We Keep Coding

Speeding up normal distribution probability mass allocation

We have N users with P avg. points per user, where each point is a single value between 0 and 1. We need to distribute the mass of each point using a normal distribution with a known density of 0.05 as the points have some uncertainty. Additionally, we need to wrap the mass around 0 and 1 such that e.g. a point at 0.95 will also allocate mass around 0. I've provided a working example below, which bins the normal distribution into D=50 bins. The example uses the Python typing module, but you can ignore that if you'd like.
from typing import List, Any
import numpy as np
import scipy.stats
import matplotlib.pyplot as plt
D = 50
BINS: List[float] = np.linspace(0, 1, D + 1).tolist()
def probability_mass(distribution: Any, x0: float, x1: float) -> float:
"""
Computes the area under the distribution, wrapping at 1.
The wrapping is done by adding the PDF at +- 1.
"""
assert x1 > x0
return (
(distribution.cdf(x1) - distribution.cdf(x0))
+ (distribution.cdf(x1 + 1) - distribution.cdf(x0 + 1))
+ (distribution.cdf(x1 - 1) - distribution.cdf(x0 - 1))
)
def point_density(x: float) -> List[float]:
distribution: Any = scipy.stats.norm(loc=x, scale=0.05)
density: List[float] = []
for i in range(D):
density.append(probability_mass(distribution, BINS[i], BINS[i + 1]))
return density
def user_density(points: List[float]) -> Any:
# Find the density of each point
density: Any = np.array([point_density(p) for p in points])
# Combine points and normalize
combined = density.sum(axis=0)
return combined / combined.sum()
if __name__ == "__main__":
# Example for one user
data: List[float] = [.05, .3, .5, .5]
density = user_density(data)
# Example for multiple users (N = 2)
print([user_density(x) for x in [[.3, .5], [.7, .7, .7, .9]]])
### NB: THE REMAINING CODE IS FOR ILLUSTRATION ONLY!
### NB: THE IMPORTANT THING IS TO COMPUTE THE DENSITY FAST!
middle: List[float] = []
for i in range(D):
middle.append((BINS[i] + BINS[i + 1]) / 2)
plt.bar(x=middle, height=density, width=1.0 / D + 0.001)
plt.xlim(0, 1)
plt.xlabel("x")
plt.ylabel("Density")
plt.show()
In this example N=1, D=50, P=4. However, we want to scale this approach to N=10000 and P=100 while being as fast as possible. It's unclear to me how we'd vectorize this approach. How do we best speed up this?
EDIT
The faster solution can have slightly different results. For instance, it could approximate the normal distribution instead of using the precise normal distribution.
EDIT2
We only care about computing density using the user_density() function. The plot is only to help explain the approach. We do not care about the plot itself :)
EDIT3
Note that P is the avg. points per user. Some users may have more and some may have less. If it helps, you can assume that we can throw away points such that all users have a max of 2 * P points. It's fine to ignore this part while benchmarking as long as the solution can handle a flexible # of points per user.

You could get below 50ms for largest case (N=10000, AVG[P]=100, D=50) by using using FFT and creating data in numpy friendly format. Otherwise it will be closer to 300 msec.
The idea is to convolve a single normal distribution centered at 0 with a series Dirac deltas.
See image below:
Using circular convolution solves two issues.
naturally deals with wrapping at the edges
can be efficiently computed with FFT and Convolution Theorem
First one must create a distribution to be copied. Function mk_bell() created a histogram of a normal distribution of stddev 0.05 centered at 0.
The distribution wraps around 1. One could use arbitrary distribution here. The spectrum of the distribution is computed are used for fast convolution.
Next a comb-like function is created. The peaks are placed at indices corresponding to peaks in user density. E.g.
peaks_location = [0.1, 0.3, 0.7]
D = 10
maps to
peak_index = (D * peak_location).astype(int) = [1, 3, 7]
dist = [0, 1, 0, 1, 0, 0, 0, 1, 0, 0, 0] # ones at [1, 3, 7]
You can quickly create a composition of Diract Deltas by computing indices of the bins for each peak location with help of np.bincount() function.
In order to speed things even more one can compute comb-functions for user-peaks in parallel.
Array dist is 2D-array of shape NxD. It can be linearized to 1D array of shape (N*D). After this change element on position [user_id, peak_index] will be accessible from index user_id*D + peak_index.
With numpy-friendly input format (described below) this operation is easily vectorized.
The convolution theorem says that spectrum of convolution of two signals is equal to product of spectrums of each signal.
The spectrum is compute with numpy.fft.rfft which is a variant of Fast Fourier Transfrom dedicated to real-only signals (no imaginary part).
Numpy allows to compute FFT of each row of the larger matrix with one command.
Next, the spectrum of convolution is computed by simple multiplication and use of broadcasting.
Next, the spectrum is computed back to "time" domain by Inverse Fourier Transform implemented in numpy.fft.irfft.
To use the full speed of numpy one should avoid variable size data structure and keep to fixed size arrays. I propose to represent input data as three arrays.
uids the identifier for user, integer 0..N-1
peaks, the location of the peak
mass, the mass of the peek, currently it is 1/numer-of-peaks-for-user
This representation of data allows quick vectorized processing.
Eg:
user_data = [[0.1, 0.3], [0.5]]
maps to:
uids = [0, 0, 1] # 2 points for user_data[0], one from user_data[1]
peaks = [0.1, 0.3, 0.5] # serialized user_data
mass = [0.5, 0.5, 1] # scaling factors for each peak, 0.5 means 2 peaks for user 0
The code:
import numpy as np
import matplotlib.pyplot as plt
import time
def mk_bell(D, SIGMA):
# computes normal distribution wrapped and centered at zero
x = np.linspace(0, 1, D, endpoint=False);
x = (x + 0.5) % 1 - 0.5
bell = np.exp(-0.5*np.square(x / SIGMA))
return bell / bell.sum()
def user_densities_by_fft(uids, peaks, mass, D, N=None):
bell = mk_bell(D, 0.05).astype('f4')
sbell = np.fft.rfft(bell)
if N is None:
N = uids.max() + 1
# ensure that peaks are in [0..1) internal
peaks = peaks - np.floor(peaks)
# convert peak location from 0-1 to the indices
pidx = (D * (peaks + uids)).astype('i4')
dist = np.bincount(pidx, mass, N * D).reshape(N, D)
# process all users at once with Convolution Theorem
sdist = np.fft.rfft(dist)
sdist *= sbell
res = np.fft.irfft(sdist)
return res
def generate_data(N, Pmean):
# generateor for large data
data = []
for n in range(N):
# select P uniformly from 1..2*Pmean
P = np.random.randint(2 * Pmean) + 1
# select peak locations
chunk = np.random.uniform(size=P)
data.append(chunk.tolist())
return data
def make_data_numpy_friendly(data):
uids = []
chunks = []
mass = []
for uid, peaks in enumerate(data):
uids.append(np.full(len(peaks), uid))
mass.append(np.full(len(peaks), 1 / len(peaks)))
chunks.append(peaks)
return np.hstack(uids), np.hstack(chunks), np.hstack(mass)
D = 50
# demo for simple multi-distribution
data, N = [[0, .5], [.7, .7, .7, .9], [0.05, 0.3, 0.5, 0.5]], None
uids, peaks, mass = make_data_numpy_friendly(data)
dist = user_densities_by_fft(uids, peaks, mass, D, N)
plt.plot(dist.T)
plt.show()
# the actual measurement
N = 10000
P = 100
data = generate_data(N, P)
tic = time.time()
uids, peaks, mass = make_data_numpy_friendly(data)
toc = time.time()
print(f"make_data_numpy_friendly: {toc - tic}")
tic = time.time()
dist = user_densities_by_fft(uids, peaks, mass, D, N)
toc = time.time()
print(f"user_densities_by_fft: {toc - tic}")
The results on my 4-core Haswell machine are:
make_data_numpy_friendly: 0.2733159065246582
user_densities_by_fft: 0.04064297676086426
It took 40ms to process the data. Notice that processing data to numpy friendly format takes 6 times more time than the actual computation of distributions.
Python is really slow when it comes to looping.
Therefore I strongly recommend to generate input data directly in numpy-friendly way in the first place.
There are some issues to be fixed:
precision, can be improved by using larger D and downsampling
accuracy of peak location could be further improved by widening the spikes.
performance, scipy.fft offers move variants of FFT implementation that may be faster

This would be my vectorized approach:
data = np.array([0.05, 0.3, 0.5, 0.5])
np.random.seed(31415)
# random noise
randoms = np.random.normal(0,1,(len(data), int(1e5))) * 0.05
# samples with noise
samples = data[:,None] + randoms
# wrap [0,1]
samples = (samples % 1).ravel()
# histogram
hist, bins, patches = plt.hist(samples, bins=BINS, density=True)
Output:

I was able to reduce the time from about 4 seconds per sample of 100 datapoints to about 1 ms per sample.
It looks to me like you're spending quite a lot of time simulating a very large number of normal distributions. Since you're dealing with a very large sample size anyway, you may as well just use standard normal distribution values, because it'll all just average out anyway.
I recreated your approach (BaseMethod class), then created an optimized class (OptimizedMethod class), and evaluated them using a timeit decorator. The primary difference in my approach is the following line:
# Generate a standardized set of values to add to each sample to simulate normal distribution
self.norm_vals = np.array([norm.ppf(x / norm_val_n) * 0.05 for x in range(1, norm_val_n, 1)])
This creates a generic set of datapoints based on an inverse normal cumulative distribution function that we can add to each datapoint to simulate a normal distribution around that point. Then we just reshape the data into user samples and run np.histogram on the samples.
import numpy as np
import scipy.stats
from scipy.stats import norm
import time
# timeit decorator for evaluating performance
def timeit(method):
def timed(*args, **kw):
ts = time.time()
result = method(*args, **kw)
te = time.time()
print('%r %2.2f ms' % (method.__name__, (te - ts) * 1000 ))
return result
return timed
# Define Variables
N = 10000
D = 50
P = 100
# Generate sample data
np.random.seed(0)
data = np.random.rand(N, P)
# Run OP's method for comparison
class BaseMethod:
def __init__(self, d=50):
self.d = d
self.bins = np.linspace(0, 1, d + 1).tolist()
def probability_mass(self, distribution, x0, x1):
"""
Computes the area under the distribution, wrapping at 1.
The wrapping is done by adding the PDF at +- 1.
"""
assert x1 > x0
return (
(distribution.cdf(x1) - distribution.cdf(x0))
+ (distribution.cdf(x1 + 1) - distribution.cdf(x0 + 1))
+ (distribution.cdf(x1 - 1) - distribution.cdf(x0 - 1))
)
def point_density(self, x):
distribution = scipy.stats.norm(loc=x, scale=0.05)
density = []
for i in range(self.d):
density.append(self.probability_mass(distribution, self.bins[i], self.bins[i + 1]))
return density
#timeit
def base_user_density(self, data):
n = data.shape[0]
density = np.empty((n, self.d))
for i in range(data.shape[0]):
# Find the density of each point
row_density = np.array([self.point_density(p) for p in data[i]])
# Combine points and normalize
combined = row_density.sum(axis=0)
density[i, :] = combined / combined.sum()
return density
base = BaseMethod(d=D)
# Only running base method on first 2 rows of data because it's slow
density = base.base_user_density(data[:2])
print(density[:2, :5])
class OptimizedMethod:
def __init__(self, d=50, norm_val_n=50):
self.d = d
self.norm_val_n = norm_val_n
self.bins = np.linspace(0, 1, d + 1).tolist()
# Generate a standardized set of values to add to each sample to simulate normal distribution
self.norm_vals = np.array([norm.ppf(x / norm_val_n) * 0.05 for x in range(1, norm_val_n, 1)])
#timeit
def optimized_user_density(self, data):
samples = np.empty((data.shape[0], data.shape[1], self.norm_val_n - 1))
# transform datapoints to normal distributions around datapoint
for i in range(self.norm_vals.shape[0]):
samples[:, :, i] = data + self.norm_vals[i]
samples = samples.reshape(samples.shape[0], -1)
#wrap around [0, 1]
samples = samples % 1
#loop over samples for density
density = np.empty((data.shape[0], self.d))
for i in range(samples.shape[0]):
hist, bins = np.histogram(samples[i], bins=self.bins)
density[i, :] = hist / hist.sum()
return density
om = OptimizedMethod()
#Run optimized method on first 2 rows for apples to apples comparison
density = om.optimized_user_density(data[:2])
#Run optimized method on full data
density = om.optimized_user_density(data)
print(density[:2, :5])
Running on my system, the original method took about 8.4 seconds to run on 2 rows of data, while the optimized method took 1 millisecond to run on 2 rows of data and completed 10,000 rows in 4.7 seconds. I printed the first five values of the first 2 samples for each method.
'base_user_density' 8415.03 ms
[[0.02176227 0.02278653 0.02422535 0.02597123 0.02745976]
[0.0175103 0.01638513 0.01524853 0.01432158 0.01391156]]
'optimized_user_density' 1.09 ms
'optimized_user_density' 4755.49 ms
[[0.02142857 0.02244898 0.02530612 0.02612245 0.0277551 ]
[0.01673469 0.01653061 0.01510204 0.01428571 0.01326531]]

Find time shift of two signals using cross correlation

I have two signals which are related to each other and have been captured by two different measurement devices simultaneously.
Since the two measurements are not time synchronized there is a small time delay between them which I want to calculate. Additionally, I need to know which signal is the leading one.
The following can be assumed:
no or only very less noise present
speed of the algorithm is not an issue, only accuracy and robustness
signals are captured with an high sampling rate (>10 kHz) for several seconds
expected time delay is < 0.5s
I though of using-cross correlation for that purpose.
Any suggestions how to implement that in Python are very appreciated.
Please let me know if I should provide more information in order to find the most suitable algorithmn.

A popular approach: timeshift is the lag corresponding to the maximum cross-correlation coefficient. Here is how it works with an example:
import matplotlib.pyplot as plt
from scipy import signal
import numpy as np
def lag_finder(y1, y2, sr):
n = len(y1)
corr = signal.correlate(y2, y1, mode='same') / np.sqrt(signal.correlate(y1, y1, mode='same')[int(n/2)] * signal.correlate(y2, y2, mode='same')[int(n/2)])
delay_arr = np.linspace(-0.5*n/sr, 0.5*n/sr, n)
delay = delay_arr[np.argmax(corr)]
print('y2 is ' + str(delay) + ' behind y1')
plt.figure()
plt.plot(delay_arr, corr)
plt.title('Lag: ' + str(np.round(delay, 3)) + ' s')
plt.xlabel('Lag')
plt.ylabel('Correlation coeff')
plt.show()
# Sine sample with some noise and copy to y1 and y2 with a 1-second lag
sr = 1024
y = np.linspace(0, 2*np.pi, sr)
y = np.tile(np.sin(y), 5)
y += np.random.normal(0, 5, y.shape)
y1 = y[sr:4*sr]
y2 = y[:3*sr]
lag_finder(y1, y2, sr)
In the case of noisy signals, it is common to apply band-pass filters first. In the case of harmonic noise, they can be removed by identifying and removing frequency spikes present in the frequency spectrum.

Numpy has function correlate which suits your needs: https://docs.scipy.org/doc/numpy/reference/generated/numpy.correlate.html

To complement Reveille's answer above (I reproduce his algorithm), I would like to point out some ideas for preprocessing the input signals.
Since there seems to be no fit-for-all (duration in periods, resolution, offset, noise, signal type, ...) you may play with it.
In my example the application of a window function improves the detected phase shift (within resolution of the discretization).
import numpy as np
from scipy import signal
import matplotlib.pyplot as plt
r2d = 180.0/np.pi # conversion factor RAD-to-DEG
delta_phi_true = 50.0/r2d
def detect_phase_shift(t, x, y):
'''detect phase shift between two signals from cross correlation maximum'''
N = len(t)
L = t[-1] - t[0]
cc = signal.correlate(x, y, mode="same")
i_max = np.argmax(cc)
phi_shift = np.linspace(-0.5*L, 0.5*L , N)
delta_phi = phi_shift[i_max]
print("true delta phi = {} DEG".format(delta_phi_true*r2d))
print("detected delta phi = {} DEG".format(delta_phi*r2d))
print("error = {} DEG resolution for comparison dphi = {} DEG".format((delta_phi-delta_phi_true)*r2d, dphi*r2d))
print("ratio = {}".format(delta_phi/delta_phi_true))
return delta_phi
L = np.pi*10+2 # interval length [RAD], for generality not multiple period
N = 1001 # interval division, odd number is better (center is integer)
noise_intensity = 0.0
X = 0.5 # amplitude of first signal..
Y = 2.0 # ..and second signal
phi = np.linspace(0, L, N)
dphi = phi[1] - phi[0]
'''generate signals'''
nx = noise_intensity*np.random.randn(N)*np.sqrt(dphi)
ny = noise_intensity*np.random.randn(N)*np.sqrt(dphi)
x_raw = X*np.sin(phi) + nx
y_raw = Y*np.sin(phi+delta_phi_true) + ny
'''preprocessing signals'''
x = x_raw.copy()
y = y_raw.copy()
window = signal.windows.hann(N) # Hanning window
#x -= np.mean(x) # zero mean
#y -= np.mean(y) # zero mean
#x /= np.std(x) # scale
#y /= np.std(y) # scale
x *= window # reduce effect of finite length
y *= window # reduce effect of finite length
print(" -- using raw data -- ")
delta_phi_raw = detect_phase_shift(phi, x_raw, y_raw)
print(" -- using preprocessed data -- ")
delta_phi_preprocessed = detect_phase_shift(phi, x, y)
Without noise (to be deterministic) the output is
-- using raw data --
true delta phi = 50.0 DEG
detected delta phi = 47.864788975654 DEG
...
-- using preprocessed data --
true delta phi = 50.0 DEG
detected delta phi = 49.77938053468019 DEG
...

Numpy has a useful function, called correlation_lags for this, which uses the underlying correlate function mentioned by other answers to find the time lag. The example displayed at the bottom of that page is useful:
from scipy import signal
from numpy.random import default_rng
rng = default_rng()
x = rng.standard_normal(1000)
y = np.concatenate([rng.standard_normal(100), x])
correlation = signal.correlate(x, y, mode="full")
lags = signal.correlation_lags(x.size, y.size, mode="full")
lag = lags[np.argmax(correlation)]
Then lag would be -100

Is there Implementation of Hawkes Process in PyMC?

I want to use Hawkes process to model some data. I could not find whether PyMC supports Hawkes process. More specifically I want an observed variable with Hawkes Process and learn a posterior on its params.
If it is not there, then could I define it in PyMC in some way e.g. #deterministic etc.??

It's been quite a long time since your question, but I've worked it out on PyMC today so I'd thought I'd share the gist of my implementation for the other people who might get across the same problem. We're going to infer the parameters λ and α of a Hawkes process. I'm not going to cover the temporal scale parameter β, I'll leave that as an exercise for the readers.
First let's generate some data :
def hawkes_intensity(mu, alpha, points, t):
p = np.array(points)
p = p[p <= t]
p = np.exp(p - t)
return mu + alpha * np.sum(p)
def simulate_hawkes(mu, alpha, window):
t = 0
points = []
lambdas = []
while t < window:
m = hawkes_intensity(mu, alpha, points, t)
s = np.random.exponential(scale=1/m)
ratio = hawkes_intensity(mu, alpha, points, t + s)
t = t + s
if t < window:
points.append(t)
lambdas.append(ratio)
else:
break
points = np.sort(np.array(points, dtype=np.float32))
lambdas = np.array(lambdas, dtype=np.float32)
return points, lambdas
# parameters
window = 1000
mu = 8
alpha = 0.25
points, lambdas = simulate_hawkes(mu, alpha, window)
num_points = len(points)
We just generated some temporal points using some functions that I adapted from there : https://nbviewer.jupyter.org/github/MatthewDaws/PointProcesses/blob/master/Temporal%20points%20processes.ipynb
Now, the trick is to create a matrix of size (num_points, num_points) that contains the temporal distance of the ith point from all the other points. So the (i, j) point of the matrix is the temporal interval separating the ith point to the jth. This matrix will be used to compute the sum of the exponentials of the Hawkes process, ie. the self-exciting part. The way to create this matrix as well as the sum of the exponentials is a bit tricky. I'd recommend to check every line yourself so you can see what they do.
tile = np.tile(points, num_points).reshape(num_points, num_points)
tile = np.clip(points[:, None] - tile, 0, np.inf)
tile = np.tril(np.exp(-tile), k=-1)
Σ = np.sum(tile, axis=1)[:-1] # this is our self-exciting sum term
We have points and we have a matrix containg the sum of the excitations term.
The duration between two consecutive events of a Hawkes process follow an exponential distribution of parameter λ = λ0 + ∑ excitation. This is what we are going to model, but first we have to compute the duration between two consecutive points of our generated data.
interval = points[1:] - points[:-1]
We're now ready for inference:
with pm.Model() as model:
λ = pm.Exponential("λ", 1)
α = pm.Uniform("α", 0, 1)
lam = pm.Deterministic("lam", λ + α * Σ)
interarrival = pm.Exponential(
"interarrival", lam, observed=interval)
trace = pm.sample(2000, tune=4000)
pm.plot_posterior(trace, var_names=["λ", "α"])
plt.show()
print(np.mean(trace["λ"]))
print(np.mean(trace["α"]))
7.829
0.284
Note: the tile matrix can become quite large if you have many data points.

Pythonic way of detecting outliers in one dimensional observation data

For the given data, I want to set the outlier values (defined by 95% confidense level or 95% quantile function or anything that is required) as nan values. Following is the my data and code that I am using right now. I would be glad if someone could explain me further.
import numpy as np, matplotlib.pyplot as plt
data = np.random.rand(1000)+5.0
plt.plot(data)
plt.xlabel('observation number')
plt.ylabel('recorded value')
plt.show()

The problem with using percentile is that the points identified as outliers is a function of your sample size.
There are a huge number of ways to test for outliers, and you should give some thought to how you classify them. Ideally, you should use a-priori information (e.g. "anything above/below this value is unrealistic because...")
However, a common, not-too-unreasonable outlier test is to remove points based on their "median absolute deviation".
Here's an implementation for the N-dimensional case (from some code for a paper here: https://github.com/joferkington/oost_paper_code/blob/master/utilities.py):
def is_outlier(points, thresh=3.5):
"""
Returns a boolean array with True if points are outliers and False
otherwise.
Parameters:
-----------
points : An numobservations by numdimensions array of observations
thresh : The modified z-score to use as a threshold. Observations with
a modified z-score (based on the median absolute deviation) greater
than this value will be classified as outliers.
Returns:
--------
mask : A numobservations-length boolean array.
References:
----------
Boris Iglewicz and David Hoaglin (1993), "Volume 16: How to Detect and
Handle Outliers", The ASQC Basic References in Quality Control:
Statistical Techniques, Edward F. Mykytka, Ph.D., Editor.
"""
if len(points.shape) == 1:
points = points[:,None]
median = np.median(points, axis=0)
diff = np.sum((points - median)**2, axis=-1)
diff = np.sqrt(diff)
med_abs_deviation = np.median(diff)
modified_z_score = 0.6745 * diff / med_abs_deviation
return modified_z_score > thresh
This is very similar to one of my previous answers, but I wanted to illustrate the sample size effect in detail.
Let's compare a percentile-based outlier test (similar to #CTZhu's answer) with a median-absolute-deviation (MAD) test for a variety of different sample sizes:
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
def main():
for num in [10, 50, 100, 1000]:
# Generate some data
x = np.random.normal(0, 0.5, num-3)
# Add three outliers...
x = np.r_[x, -3, -10, 12]
plot(x)
plt.show()
def mad_based_outlier(points, thresh=3.5):
if len(points.shape) == 1:
points = points[:,None]
median = np.median(points, axis=0)
diff = np.sum((points - median)**2, axis=-1)
diff = np.sqrt(diff)
med_abs_deviation = np.median(diff)
modified_z_score = 0.6745 * diff / med_abs_deviation
return modified_z_score > thresh
def percentile_based_outlier(data, threshold=95):
diff = (100 - threshold) / 2.0
minval, maxval = np.percentile(data, [diff, 100 - diff])
return (data < minval) | (data > maxval)
def plot(x):
fig, axes = plt.subplots(nrows=2)
for ax, func in zip(axes, [percentile_based_outlier, mad_based_outlier]):
sns.distplot(x, ax=ax, rug=True, hist=False)
outliers = x[func(x)]
ax.plot(outliers, np.zeros_like(outliers), 'ro', clip_on=False)
kwargs = dict(y=0.95, x=0.05, ha='left', va='top')
axes[0].set_title('Percentile-based Outliers', **kwargs)
axes[1].set_title('MAD-based Outliers', **kwargs)
fig.suptitle('Comparing Outlier Tests with n={}'.format(len(x)), size=14)
main()
Notice that the MAD-based classifier works correctly regardless of sample-size, while the percentile based classifier classifies more points the larger the sample size is, regardless of whether or not they are actually outliers.

Detection of outliers in one dimensional data depends on its distribution
1- Normal Distribution :
Data values are almost equally distributed over the expected range :
In this case you easily use all the methods that include mean ,like the confidence interval of 3 or 2 standard deviations(95% or 99.7%) accordingly for a normally distributed data (central limit theorem and sampling distribution of sample mean).I is a highly effective method.
Explained in Khan Academy statistics and Probability - sampling distribution library.
One other way is prediction interval if you want confidence interval of data points rather than mean.
Data values are are randomly distributed over a range:
mean may not be a fair representation of the data, because the average is easily influenced by outliers (very small or large values in the data set that are not typical)
The median is another way to measure the center of a numerical data set.
Median Absolute deviation - a method which measures the distance of all points from the median in terms of median distance
http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h.htm - has a good explanation as explained in Joe Kington's answer above
2 - Symmetric Distribution : Again Median Absolute Deviation is a good method if the z-score calculation and threshold is changed accordingly
Explanation :
http://eurekastatistics.com/using-the-median-absolute-deviation-to-find-outliers/
3 - Asymmetric Distribution : Double MAD - Double Median Absolute Deviation
Explanation in the above attached link
Attaching my python code for reference :
def is_outlier_doubleMAD(self,points):
"""
FOR ASSYMMETRIC DISTRIBUTION
Returns : filtered array excluding the outliers
Parameters : the actual data Points array
Calculates median to divide data into 2 halves.(skew conditions handled)
Then those two halves are treated as separate data with calculation same as for symmetric distribution.(first answer)
Only difference being , the thresholds are now the median distance of the right and left median with the actual data median
"""
if len(points.shape) == 1:
points = points[:,None]
median = np.median(points, axis=0)
medianIndex = (points.size/2)
leftData = np.copy(points[0:medianIndex])
rightData = np.copy(points[medianIndex:points.size])
median1 = np.median(leftData, axis=0)
diff1 = np.sum((leftData - median1)**2, axis=-1)
diff1 = np.sqrt(diff1)
median2 = np.median(rightData, axis=0)
diff2 = np.sum((rightData - median2)**2, axis=-1)
diff2 = np.sqrt(diff2)
med_abs_deviation1 = max(np.median(diff1),0.000001)
med_abs_deviation2 = max(np.median(diff2),0.000001)
threshold1 = ((median-median1)/med_abs_deviation1)*3
threshold2 = ((median2-median)/med_abs_deviation2)*3
#if any threshold is 0 -> no outliers
if threshold1==0:
threshold1 = sys.maxint
if threshold2==0:
threshold2 = sys.maxint
#multiplied by a factor so that only the outermost points are removed
modified_z_score1 = 0.6745 * diff1 / med_abs_deviation1
modified_z_score2 = 0.6745 * diff2 / med_abs_deviation2
filtered1 = []
i = 0
for data in modified_z_score1:
if data < threshold1:
filtered1.append(leftData[i])
i += 1
i = 0
filtered2 = []
for data in modified_z_score2:
if data < threshold2:
filtered2.append(rightData[i])
i += 1
filtered = filtered1 + filtered2
return filtered

I've adapted the code from http://eurekastatistics.com/using-the-median-absolute-deviation-to-find-outliers and it gives the same results as Joe Kington's, but uses L1 distance instead of L2 distance, and has support for asymmetric distributions. The original R code did not have Joe's 0.6745 multiplier, so I also added that in for consistency within this thread. Not 100% sure if it's necessary, but makes the comparison apples-to-apples.
def doubleMADsfromMedian(y,thresh=3.5):
# warning: this function does not check for NAs
# nor does it address issues when
# more than 50% of your data have identical values
m = np.median(y)
abs_dev = np.abs(y - m)
left_mad = np.median(abs_dev[y <= m])
right_mad = np.median(abs_dev[y >= m])
y_mad = left_mad * np.ones(len(y))
y_mad[y > m] = right_mad
modified_z_score = 0.6745 * abs_dev / y_mad
modified_z_score[y == m] = 0
return modified_z_score > thresh

Well a simple solution can also be, removing something which outside 2 standard deviations(or 1.96):
import random
def outliers(tmp):
"""tmp is a list of numbers"""
outs = []
mean = sum(tmp)/(1.0*len(tmp))
var = sum((tmp[i] - mean)**2 for i in range(0, len(tmp)))/(1.0*len(tmp))
std = var**0.5
outs = [tmp[i] for i in range(0, len(tmp)) if abs(tmp[i]-mean) > 1.96*std]
return outs
lst = [random.randrange(-10, 55) for _ in range(40)]
print lst
print outliers(lst)

Use np.percentile as #Martin suggested:
percentiles = np.percentile(data, [2.5, 97.5])
# or =>, <= for within 95%
data[(percentiles[0]<data) & (percentiles[1]>data)]
# set the outliners to np.nan
data[(percentiles[0]>data) | (percentiles[1]<data)] = np.nan

Fundamental Frequency by Cepstral Method

I'm trying to find frequencies by the Cepstral method. For my tests I got the following file http://www.mediacollege.com/audio/tone/files/440Hz_44100Hz_16bit_05sec.wav, an audio signal with a frequency of 440Hz.
I've applied the following formula:
cepstrum = IFFT (log FFT (s))
I'm getting 256 chunk, But my results are always wrong ...
from numpy.fft import fft, ifft
import math
import wave
import numpy as np
from scipy.signal import hamming
index1=15000;
frameSize=256;
spf = wave.open('440.wav','r');
fs = spf.getframerate();
signal = spf.readframes(-1);
signal = np.fromstring(signal, 'Int16');
index2=index1+frameSize-1;
frames=signal[index1:int(index2)+1]
zeroPaddedFrameSize=16*frameSize;
frames2=frames*hamming(len(frames));
frameSize=len(frames);
if (zeroPaddedFrameSize>frameSize):
zrs= np.zeros(zeroPaddedFrameSize-frameSize);
frames2=np.concatenate((frames2, zrs), axis=0)
fftResult=np.log(abs(fft(frames2)));
ceps=ifft(fftResult);
posmax = ceps.argmax();
result = fs/zeroPaddedFrameSize*(posmax-1)
print result
For this case how get the result = 440?
**
UPDATE:
**
Well i rewrote my source in matlab, and now everything seems to work, i did tests with frequencies of 440 Hz and 250 Hz ...
For 440Hz i get 441Hz not bad
For 250Hz i get 249.1525Hz near result
I did make one simple way to get the peaks into cepstral values.
I think I can find better results using quadract interpolation to find the maximum !
I'm plotting my results for the estimation of 440Hz
Sharing the source for Cepstral Frequency estimation:
%% ederwander Cepstral Frequency (Matlab)
waveFile='440.wav';
[y, fs, nbits]=wavread(waveFile);
subplot(4,2,1); plot(y); legend('Original signal');
startIndex=15000;
frameSize=4096;
endIndex=startIndex+frameSize-1;
frame = y(startIndex:endIndex);
subplot(4,2,2); plot(frame); legend('4096 CHUNK signal');
%make hamming window
win = hamming(length(frame));
%samples multplied by hamming window
windowedSignal = frame.*win;
fftResult=log(abs(fft(windowedSignal)));
subplot(4,2,3); plot(fftResult); legend('FFT signal');
ceps=ifft(fftResult);
subplot(4,2,4); plot(ceps); legend('ceps signal');
nceps=length(ceps)
%find the peaks in ceps
peaks = zeros(nceps,1);
k=3;
while(k <= nceps - 1)
y1 = ceps(k - 1);
y2 = ceps(k);
y3 = ceps(k + 1);
if (y2 > y1 && y2 >= y3)
peaks(k)=ceps(k);
end
k=k+1;
end
subplot(4,2,5); plot(peaks); legend('PEAKS');
%get the maximum ...
[maxivalue, maxi]=max(peaks)
result = fs/(maxi+1)
subplot(4,2,6); plot(result); %legend('Frequency is' result);
legend(sprintf('Final Result Frequency =====>>> (%8.3f)',result))

256 is probably too small to do anything useful if your sample rate is 44.1 kHz. The resolution of your FFT in this case will be 44100 / 256 = 172 Hz. If you want a resolution of the order of say 10 Hz then you might use an FFT size of 4096.

Cepstral methods work best with signals that have a high harmonic content, not as well on signals that are close to pure sinusoids.
The best test signal might be something more like repetitive, very near equally spaced, impulses in the time domain (the more per FFT window the better), which should produce something close to repetitive equally spaced peaks in the frequency domain, which should show up as the exciter portion of a cepstrum. The impulse response will be represented in the lower formant portion of the cepstrum.

I had a similar problem and so I have reused a part of your code and improvement the quality of the result by performing contiguous evaluation of the same frame and then selecting the medium value from the
I am getting a consistent result.
def fondamentals(frames0, samplerate):
mid = 16
sample = mid*2+1
res = []
for first in xrange(sample):
last = first-sample
frames = frames0[first:last]
res.append(_fondamentals(frames, samplerate))
res = sorted(res)
return res[mid] # We use the medium value
def _fondamentals(frames, samplerate):
frames2=frames*hamming(len(frames));
frameSize=len(frames);
ceps=ifft(np.log(np.abs(fft(frames2))))
nceps=ceps.shape[-1]*2/3
peaks = []
k=3
while(k < nceps - 1):
y1 = (ceps[k - 1])
y2 = (ceps[k])
y3 = (ceps[k + 1])
if (y2 > y1 and y2 >= y3): peaks.append([float(samplerate)/(k+2),abs(y2), k, nceps])
k=k+1
maxi=max(peaks, key=lambda x: x[1])
return maxi[0]