratio_in / (1. + ratio_in)
factor2 = 1. / (1. + ratio_in)
ax.plot(x, gm.pdf(x), '-k', label='true distribution')
ax.plot(x, gm.pdf_individual(x), ':k')
ax.plot(x, norm.pdf(x, sample_mu, sample_std), '--k', label='best fit normal')
ax.legend(loc=1)
ax.set_xlim(-2.1, 4.1)
ax.set_xlabel('$x$')
ax.set_ylabel('$p(x)$')
ax.set_title('Input pdf and sampled data')
ax.text(0.95,
0.80, ('$\mu_1 = 0;\ \sigma_1=0.3$\n'
'$\mu_2=1;\ \sigma_2=1.0$\n'
'$\mathrm{ratio}=1.5$'),
transform=ax.transAxes,
ha='right',
va='top')
plt.rc('figure', titlesize=BIGGER_SIZE) # fontsize of the figure title
plt.close('all')
plt.figure(1, figsize=(15, 7))
x = [(i + 1) * 0.1 for i in range(9)]
y = [norm.ppf(e) for e in x]
plt.subplot(121)
plt.plot(x, y, '-*b')
plt.ylabel('percentile', labelpad=18)
plt.xlabel(r'$\alpha$', labelpad=18)
# plot our distribution
xx = np.linspace(norm.ppf(0.01),
norm.ppf(0.99), 100)
plt.figure(1)
plt.subplot(122)
plt.plot(xx, norm.pdf(xx),
'r-', lw=2, alpha=0.6)
plt.ylim([0, 0.5])
for e in y:
print(norm.pdf(e))
plt.axvline(e, 0, norm.pdf(e) / 0.5,
color='black', alpha=0.5)
plt.ylabel('$f(x)$', labelpad=18)
plt.xlabel('$x$', labelpad=18)
plt.savefig('percentile_normal.png',
bbox_inches='tight', dpi=300)
amp = np.max(px_spec_avg)
new_dict[pixel] = amp
# try baseline subtraction (from Brad's Renishaw-Reniawesome)
base = baseline(px_spec[:,1], 4)
plt.plot(base)
base_sub = px_spec[:,1]-base
plt.plot(base_sub)
# gaussian fit try 1
popt, _ = optimize.curve_fit(gaussian, px_spec[:,0], px_spec[:,1])
plt.plot(px_spec[:,0], px_spec[:,1])
plt.plot(px_spec[:,0], gaussian(px_spec[:,0], *popt))
# gaussian fit try 2
mean,std=norm.fit(px_spec[:,1])
px_fit = norm.pdf(px_spec[:,0], mean, std)
plt.plot(px_spec[:,0], px_spec[:,1])
plt.plot(px_spec[:,0], px_fit)
# gaussian fit try 3
fitter = modeling.fitting.LevMarLSQFitter()
model = modeling.models.Gaussian1D() # depending on the data you need to give some initial values
fitted_model = fitter(model, px_spec[:,0], px_spec[:,1])
plt.plot(px_spec[:,0], px_spec[:,1])
plt.plot(px_spec[:,0], fitted_model(px_spec[:,0]))
#pixel_spectrum_arr_5avg = moving_average(pixel_spectrum_arr[:,1], 5)
# copy dictionary key
# set value of copy dictionary key to max amplitude
# store this key-value pair in new dictionary
def get_emission(self, n, state):
X = self.X
phi = self.phi
prior = norm.pdf(X[n], phi[state][0], phi[state][1])
return (prior)
import os
figdir = os.path.join(os.environ["PYPROBML"], "figures")
def save_fig(fname): plt.savefig(os.path.join(figdir, fname))
from scipy.stats import t, laplace, norm
a = np.random.randn(30)
outliers = np.array([8, 8.75, 9.5])
plt.hist(a, 7, weights=[1 / 30] * 30, rwidth=0.8)
#fit without outliers
x = np.linspace(-5, 10, 500)
loc, scale = norm.fit(a)
n = norm.pdf(x, loc=loc, scale=scale)
loc, scale = laplace.fit(a)
l = laplace.pdf(x, loc=loc, scale=scale)
fd, loc, scale = t.fit(a)
s = t.pdf(x, fd, loc=loc, scale=scale)
plt.plot(x, n, 'k>',
x, s, 'r-',
x, l, 'b--')
plt.legend(('Gauss', 'Student', 'Laplace'))
save_fig('robustDemoNoOutliers.pdf')
#add the outliers
plt.figure()
plt.hist(a, 7, weights=[1 / 33] * 30, rwidth=0.8)
def getNegLogProbNorm(Param):
avg = Param[0]
std = Param[1]
getNegLogProbNorm = -np.sum(np.log(norm.pdf(x=Data, loc=avg, scale=std)))
return getNegLogProbNorm
def editsolvent_removal2(solvent, y_data, x_data, picked_peaks, peak_regions, grouped_peaks, total_params, uc):
picked_peaks = np.array(picked_peaks)
# define the solvents
if solvent == 'chloroform':
exp_ppm = [7.26]
Jv = [[]]
elif solvent == 'dimethylsulfoxide':
exp_ppm = [2.50]
Jv = [[1.9, 1.9]]
elif solvent == 'methanol':
exp_ppm = [4.78, 3.31]
Jv = [[], [1.7, 1.7]]
elif solvent == 'benzene':
exp_ppm = [7.16]
Jv = [[]]
elif solvent == 'pyridine':
exp_ppm = [8.74, 7.58, 7.22]
Jv = [[], [], []]
else:
exp_ppm = []
Jv = [[]]
# find picked peaks ppm values
picked_peaks_ppm = x_data[picked_peaks]
# make differences vector for referencing against multiple solvent peaks
differences = []
peaks_to_remove = []
solvent_regions = []
# now remove each peak in turn
for ind1, speak_ppm in enumerate(exp_ppm):
# if only a singlet is expected for this peak find solvent peak based on amplitude and position
if len(Jv[ind1]) == 0:
probs = norm.pdf(abs(picked_peaks_ppm - speak_ppm), loc=0, scale=0.1) * y_data[picked_peaks]
# find the maximum probability
w = np.argmax(probs)
# append this to the list to remove
peaks_to_remove.append(picked_peaks[w])
# append this to the list of differences
differences.append(speak_ppm - picked_peaks_ppm[w])
# if the peak displays a splitting pattern then we have to be a bit more selective
# do optimisation problem with projected peaks
else:
amp_res = []
dist_res = []
pos_res = []
# limit the search to peaks +- 1 ppm either side
srange = (picked_peaks_ppm > speak_ppm - 1) * (picked_peaks_ppm < speak_ppm + 1)
for peak in picked_peaks_ppm[srange]:
# print("picked ppm ", peak)
fit_s_peaks, amp_vector, fit_s_y = new_first_order_peak(peak, Jv[ind1], np.arange(len(x_data)), 0.1, uc,
1)
diff_matrix = np.zeros((len(fit_s_peaks), len(picked_peaks)))
for i, f in enumerate(fit_s_peaks):
for j, g in enumerate(picked_peaks):
diff_matrix[i, j] = abs(f - g)
# minimise these distances
vertical_ind, horizontal_ind = optimise(diff_matrix)
closest_peaks = np.sort(picked_peaks[horizontal_ind])
closest_amps = []
for cpeak in closest_peaks:
closest_amps.append(total_params['A' + str(cpeak)])
# find the amplitude residual between the closest peaks and the predicted pattern
# normalise these amplitudes
amp_vector = [i / max(amp_vector) for i in amp_vector]
closest_amps = [i / max(closest_amps) for i in closest_amps]
# append to the vector
amp_res.append(sum([abs(amp_vector[i] - closest_amps[i]) for i in range(len(amp_vector))]))
dist_res.append(np.sum(np.abs(closest_peaks - fit_s_peaks)))
pos_res.append(norm.pdf(abs(peak - speak_ppm), loc=0, scale=0.5))
# use the gsd data to find amplitudes of these peaks
pos_res = [1 - i / max(pos_res) for i in pos_res]
dist_res = [i / max(dist_res) for i in dist_res]
amp_res = [i / max(amp_res) for i in amp_res]
# calculate geometric mean of metrics for each peak
g_mean = [(dist_res[i] + amp_res[i] + pos_res[i]) / 3 for i in range(0, len(amp_res))]
# compare the residuals and find the minimum
minres = np.argmin(g_mean)
# append the closest peaks to the vector
fit_s_peaks, amp_vector, fit_s_y = new_first_order_peak(picked_peaks_ppm[srange][minres], Jv[ind1],
np.arange(len(x_data)), 0.1, uc, 1)
diff_matrix = np.zeros((len(fit_s_peaks), len(picked_peaks)))
for i, f in enumerate(fit_s_peaks):
for j, g in enumerate(picked_peaks):
diff_matrix[i, j] = abs(f - g)
# minimise these distances
vertical_ind, horizontal_ind = optimise(diff_matrix)
closest_peaks = np.sort(picked_peaks[horizontal_ind])
for peak in closest_peaks:
ind3 = np.abs(picked_peaks - peak).argmin()
peaks_to_remove.append(picked_peaks[ind3])
differences.append(picked_peaks_ppm[ind3] - uc.ppm(peak))
# find the region this peak is in and append it to the list
for peak in peaks_to_remove:
for ind2, region in enumerate(peak_regions):
if (peak > region[0]) & (peak < region[-1]):
solvent_regions.append(ind2)
break
# now remove the selected peaks from the picked peaks list and grouped peaks
w = np.searchsorted(picked_peaks, peaks_to_remove)
picked_peaks = np.delete(picked_peaks, w)
for ind4, peak in enumerate(peaks_to_remove):
grouped_peaks[solvent_regions[ind4]] = np.delete(grouped_peaks[solvent_regions[ind4]],
np.where(grouped_peaks[solvent_regions[ind4]] == peak))
# resimulate the solvent regions
solvent_region_ind = sorted(list(set(solvent_regions)))
# now need to reference the spectrum
# differences = list of differences in ppm found_solvent_peaks - expected_solvent_peaks
s_differences = sum(differences)
x_data = x_data + s_differences
return peak_regions, picked_peaks, grouped_peaks, x_data, solvent_region_ind
#!/usr/bin/env python3
import numpy as np
from scipy.stats import norm
import matplotlib.pyplot as plt
plt.style.use('seaborn-darkgrid')
# compute cumulative distribution function (CDF): P(-1.45 < Z < 1.45)
x = -1.45; y = 1.45; loc = 0; scale = 1
pls = norm.cdf(y, loc, scale) - norm.cdf(x, loc, scale)
print("Area under the standard normal curve P(-1.45 < Z < 1.45): {:.5f}".format(pls))
x1=np.linspace(x,y, 1000); y1=norm.pdf(x1,loc,scale)
x2=np.linspace(-4,4,1000); y2=norm.pdf(x2,loc,scale)
ax = plt.figure().add_subplot(111); ax.minorticks_on()
ax.plot(x2,y2,c='r',label='P(-1.45 < Z < 1.45) = {:.5f}'.format(pls))
ax.set(xlabel='X',ylabel='P(X)'); ax.legend(handlelength=0)
ax.fill_between(x1,y1,alpha=0.5,color='r'); ax.set(xlim=[-4,4])
plt.savefig('norm1f.pdf',dpi=72,bbox_inches='tight'); plt.show()
def test_probit():
for Y, T in [(np.random.binomial(1, 0.5, size=(10, )), np.ones(10)),
(np.random.binomial(1, 0.5, size=(10, )), None),
(np.random.binomial(3, 0.5, size=(10, )), 3 * np.ones(10))]:
X = np.random.standard_normal((10, 5))
for case_weights in [None, np.ones(10)]:
L = glm.glm.probit(X, Y, trials=T, case_weights=case_weights)
L.smooth_objective(np.zeros(L.shape), 'both')
L.hessian(np.zeros(L.shape))
sat_sub = L.saturated_loss.subsample(np.arange(
5)) # check that subsample of saturated loss at least works
sat_sub.smooth_objective(np.zeros(sat_sub.shape))
# check that subsample is getting correct answer
Xsub = X[np.arange(5)]
Ysub = Y[np.arange(5)]
if T is not None:
Tsub = T[np.arange(5)]
T_num = T
else:
Tsub = np.ones(5)
T_num = np.ones(10)
beta = np.ones(L.shape)
if case_weights is not None:
Lsub2 = glm.glm.probit(Xsub,
Ysub,
trials=Tsub,
case_weights=case_weights[np.arange(5)])
Lsub3 = glm.glm.probit(Xsub,
Ysub,
trials=Tsub,
case_weights=case_weights[np.arange(5)])
case_cp = case_weights.copy() * 0
case_cp[np.arange(5)] = 1
Lsub4 = glm.glm.probit(X, Y, trials=T, case_weights=case_cp)
else:
Lsub2 = glm.glm.probit(Xsub, Ysub, trials=Tsub)
Lsub3 = glm.glm.probit(Xsub, Ysub, trials=Tsub)
Lsub3.coef *= 2.
f2, g2 = Lsub2.smooth_objective(beta, 'both')
f3, g3 = Lsub3.smooth_objective(beta, 'both')
f4, g4 = Lsub2.smooth_objective(beta, 'both')
np.testing.assert_allclose(f3, 2 * f2)
np.testing.assert_allclose(g3, 2 * g2)
np.testing.assert_allclose(f2, f4)
np.testing.assert_allclose(g2, g4)
Lcp = copy(L)
prev_value = L.smooth_objective(np.zeros(L.shape), 'func')
L_sub = L.subsample(np.arange(5))
L_sub.coef *= 45
new_value = L.smooth_objective(np.zeros(L.shape), 'func')
assert (prev_value == new_value)
np.testing.assert_allclose(L_sub.gradient(beta),
45 * Lsub2.gradient(beta))
linpred = X.dot(beta)
np.testing.assert_allclose(
L.gradient(beta),
X.T.dot(-normal_dbn.pdf(linpred) *
(Y / normal_dbn.cdf(linpred) -
(T_num - Y) / normal_dbn.sf(linpred))))
linpred = Xsub.dot(beta)
np.testing.assert_allclose(
L_sub.gradient(beta),
45 * Xsub.T.dot(-normal_dbn.pdf(linpred) *
(Ysub / normal_dbn.cdf(linpred) -
(Tsub - Ysub) / normal_dbn.sf(linpred))))
# other checks on gradient
if T is None:
sat = L.saturated_loss
np.testing.assert_allclose(
sat.smooth_objective(np.zeros(sat.shape), 'grad'),
(0.5 - Y) * normal_dbn.pdf(0) / 0.25)
np.testing.assert_allclose(
L.gradient(np.zeros(L.shape)),
X.T.dot(0.5 - Y) * normal_dbn.pdf(0) / 0.25)
np.testing.assert_allclose(
L.hessian(np.zeros(L.shape)),
X.T.dot(X) / 0.25 * normal_dbn.pdf(0)**2)
else:
L.gradient(np.zeros(L.shape))
L.hessian(np.zeros(L.shape))
L.objective(np.zeros(L.shape))
L.latexify()
L.saturated_loss.data = (Y, T)
L.saturated_loss.data
L.data = (X, (Y, T))
L.data
from scipy.stats import norm
from scipy.misc import comb
from scipy.stats import beta
from scipy import integrate
import matplotlib.pyplot as plt
import numpy
import math
import pylab
alpha=[5,3,1] # define alpha pars for inverse gamma dist
betaa=[5,15,30] # define beta pars for inverse gamma dist
y_value=[10,3,3] # number of heads
N=[20,10,10] #number of tosses
x=numpy.linspace(-0.25,1,100) # intervals for r ( it can't be megative, but negative r is considered for visul purposeis )
for i in range( len(alpha)):
a=' Inverse Gamma :alpha={} beta={}'.format(alpha[i],betaa[i])
b=' Laplace Approx :alpha={} beta={}'.format(alpha[i],betaa[i])
real_dist= lambda x : comb(N[i],y_value[i])*(x**y_value[i])*((1-x)**(N[i]-y_value[i]))*beta.pdf(x,alpha[i],betaa[i]) # prior * liklihood
marginal_liklihood=integrate.quad(real_dist, x[0], x[-1])[0] # calculating marginal liklihood (i.e, normalization tep)
map_r=(y_value[i]+alpha[i]-1)/(alpha[i]+N[i]+betaa[i]-2) # MAP for r
print('r={}',format(map_r))
g_2d=((-alpha[i]-y_value[i]+1)/(map_r**2))+((-betaa[i]-N[i]+y_value[i]+1)/(1-map_r)**2) # calculating hessian matrix
variance_laplace=-(1/g_2d)
plt.plot(x,norm.pdf(x,map_r, math.sqrt(variance_laplace)),label=b)
print('sigma={}',format(math.sqrt(variance_laplace)))
plt.plot(x,real_dist(x)/marginal_liklihood,label=a)
pylab.legend(loc='upper right')
#normal_pdf=norm.pdf(x,map_r, math.sqrt(variance_laplace))
# real_dist=comb(N[i],y_value[i])*(x**y_value[i])*((1-x)**(N[i]-y_value[i]))*beta.pdf(x,alpha[i],betaa[i])
plt.xlabel('r value')
plt.ylabel('p(r)')
def out(w1, w2):
err = y0 - (w1 + w2 * x0)
return norm.pdf(err, loc=0, scale=likelihoodSD)
fig, ax = plt.subplots()
ax.plot(x, y, 'r-', linewidth=2, label='sine function', alpha=0.9)
ax.legend(loc='upper center')
plt.show()
fig, ax = plt.subplots()
ax.plot(x, y, 'r-', linewidth=2, label='$y=\sin(x)$', alpha=0.6)
ax.legend(loc='upper center')
ax.set_yticks([-1,-.5, 0, .5, 1])
plt.title('Test plot')
plt.show()
##More plot on one axis
from scipy.stats import norm
from random import uniform
fig, ax = plt.subplots()
x = np.linspace(-4, 4, 150)
for i in range(3):
print(i)
m, s = uniform(-1, 1), uniform(1, 2)
y = norm.pdf(x, loc=m, scale=s)
current_label = '$\mu = {m:.2}$'
ax.plot(x, y, linewidth=2, alpha=0.6, label=current_label)
ax.legend()
plt.show()
def precompute_sensor_model(self):
print "Precomputing sensor model"
table_width = int(self.MAX_RANGE_PX) + 1
#meshgrid of data
(x, y) = np.meshgrid(np.linspace(0, self.MAX_RANGE_PX, table_width),
np.linspace(0, self.MAX_RANGE_PX, table_width))
#normal along identity
z = 2 * norm.pdf(x, y, 5)
#uniform
z += 2.0 / self.MAX_RANGE_PX
#ramp
for row in xrange(table_width):
z[row][0:row] += .01 - .01 * np.arange(row, dtype=np.float32) / row
#max_dist
z[:, -1:] = .3
#normalize
#for i in range(len(z)):
# z[i]= z[i]/sum(z[i])
z / z.sum(axis=1, keepdims=True)
#transpose and save it had to use ascontiguousarray for cpython to be happy
self.sensor_model_table = np.ascontiguousarray(z.T)
# upload the sensor model to RangeLib for ultra fast resolution later
self.range_method.set_sensor_model(self.sensor_model_table)
# code to generate visualizations of the sensor model
if False:
# visualize the sensor model
fig = plt.figure()
ax = fig.gca(projection='3d')
# Make data.
X = np.arange(0, table_width, 1.0)
Y = np.arange(0, table_width, 1.0)
X, Y = np.meshgrid(X, Y)
# Plot the surface.
surf = ax.plot_surface(X,
Y,
self.sensor_model_table,
cmap="bone",
rstride=2,
cstride=2,
linewidth=0,
antialiased=True)
ax.text2D(0.05,
0.95,
"Precomputed Sensor Model",
transform=ax.transAxes)
ax.set_xlabel('Ground truth distance (in px)')
ax.set_ylabel('Measured Distance (in px)')
ax.set_zlabel('P(Measured Distance | Ground Truth)')
plt.show()
elif False:
plt.imshow(self.sensor_model_table * 255, cmap="gray")
plt.show()
elif False:
plt.plot(self.sensor_model_table[:, 140])
plt.plot([139, 139], [0.0, 0.08], label="test")
plt.ylim(0.0, 0.08)
plt.xlabel("Measured Distance (in px)")
plt.ylabel("P(Measured Distance | Ground Truth Distance = 140px)")
plt.show()
def prob(val, mu, sig, lam):
p = lam
for i in range(len(val)):
p *= norm.pdf(val[i], mu[i], sig[i][i])
return p
def sigmoid(x):
return 1 / (1 + np.exp(-x))
nT = 30 # Temporal filter length
nX = 40 # Spatial filter length
N = 30 # Number of neurons
temp_filt = np.zeros((N, nT))
spat_filt = np.zeros((N, nX))
# Define filters manually
for i in range(N - 15):
temp_filt[i, :] = norm.pdf(np.linspace(-15, 15, nT),
loc=i - 12.5 + i / 1.5,
scale=.8)
spat_filt[i, :] = norm.pdf(
np.linspace(-10, 10, nX), i - 7.5,
scale=.8) - .8 * norm.pdf(np.linspace(-10, 10, nX), i - 7.5, 1)
for i in range(N - 15, N):
temp_filt[i, :] = norm.pdf(
np.linspace(-5, 5, nT), loc=.45 * i - 10, scale=.8) - .6 * norm.pdf(
np.linspace(-5, 5, nT), loc=.45 * i - 9, scale=.8)
spat_filt[i, :] = norm.pdf(
np.linspace(-15, 15, nX), loc=1.5 * i - 33, scale=.7) - .5 * norm.pdf(
np.linspace(-15, 15, nX), loc=1.5 * i - 33, scale=1.5)
bias = -3 - 2 * np.random.uniform(size=N) # Bias
print(bias)
for iteration in range(0, 250):
# We need to have observed at least 3 items for the model to be able to predict
surr_predictions = np.zeros_like(test_index)
if iteration > 2 and alpha < 1:
surr_estimator.train(
np.array(observed_X).astype(float),
np.array(observed_y))
mu, var = surr_estimator.predict(
np.array(surr_X.iloc[test_index]).astype(float))
mu = mu.reshape(-1)
var = var.reshape(-1)
sigma = np.sqrt(var)
diff = mu - np.max(observed_y)
Z = diff / sigma
ei = diff * norm.cdf(Z) + sigma * norm.pdf(Z)
surr_predictions = ei
# surr_predictions = surr_estimator.predict(np.array(surr_X.iloc[test_index]).astype(float))
# print(iteration, "\t", np.std(surr_predictions), "\t", np.std(meta_predictions))
m_corr, m_pvalue = kendalltau(meta_predictions, y[test_index])
s_corr, s_pvalue = kendalltau(surr_predictions, y[test_index])
if s_corr > m_corr:
surpassed = iteration if surpassed is None else surpassed
# alpha == 0: Only surrogate predictions
# alpha == 1: Only meta-model predictions
corrected_iteration = np.maximum(1, iteration - 2)
# scaled_meta_predictions = MinMaxScaler().fit_transform(meta_predictions.reshape(-1, 1)).reshape(-1)
r_list = np.random.random(N) # list with N random numbers between 0 and 1
for i in range(N):
if r_list[i] >= 0.5:
part_pos_list[i] += h # One step to the right
else:
part_pos_list[i] -= h # One step to the left
# find standard deviation mu and variance sigma to the normal distribution best suited to part_pos_list
mu, sigma = norm.fit(part_pos_list)
print("mu =", mu, "sigma =", sigma)
# pre-plotting
xMax = np.max(np.abs(part_pos_list)) # maximum absolute x position value
xRange = (-xMax * 1.1, xMax * 1.1) # Range for the plot
xAx = np.linspace(*xRange, 1000) # list og x values for normal distribution
p = norm.pdf(xAx, mu, sigma) # normal distribution
# plotting
savename = "RandomWalkIn1D"
fig, ax = plt.subplots(1, 1, num=savename)
# new axis for p
ax2 = ax.twinx()
# Set ax's patch invisible
ax.patch.set_visible(False)
# move ax in front
ax.set_zorder(ax2.get_zorder() + 1)
# Histogram, bins is given as the number of possible positions for a particle
# distribution=False because True will mess with scaling
hist, bins = np.histogram(part_pos_list,
bins=(2 * Ntime + 1),
density=True,
def EI(mean, std, max_val, tradeoff):
z = (mean - max_val - tradeoff) / std
return (mean - max_val - tradeoff) * ndtr(z) + std * norm.pdf(z)
# mean - get_error_conf(sample_size, confidence, std), mean + get_error_conf(sample_size, confidence, std)
interval = norm.interval(alpha=confidence,
loc=mean,
scale=std / np.sqrt(sample_size))
return interval
if __name__ == "__main__":
""" If the module is called as script, plot the probability density function
and the cumulative distribution function.
Modified from: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.norm.html
"""
from scipy.stats import norm
import matplotlib.pyplot as plt
import numpy as np
print('Plotting the Probability Density Function' +
'\n and the Cumulative Density Function')
x = np.linspace(norm.ppf(0.001), norm.ppf(0.999), 100)
plt.subplot(1, 2, 1)
plt.plot(x, norm.pdf(x), 'r-', lw=5, alpha=0.6)
plt.title("Probability Density Function")
plt.xlabel("norm.ppf(0.001) <= x <= norm.ppf(0.999)")
plt.ylabel("norm.pdf(x)")
plt.subplot(1, 2, 2)
plt.plot(x, norm.cdf(x), 'b-', lw=5, alpha=0.6)
plt.title("Cumulative Density Function")
plt.xlabel("norm.ppf(0.001) <= x <= norm.ppf(0.999)")
plt.ylabel("norm.cdf(x)")
plt.show()
def detect_knee(data, window_size=1, s=10):
"""
Detect the so-called knee in the data.
The implementation is based on paper [1] and code here (https://github.com/jagandecapri/kneedle).
@param data: The 2d data to find an knee in.
@param window_size: The data is smoothed using Gaussian kernel average smoother, this parameter is the window used for averaging (higher values mean more smoothing, try 3 to begin with).
@param s: How many "flat" points to require before we consider it a knee.
@return: The knee values.
"""
data_size = len(data)
data = np.array(data)
if data_size == 1:
return None
# smooth
smoothed_data = []
for i in range(data_size):
if 0 < i - window_size:
start_index = i - window_size
else:
start_index = 0
if i + window_size > data_size - 1:
end_index = data_size - 1
else:
end_index = i + window_size
sum_x_weight = 0
sum_y_weight = 0
sum_index_weight = 0
for j in range(start_index, end_index):
index_weight = norm.pdf(abs(j - i) / window_size, 0, 1)
sum_index_weight += index_weight
sum_x_weight += index_weight * data[j][0]
sum_y_weight += index_weight * data[j][1]
smoothed_x = sum_x_weight / sum_index_weight
smoothed_y = sum_y_weight / sum_index_weight
smoothed_data.append((smoothed_x, smoothed_y))
smoothed_data = np.array(smoothed_data)
# normalize
normalized_data = MinMaxScaler().fit_transform(smoothed_data)
# difference
differed_data = [(x, y - x) for x, y in normalized_data]
# find indices for local maximums
candidate_indices = []
for i in range(1, data_size - 1):
if (differed_data[i - 1][1] < differed_data[i][1]) and (
differed_data[i][1] > differed_data[i + 1][1]):
candidate_indices.append(i)
# threshold
step = s * (normalized_data[-1][0] - data[0][0]) / (data_size - 1)
# knees
knee_indices = []
for i in range(len(candidate_indices)):
candidate_index = candidate_indices[i]
if i + 1 < len(candidate_indices): # not last second
end_index = candidate_indices[i + 1]
else:
end_index = data_size
threshold = differed_data[candidate_index][1] - step
for j in range(candidate_index, end_index):
if differed_data[j][1] < threshold:
knee_indices.append(candidate_index)
break
if knee_indices != []:
return knee_indices #data[knee_indices]
else:
return None
def gaussian_ei(X, model, y_opt=0.0, xi=0.01, return_grad=False):
"""
Use the expected improvement to calculate the acquisition values.
The conditional probability `P(y=f(x) | x)`form a gaussian with a certain
mean and standard deviation approximated by the model.
The EI condition is derived by computing ``E[u(f(x))]``
where ``u(f(x)) = 0``, if ``f(x) > y_opt`` and ``u(f(x)) = y_opt - f(x)``,
if``f(x) < y_opt``.
This solves one of the issues of the PI condition by giving a reward
proportional to the amount of improvement got.
Note that the value returned by this function should be maximized to
obtain the ``X`` with maximum improvement.
Parameters
----------
* `X` [array-like, shape=(n_samples, n_features)]:
Values where the acquisition function should be computed.
* `model` [sklearn estimator that implements predict with ``return_std``]:
The fit estimator that approximates the function through the
method ``predict``.
It should have a ``return_std`` parameter that returns the standard
deviation.
* `y_opt` [float, default 0]:
Previous minimum value which we would like to improve upon.
* `xi`: [float, default=0.01]:
Controls how much improvement one wants over the previous best
values. Useful only when ``method`` is set to "EI"
* `return_grad`: [boolean, optional]:
Whether or not to return the grad. Implemented only for the case where
``X`` is a single sample.
Returns
-------
* `values`: [array-like, shape=(X.shape[0],)]:
Acquisition function values computed at X.
"""
with warnings.catch_warnings():
warnings.simplefilter("ignore")
if return_grad:
mu, std, mu_grad, std_grad = model.predict(X,
return_std=True,
return_mean_grad=True,
return_std_grad=True)
else:
mu, std = model.predict(X, return_std=True)
values = np.zeros_like(mu)
mask = std > 0
improve = y_opt - xi - mu[mask]
scaled = improve / std[mask]
cdf = norm.cdf(scaled)
pdf = norm.pdf(scaled)
exploit = improve * cdf
explore = std[mask] * pdf
values[mask] = exploit + explore
if return_grad:
if not np.all(mask):
return values, np.zeros_like(std_grad)
# Substitute (y_opt - xi - mu) / sigma = t and apply chain rule.
# improve_grad is the gradient of t wrt x.
improve_grad = -mu_grad * std - std_grad * improve
improve_grad /= std**2
cdf_grad = improve_grad * pdf
pdf_grad = -improve * cdf_grad
exploit_grad = -mu_grad * cdf - pdf_grad
explore_grad = std_grad * pdf + pdf_grad
grad = exploit_grad + explore_grad
return values, grad
return values
b: upper bound of the interval of test values
n: number of test values
name: name of the test to be printed by the test program
c_fn_name: name of the C function being tested
args: additional arguments of the C function being tested
"""
TestParams = namedtuple('TestParams', 'a b n f name c_fn_name args'.split())
test_params_list = [
TestParams(-1, 2, 100, uniform.pdf, 'uniform(0, 1) pdf', 'uniform_pdf',
(0, 1)),
TestParams(-1, 2, 100, uniform.cdf, 'uniform(0, 1) cdf', 'uniform_cdf',
(0, 1)),
TestParams(-10, 10, 1000, norm.pdf, 'gaussian(0, 1) pdf', 'gaussian_pdf',
(0, 1)),
TestParams(-10, 20, 1000, lambda x: norm.pdf(x, 10, 2**.5),
'gaussian(10, 2) pdf', 'gaussian_pdf', (10, 2)),
TestParams(-10, 10, 1000, norm.cdf, 'gaussian(0, 1) cdf', 'gaussian_cdf',
(0, 1)),
TestParams(-10, 20, 1000, lambda x: norm.cdf(x, 10, 2**.5),
'gaussian(10, 2) cdf', 'gaussian_cdf', (10, 2)),
TestParams(-1, 10, 1000, lambda x: gamma.pdf(x, 1), 'gamma(1, 1) pdf',
'gamma_pdf', (1, 1)),
TestParams(-1, 10, 1000, lambda x: gamma.pdf(x, 1.25, 0, 1),
'gamma(1.25, 1) pdf', 'gamma_pdf', (1.25, 1)),
TestParams(-1, 10, 1000, lambda x: gamma.pdf(x, 1.25, 0, 1 / 2),
'gamma(1.25, 2) pdf', 'gamma_pdf', (1.25, 2)),
TestParams(-1, 10, 1000, lambda x: gamma.cdf(x, 1), 'gamma(1, 1) cdf',
'gamma_cdf', (1, 1)),
TestParams(-1, 10, 1000, lambda x: gamma.cdf(x, 1.25, 0, 1),
'gamma(1.25, 1) cdf', 'gamma_cdf', (1.25, 1)),
import matplotlib.pyplot as plt
import numpy as np
from scipy.stats import norm
x = np.linspace(-4, 4, num=100)
fig = plt.figure(figsize=(8, 5))
ax = fig.add_subplot()
ax.plot(x, norm.pdf(x, loc=-1, scale=1), color="magenta")
ax.plot(x, norm.pdf(x, loc=0, scale=1), color=(0.85, 0.64, 0.12))
ax.plot(x, norm.pdf(x, loc=1, scale=1), color="#228B22")
plt.savefig('colours.svg', bbox_inches='tight')
plt.show()
