Related
I am trying to learn how to sample truncated distributions. To begin with I decided to try a simple example I found here example
I didn't really understand the division by the CDF, therefore I decided to tweak the algorithm a bit. Being sampled is an exponential distribution for values x>0 Here is an example python code:
# Sample exponential distribution for the case x>0
import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import norm
def pdf(x):
return x*np.exp(-x)
xvec=np.zeros(1000000)
x=1.
for i in range(1000000):
a=x+np.random.normal()
xs=x
if a > 0. :
xs=a
A=pdf(xs)/pdf(x)
if np.random.uniform()<A :
x=xs
xvec[i]=x
x=np.linspace(0,15,1000)
plt.plot(x,pdf(x))
plt.hist([x for x in xvec if x != 0],bins=150,normed=True)
plt.show()
Ant the output is:
The code above seems to work fine only for when using the condition if a > 0. :, i.e. positive x, choosing another condition (e.g. if a > 0.5 :) produces wrong results.
Since my final goal was to sample a 2D-Gaussian - pdf on a truncated interval I tried extending the simple example using the exponential distribution (see the code below). Unfortunately, since the simple case didn't work, I assume that the code given below would yield wrong results.
I assume that all this can be done using the advanced tools of python. However, since my primary idea was to understand the principle behind, I would greatly appreciate your help to understand my mistake.
Thank you for your help.
EDIT:
# code updated according to the answer of CrazyIvan
from scipy.stats import multivariate_normal
RANGE=100000
a=2.06072E-02
b=1.10011E+00
a_range=[0.001,0.5]
b_range=[0.01, 2.5]
cov=[[3.1313994E-05, 1.8013737E-03],[ 1.8013737E-03, 1.0421529E-01]]
x=a
y=b
j=0
for i in range(RANGE):
a_t,b_t=np.random.multivariate_normal([a,b],cov)
# accept if within bounds - all that is neded to truncate
if a_range[0]<a_t and a_t<a_range[1] and b_range[0]<b_t and b_t<b_range[1]:
print(dx,dy)
EDIT:
I changed the code by norming the analytic pdf according to this scheme, and according to the answers given by, #Crazy Ivan and #Leandro Caniglia , for the case where the bottom of the pdf is removed. That is dividing by (1-CDF(0.5)) since my accept condition is x>0.5. This seems again to show some discrepancies. Again the mystery prevails ..
import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import norm
def pdf(x):
return x*np.exp(-x)
# included the corresponding cdf
def cdf(x):
return 1. -np.exp(-x)-x*np.exp(-x)
xvec=np.zeros(1000000)
x=1.
for i in range(1000000):
a=x+np.random.normal()
xs=x
if a > 0.5 :
xs=a
A=pdf(xs)/pdf(x)
if np.random.uniform()<A :
x=xs
xvec[i]=x
x=np.linspace(0,15,1000)
# new part norm the analytic pdf to fix the area
plt.plot(x,pdf(x)/(1.-cdf(0.5)))
plt.hist([x for x in xvec if x != 0],bins=200,normed=True)
plt.savefig("test_exp.png")
plt.show()
It seems that this can be cured by choosing larger shift size
shift=15.
a=x+np.random.normal()*shift.
which is in general an issue of the Metropolis - Hastings. See the graph below:
I also checked shift=150
Bottom line is that changing the shift size definitely improves the convergence. The misery is why, since the Gaussian is unbounded.
You say you want to learn the basic idea of sampling a truncated distribution, but your source is a blog post about
Metropolis–Hastings algorithm? Do you actually need this "method for obtaining a sequence of random samples from a probability distribution for which direct sampling is difficult"? Taking this as your starting point is like learning English by reading Shakespeare.
Truncated normal
For truncated normal, basic rejection sampling is all you need: generate samples for original distribution, reject those outside of bounds. As Leandro Caniglia noted, you should not expect truncated distribution to have the same PDF except on a shorter interval — this is plain impossible because the area under the graph of a PDF is always 1. If you cut off stuff from sides, there has to be more in the middle; the PDF gets rescaled.
It's quite inefficient to gather samples one by one, when you need 100000. I would grab 100000 normal samples at once, accept only those that fit; then repeat until I have enough. Example of sampling truncated normal between amin and amax:
import numpy as np
n_samples = 100000
amin, amax = -1, 2
samples = np.zeros((0,)) # empty for now
while samples.shape[0] < n_samples:
s = np.random.normal(0, 1, size=(n_samples,))
accepted = s[(s >= amin) & (s <= amax)]
samples = np.concatenate((samples, accepted), axis=0)
samples = samples[:n_samples] # we probably got more than needed, so discard extra ones
And here is the comparison with the PDF curve, rescaled by division by cdf(amax) - cdf(amin) as explained above.
from scipy.stats import norm
_ = plt.hist(samples, bins=50, density=True)
t = np.linspace(-2, 3, 500)
plt.plot(t, norm.pdf(t)/(norm.cdf(amax) - norm.cdf(amin)), 'r')
plt.show()
Truncated multivariate normal
Now we want to keep the first coordinate between amin and amax, and the second between bmin and bmax. Same story, except there will be a 2-column array and the comparison with bounds is done in a relatively sneaky way:
(np.min(s - [amin, bmin], axis=1) >= 0) & (np.max(s - [amax, bmax], axis=1) <= 0)
This means: subtract amin, bmin from each row and keep only the rows where both results are nonnegative (meaning we had a >= amin and b >= bmin). Also do a similar thing with amax, bmax. Accept only the rows that meet both criteria.
n_samples = 10
amin, amax = -1, 2
bmin, bmax = 0.2, 2.4
mean = [0.3, 0.5]
cov = [[2, 1.1], [1.1, 2]]
samples = np.zeros((0, 2)) # 2 columns now
while samples.shape[0] < n_samples:
s = np.random.multivariate_normal(mean, cov, size=(n_samples,))
accepted = s[(np.min(s - [amin, bmin], axis=1) >= 0) & (np.max(s - [amax, bmax], axis=1) <= 0)]
samples = np.concatenate((samples, accepted), axis=0)
samples = samples[:n_samples, :]
Not going to plot, but here are some values: naturally, within bounds.
array([[ 0.43150033, 1.55775629],
[ 0.62339265, 1.63506963],
[-0.6723598 , 1.58053835],
[-0.53347361, 0.53513105],
[ 1.70524439, 2.08226558],
[ 0.37474842, 0.2512812 ],
[-0.40986396, 0.58783193],
[ 0.65967087, 0.59755193],
[ 0.33383214, 2.37651975],
[ 1.7513789 , 1.24469918]])
To compute the truncated density function pdf_t from the entire density function pdf, do the following:
Let [a, b] be the truncation interval; (x axis)
Let A := cdf(a) and B := cdf(b); (cdf = non-truncated cumulative distribution function)
Then pdf_t(x) := pdf(x) / (B - A) if x in [a, b] and 0 elsewhere.
In cases where a = -infinity (resp. b = +infinity), take A := 0 (resp. B := 1).
As for the "mystery" you see
please note that your blue curve is wrong. It is not the pdf of your truncated distribution, it is just the pdf of the non-truncated one, scaled by the correct amount (division by 1-cdf(0.5)). The actual truncated pdf curve starts with a vertical line on x = 0.5 which goes up until it reaches your current blue curve. In other words, you only scaled the curve but forgot to truncate it, in this case to the left. Such a truncation corresponds to the "0 elsewhere" part of step 3 in the algorithm above.
This is my first question on Stackoverflow, so I apologize if I word it poorly. I am writing code to take raw acceleration data from an IMU and then integrate it to update the position of an object. Currently this code takes a new accelerometer reading every milisecond, and uses that to update the position. My system has a lot of noise, which results in crazy readings due to compounding error, even with the ZUPT scheme I implemented. I know that a Kalman filter is theoretically ideal for this scenario, and I would like to use the pykalman module instead of building one myself.
My first question is, can pykalman be used in real time like this? From the documentation it looks to me like you have to have a record of all measurements and then perform the smooth operation, which would not be practical as I want to filter recursively every milisecond.
My second question is, for the transition matrix can I only apply pykalman to the acceleration data by itself, or can I somehow include the double integration to position? What would that matrix look like?
If pykalman is not practical for this situation, is there another way I can implement a Kalman Filter? Thank you in advance!
You can use a Kalman Filter in this case, but your position estimation will strongly depend on the precision of your acceleration signal. The Kalman Filter is actually useful for a fusion of several signals. So error of one signal can be compensated by another signal. Ideally you need to use sensors based on different physical effects (for example an IMU for acceleration, GPS for position, odometry for velocity).
In this answer I'm going to use readings from two acceleration sensors (both in X direction). One of these sensors is an expansive and precise. The second one is much cheaper. So you will see the sensor precision influence on the position and velocity estimations.
You already mentioned the ZUPT scheme. I just want to add some notes: it is very important to have a good estimation of the pitch angle, to get rid of the gravitation component in your X-acceleration. If you use Y- and Z-acceleration you need both pitch and roll angles.
Let's start with modelling. Assume you have only acceleration readings in X-direction. So your observation will look like
Now you need to define the smallest data set, which completely describes your system in each point of time. It will be the system state.
The mapping between the measurement and state domains is defined by the observation matrix:
Now you need to describe the system dynamics. According to this information the Filter will predict a new state based on the previous one.
In my case dt=0.01s. Using this matrix the Filter will integrate the acceleration signal to estimate the velocity and position.
The observation covariance R can be described by the variance of your sensor readings. In my case I have only one signal in my observation, so the observation covariance is equal to the variance of the X-acceleration (the value can be calculated based on your sensors datasheet).
Through the transition covariance Q you describe the system noise. The smaller the matrix values, the smaller the system noise. The Filter will become stiffer and the estimation will be delayed. The weight of the system's past will be higher compared to new measurement. Otherwise the filter will be more flexible and will react strongly on each new measurement.
Now everything is ready to configure the Pykalman. In order to use it in real time, you have to use the filter_update function.
from pykalman import KalmanFilter
import numpy as np
import matplotlib.pyplot as plt
load_data()
# Data description
# Time
# AccX_HP - high precision acceleration signal
# AccX_LP - low precision acceleration signal
# RefPosX - real position (ground truth)
# RefVelX - real velocity (ground truth)
# switch between two acceleration signals
use_HP_signal = 1
if use_HP_signal:
AccX_Value = AccX_HP
AccX_Variance = 0.0007
else:
AccX_Value = AccX_LP
AccX_Variance = 0.0020
# time step
dt = 0.01
# transition_matrix
F = [[1, dt, 0.5*dt**2],
[0, 1, dt],
[0, 0, 1]]
# observation_matrix
H = [0, 0, 1]
# transition_covariance
Q = [[0.2, 0, 0],
[ 0, 0.1, 0],
[ 0, 0, 10e-4]]
# observation_covariance
R = AccX_Variance
# initial_state_mean
X0 = [0,
0,
AccX_Value[0, 0]]
# initial_state_covariance
P0 = [[ 0, 0, 0],
[ 0, 0, 0],
[ 0, 0, AccX_Variance]]
n_timesteps = AccX_Value.shape[0]
n_dim_state = 3
filtered_state_means = np.zeros((n_timesteps, n_dim_state))
filtered_state_covariances = np.zeros((n_timesteps, n_dim_state, n_dim_state))
kf = KalmanFilter(transition_matrices = F,
observation_matrices = H,
transition_covariance = Q,
observation_covariance = R,
initial_state_mean = X0,
initial_state_covariance = P0)
# iterative estimation for each new measurement
for t in range(n_timesteps):
if t == 0:
filtered_state_means[t] = X0
filtered_state_covariances[t] = P0
else:
filtered_state_means[t], filtered_state_covariances[t] = (
kf.filter_update(
filtered_state_means[t-1],
filtered_state_covariances[t-1],
AccX_Value[t, 0]
)
)
f, axarr = plt.subplots(3, sharex=True)
axarr[0].plot(Time, AccX_Value, label="Input AccX")
axarr[0].plot(Time, filtered_state_means[:, 2], "r-", label="Estimated AccX")
axarr[0].set_title('Acceleration X')
axarr[0].grid()
axarr[0].legend()
axarr[0].set_ylim([-4, 4])
axarr[1].plot(Time, RefVelX, label="Reference VelX")
axarr[1].plot(Time, filtered_state_means[:, 1], "r-", label="Estimated VelX")
axarr[1].set_title('Velocity X')
axarr[1].grid()
axarr[1].legend()
axarr[1].set_ylim([-1, 20])
axarr[2].plot(Time, RefPosX, label="Reference PosX")
axarr[2].plot(Time, filtered_state_means[:, 0], "r-", label="Estimated PosX")
axarr[2].set_title('Position X')
axarr[2].grid()
axarr[2].legend()
axarr[2].set_ylim([-10, 1000])
plt.show()
When using the better IMU-sensor, the estimated position is exactly the same as the ground truth:
The cheaper sensor gives significantly worse results:
I hope I could help you. If you have some questions, I will try to answer them.
UPDATE
If you want to experiment with different data you can generate them easily (unfortunately I don't have the original data any more).
Here is a simple matlab script to generate reference, good and poor sensor set.
clear;
dt = 0.01;
t=0:dt:70;
accX_var_best = 0.0005; % (m/s^2)^2
accX_var_good = 0.0007; % (m/s^2)^2
accX_var_worst = 0.001; % (m/s^2)^2
accX_ref_noise = randn(size(t))*sqrt(accX_var_best);
accX_good_noise = randn(size(t))*sqrt(accX_var_good);
accX_worst_noise = randn(size(t))*sqrt(accX_var_worst);
accX_basesignal = sin(0.3*t) + 0.5*sin(0.04*t);
accX_ref = accX_basesignal + accX_ref_noise;
velX_ref = cumsum(accX_ref)*dt;
distX_ref = cumsum(velX_ref)*dt;
accX_good_offset = 0.001 + 0.0004*sin(0.05*t);
accX_good = accX_basesignal + accX_good_noise + accX_good_offset;
velX_good = cumsum(accX_good)*dt;
distX_good = cumsum(velX_good)*dt;
accX_worst_offset = -0.08 + 0.004*sin(0.07*t);
accX_worst = accX_basesignal + accX_worst_noise + accX_worst_offset;
velX_worst = cumsum(accX_worst)*dt;
distX_worst = cumsum(velX_worst)*dt;
subplot(3,1,1);
plot(t, accX_ref);
hold on;
plot(t, accX_good);
plot(t, accX_worst);
hold off;
grid minor;
legend('ref', 'good', 'worst');
title('AccX');
subplot(3,1,2);
plot(t, velX_ref);
hold on;
plot(t, velX_good);
plot(t, velX_worst);
hold off;
grid minor;
legend('ref', 'good', 'worst');
title('VelX');
subplot(3,1,3);
plot(t, distX_ref);
hold on;
plot(t, distX_good);
plot(t, distX_worst);
hold off;
grid minor;
legend('ref', 'good', 'worst');
title('DistX');
The simulated data looks pretty the same like the data above.
I have a set of 46 years worth of rainfall data. It's in the form of 46 numpy arrays each with a shape of 145, 192, so each year is a different array of maximum rainfall data at each lat and lon coordinate in the given model.
I need to create a global map of tau values by doing an M-K test (Mann-Kendall) for each coordinate over the 46 years.
I'm still learning python, so I've been having trouble finding a way to go through all the data in a simple way that doesn't involve me making 27840 new arrays for each coordinate.
So far I've looked into how to use scipy.stats.kendalltau and using the definition from here: https://github.com/mps9506/Mann-Kendall-Trend
EDIT:
To clarify and add a little more detail, I need to perform a test on for each coordinate and not just each file individually. For example, for the first M-K test, I would want my x=46 and I would want y=data1[0,0],data2[0,0],data3[0,0]...data46[0,0]. Then to repeat this process for every single coordinate in each array. In total the M-K test would be done 27840 times and leave me with 27840 tau values that I can then plot on a global map.
EDIT 2:
I'm now running into a different problem. Going off of the suggested code, I have the following:
for i in range(145):
for j in range(192):
out[i,j] = mk_test(yrmax[:,i,j],alpha=0.05)
print out
I used numpy.stack to stack all 46 arrays into a single array (yrmax) with shape: (46L, 145L, 192L) I've tested it out and it calculates p and tau correctly if I change the code from out[i,j] to just out. However, doing this messes up the for loop so it only takes the results from the last coordinate in stead of all of them. And if I leave the code as it is above, I get the error: TypeError: list indices must be integers, not tuple
My first guess was that it has to do with mk_test and how the information is supposed to be returned in the definition. So I've tried altering the code from the link above to change how the data is returned, but I keep getting errors relating back to tuples. So now I'm not sure where it's going wrong and how to fix it.
EDIT 3:
One more clarification I thought I should add. I've already modified the definition in the link so it returns only the two number values I want for creating maps, p and z.
I don't think this is as big an ask as you may imagine. From your description it sounds like you don't actually want the scipy kendalltau, but the function in the repository you posted. Here is a little example I set up:
from time import time
import numpy as np
from mk_test import mk_test
data = np.array([np.random.rand(145, 192) for _ in range(46)])
mk_res = np.empty((145, 192), dtype=object)
start = time()
for i in range(145):
for j in range(192):
out[i, j] = mk_test(data[:, i, j], alpha=0.05)
print(f'Elapsed Time: {time() - start} s')
Elapsed Time: 35.21990394592285 s
My system is a MacBook Pro 2.7 GHz Intel Core I7 with 16 GB Ram so nothing special.
Each entry in the mk_res array (shape 145, 192) corresponds to one of your coordinate points and contains an entry like so:
array(['no trend', 'False', '0.894546014835', '0.132554125342'], dtype='<U14')
One thing that might be useful would be to modify the code in mk_test.py to return all numerical values. So instead of 'no trend'/'positive'/'negative' you could return 0/1/-1, and 1/0 for True/False and then you wouldn't have to worry about the whole object array type. I don't know what kind of analysis you might want to do downstream but I imagine that would preemptively circumvent any headaches.
Thanks to the answers provided and some work I was able to work out a solution that I'll provide here for anyone else that needs to use the Mann-Kendall test for data analysis.
The first thing I needed to do was flatten the original array I had into a 1D array. I know there is probably an easier way to go about doing this, but I ultimately used the following code based on code Grr suggested using.
`x = 46
out1 = np.empty(x)
out = np.empty((0))
for i in range(146):
for j in range(193):
out1 = yrmax[:,i,j]
out = np.append(out, out1, axis=0) `
Then I reshaped the resulting array (out) as follows:
out2 = np.reshape(out,(27840,46))
I did this so my data would be in a format compatible with scipy.stats.kendalltau 27840 is the total number of values I have at every coordinate that will be on my map (i.e. it's just 145*192) and the 46 is the number of years the data spans.
I then used the following loop I modified from Grr's code to find Kendall-tau and it's respective p-value at each latitude and longitude over the 46 year period.
`x = range(46)
y = np.zeros((0))
for j in range(27840):
b = sc.stats.kendalltau(x,out2[j,:])
y = np.append(y, b, axis=0)`
Finally, I reshaped the data one for time as shown:newdata = np.reshape(y,(145,192,2)) so the final array is in a suitable format to be used to create a global map of both tau and p-values.
Thanks everyone for the assistance!
Depending on your situation, it might just be easiest to make the arrays.
You won't really need them all in memory at once (not that it sounds like a terrible amount of data). Something like this only has to deal with one "copied out" coordinate trend at once:
SIZE = (145,192)
year_matrices = load_years() # list of one 145x192 arrays per year
result_matrix = numpy.zeros(SIZE)
for x in range(SIZE[0]):
for y in range(SIZE[1]):
coord_trend = map(lambda d: d[x][y], year_matrices)
result_matrix[x][y] = analyze_trend(coord_trend)
print result_matrix
Now, there are things like itertools.izip that could help you if you really want to avoid actually copying the data.
Here's a concrete example of how Python's "zip" might works with data like yours (although as if you'd used ndarray.flatten on each year):
year_arrays = [
['y0_coord0_val', 'y0_coord1_val', 'y0_coord2_val', 'y0_coord2_val'],
['y1_coord0_val', 'y1_coord1_val', 'y1_coord2_val', 'y1_coord2_val'],
['y2_coord0_val', 'y2_coord1_val', 'y2_coord2_val', 'y2_coord2_val'],
]
assert len(year_arrays) == 3
assert len(year_arrays[0]) == 4
coord_arrays = zip(*year_arrays) # i.e. `zip(year_arrays[0], year_arrays[1], year_arrays[2])`
# original data is essentially transposed
assert len(coord_arrays) == 4
assert len(coord_arrays[0]) == 3
assert coord_arrays[0] == ('y0_coord0_val', 'y1_coord0_val', 'y2_coord0_val', 'y3_coord0_val')
assert coord_arrays[1] == ('y0_coord1_val', 'y1_coord1_val', 'y2_coord1_val', 'y3_coord1_val')
assert coord_arrays[2] == ('y0_coord2_val', 'y1_coord2_val', 'y2_coord2_val', 'y3_coord2_val')
assert coord_arrays[3] == ('y0_coord2_val', 'y1_coord2_val', 'y2_coord2_val', 'y3_coord2_val')
flat_result = map(analyze_trend, coord_arrays)
The example above still copies the data (and all at once, rather than a coordinate at a time!) but hopefully shows what's going on.
Now, if you replace zip with itertools.izip and map with itertools.map then the copies needn't occur — itertools wraps the original arrays and keeps track of where it should be fetching values from internally.
There's a catch, though: to take advantage itertools you to access the data only sequentially (i.e. through iteration). In your case, it looks like the code at https://github.com/mps9506/Mann-Kendall-Trend/blob/master/mk_test.py might not be compatible with that. (I haven't reviewed the algorithm itself to see if it could be.)
Also please note that in the example I've glossed over the numpy ndarray stuff and just show flat coordinate arrays. It looks like numpy has some of it's own options for handling this instead of itertools, e.g. this answer says "Taking the transpose of an array does not make a copy". Your question was somewhat general, so I've tried to give some general tips as to ways one might deal with larger data in Python.
I ran into the same task and have managed to come up with a vectorized solution using numpy and scipy.
The formula are the same as in this page: https://vsp.pnnl.gov/help/Vsample/Design_Trend_Mann_Kendall.htm.
The trickiest part is to work out the adjustment for the tied values. I modified the code as in this answer to compute the number of tied values for each record, in a vectorized manner.
Below are the 2 functions:
import copy
import numpy as np
from scipy.stats import norm
def countTies(x):
'''Count number of ties in rows of a 2D matrix
Args:
x (ndarray): 2d matrix.
Returns:
result (ndarray): 2d matrix with same shape as <x>. In each
row, the number of ties are inserted at (not really) arbitary
locations.
The locations of tie numbers in are not important, since
they will be subsequently put into a formula of sum(t*(t-1)*(2t+5)).
Inspired by: https://stackoverflow.com/a/24892274/2005415.
'''
if np.ndim(x) != 2:
raise Exception("<x> should be 2D.")
m, n = x.shape
pad0 = np.zeros([m, 1]).astype('int')
x = copy.deepcopy(x)
x.sort(axis=1)
diff = np.diff(x, axis=1)
cated = np.concatenate([pad0, np.where(diff==0, 1, 0), pad0], axis=1)
absdiff = np.abs(np.diff(cated, axis=1))
rows, cols = np.where(absdiff==1)
rows = rows.reshape(-1, 2)[:, 0]
cols = cols.reshape(-1, 2)
counts = np.diff(cols, axis=1)+1
result = np.zeros(x.shape).astype('int')
result[rows, cols[:,1]] = counts.flatten()
return result
def MannKendallTrend2D(data, tails=2, axis=0, verbose=True):
'''Vectorized Mann-Kendall tests on 2D matrix rows/columns
Args:
data (ndarray): 2d array with shape (m, n).
Keyword Args:
tails (int): 1 for 1-tail, 2 for 2-tail test.
axis (int): 0: test trend in each column. 1: test trend in each
row.
Returns:
z (ndarray): If <axis> = 0, 1d array with length <n>, standard scores
corresponding to data in each row in <x>.
If <axis> = 1, 1d array with length <m>, standard scores
corresponding to data in each column in <x>.
p (ndarray): p-values corresponding to <z>.
'''
if np.ndim(data) != 2:
raise Exception("<data> should be 2D.")
# alway put records in rows and do M-K test on each row
if axis == 0:
data = data.T
m, n = data.shape
mask = np.triu(np.ones([n, n])).astype('int')
mask = np.repeat(mask[None,...], m, axis=0)
s = np.sign(data[:,None,:]-data[:,:,None]).astype('int')
s = (s * mask).sum(axis=(1,2))
#--------------------Count ties--------------------
counts = countTies(data)
tt = counts * (counts - 1) * (2*counts + 5)
tt = tt.sum(axis=1)
#-----------------Sample Gaussian-----------------
var = (n * (n-1) * (2*n+5) - tt) / 18.
eps = 1e-8 # avoid dividing 0
z = (s - np.sign(s)) / (np.sqrt(var) + eps)
p = norm.cdf(z)
p = np.where(p>0.5, 1-p, p)
if tails==2:
p=p*2
return z, p
I assume your data come in the layout of (time, latitude, longitude), and you are examining the temporal trend for each lat/lon cell.
To simulate this task, I synthesized a sample data array of shape (50, 145, 192). The 50 time points are taken from Example 5.9 of the book Wilks 2011, Statistical methods in the atmospheric sciences. And then I simply duplicated the same time series 27840 times to make it (50, 145, 192).
Below is the computation:
x = np.array([0.44,1.18,2.69,2.08,3.66,1.72,2.82,0.72,1.46,1.30,1.35,0.54,\
2.74,1.13,2.50,1.72,2.27,2.82,1.98,2.44,2.53,2.00,1.12,2.13,1.36,\
4.9,2.94,1.75,1.69,1.88,1.31,1.76,2.17,2.38,1.16,1.39,1.36,\
1.03,1.11,1.35,1.44,1.84,1.69,3.,1.36,6.37,4.55,0.52,0.87,1.51])
# create a big cube with shape: (T, Y, X)
arr = np.zeros([len(x), 145, 192])
for i in range(arr.shape[1]):
for j in range(arr.shape[2]):
arr[:, i, j] = x
print(arr.shape)
# re-arrange into tabular layout: (Y*X, T)
arr = np.transpose(arr, [1, 2, 0])
arr = arr.reshape(-1, len(x))
print(arr.shape)
import time
t1 = time.time()
z, p = MannKendallTrend2D(arr, tails=2, axis=1)
p = p.reshape(145, 192)
t2 = time.time()
print('time =', t2-t1)
The p-value for that sample time series is 0.63341565, which I have validated against the pymannkendall module result. Since arr contains merely duplicated copies of x, the resultant p is a 2d array of size (145, 192), with all 0.63341565.
And it took me only 1.28 seconds to compute that.
Very simple regression task. I have three variables x1, x2, x3 with some random noise. And I know target equation: y = q1*x1 + q2*x2 + q3*x3. Now I want to find target coefs: q1, q2, q3 evaluate the
performance using the mean Relative Squared Error (RSE) (Prediction/Real - 1)^2 to evaluate the performance of our prediction methods.
In the research, I see that this is ordinary Least Squares Problem. But I can't get from examples on the internet how to solve this particular problem in Python. Let say I have data:
import numpy as np
sourceData = np.random.rand(1000, 3)
koefs = np.array([1, 2, 3])
target = np.dot(sourceData, koefs)
(In real life that data are noisy, with not normal distribution.) How to find this koefs using Least Squares approach in python? Any lib usage.
#ayhan made a valuable comment.
And there is a problem with your code: Actually there is no noise in the data you collect. The input data is noisy, but after the multiplication, you don't add any additional noise.
I've added some noise to your measurements and used the least squares formula to fit the parameters, here's my code:
data = np.random.rand(1000,3)
true_theta = np.array([1,2,3])
true_measurements = np.dot(data, true_theta)
noise = np.random.rand(1000) * 1
noisy_measurements = true_measurements + noise
estimated_theta = np.linalg.inv(data.T # data) # data.T # noisy_measurements
The estimated_theta will be close to true_theta. If you don't add noise to the measurements, they will be equal.
I've used the python3 matrix multiplication syntax.
You could use np.dot instead of #
That makes the code longer, so I've split the formula:
MTM_inv = np.linalg.inv(np.dot(data.T, data))
MTy = np.dot(data.T, noisy_measurements)
estimated_theta = np.dot(MTM_inv, MTy)
You can read up on least squares here: https://en.wikipedia.org/wiki/Linear_least_squares_(mathematics)#The_general_problem
UPDATE:
Or you could just use the builtin least squares function:
np.linalg.lstsq(data, noisy_measurements)
In addition to the #lhk answer I have found great scipy Least Squares function. It is easy to get the requested behavior with it.
This way we can provide a custom function that returns residuals and form Relative Squared Error instead of absolute squared difference:
import numpy as np
from scipy.optimize import least_squares
data = np.random.rand(1000,3)
true_theta = np.array([1,2,3])
true_measurements = np.dot(data, true_theta)
noise = np.random.rand(1000) * 1
noisy_measurements = true_measurements + noise
#noisy_measurements[-1] = data[-1] # (1000 * true_theta) - uncoment this outliner to see how much Relative Squared Error esimator works better then default abs diff for this case.
def my_func(params, x, y):
res = (x # params) / y - 1 # If we change this line to: (x # params) - y - we will got the same result as np.linalg.lstsq
return res
res = least_squares(my_func, x0, args=(data, noisy_measurements) )
estimated_theta = res.x
Also, we can provide custom loss with loss argument function that will process the residuals and form final loss.
This link provides code for an autocorrelation-based pitch detection algorithm. I am using it to detect pitches in simple guitar melodies.
In general, it produces very good results. For example, for the melody C4, C#4, D4, D#4, E4 it outputs:
262.743653536
272.144441273
290.826273006
310.431336809
327.094621169
Which correlates to the correct notes.
However, in some cases like this audio file (E4, F4, F#4, G4, G#4, A4, A#4, B4) it produces errors:
325.861452246
13381.6439242
367.518651703
391.479384923
414.604661221
218.345286173
466.503751322
244.994090035
More specifically, there are three errors here: 13381Hz is wrongly detected instead of F4 (~350Hz) (weird error), and also 218Hz instead of A4 (440Hz) and 244Hz instead of B4 (~493Hz), which are octave errors.
I assume the two errors are caused by something different? Here is the code:
slices = segment_signal(y, sr)
for segment in slices:
pitch = freq_from_autocorr(segment, sr)
print pitch
def segment_signal(y, sr, onset_frames=None, offset=0.1):
if (onset_frames == None):
onset_frames = remove_dense_onsets(librosa.onset.onset_detect(y=y, sr=sr))
offset_samples = int(librosa.time_to_samples(offset, sr))
print onset_frames
slices = np.array([y[i : i + offset_samples] for i
in librosa.frames_to_samples(onset_frames)])
return slices
You can see the freq_from_autocorr function in the first link above.
The only think that I have changed is this line:
corr = corr[len(corr)/2:]
Which I have replaced with:
corr = corr[int(len(corr)/2):]
UPDATE:
I noticed the smallest the offset I use (the smallest the signal segment I use to detect each pitch), the more high-frequency (10000+ Hz) errors I get.
Specifically, I noticed that the part that goes differently in those cases (10000+ Hz) is the calculation of the i_peak value. When in cases with no error it is in the range of 50-150, in the case of the error it is 3-5.
The autocorrelation function in the code snippet that you linked is not particularly robust. In order to get the correct result, it needs to locate the first peak on the left hand side of the autocorrelation curve. The method that the other developer used (calling the numpy.argmax() function) does not always find the correct value.
I've implemented a slightly more robust version, using the peakutils package. I don't promise that it's perfectly robust either, but in any case it achieves a better result than the version of the freq_from_autocorr() function that you were previously using.
My example solution is listed below:
import librosa
import numpy as np
import matplotlib.pyplot as plt
from scipy.signal import fftconvolve
from pprint import pprint
import peakutils
def freq_from_autocorr(signal, fs):
# Calculate autocorrelation (same thing as convolution, but with one input
# reversed in time), and throw away the negative lags
signal -= np.mean(signal) # Remove DC offset
corr = fftconvolve(signal, signal[::-1], mode='full')
corr = corr[len(corr)//2:]
# Find the first peak on the left
i_peak = peakutils.indexes(corr, thres=0.8, min_dist=5)[0]
i_interp = parabolic(corr, i_peak)[0]
return fs / i_interp, corr, i_interp
def parabolic(f, x):
"""
Quadratic interpolation for estimating the true position of an
inter-sample maximum when nearby samples are known.
f is a vector and x is an index for that vector.
Returns (vx, vy), the coordinates of the vertex of a parabola that goes
through point x and its two neighbors.
Example:
Defining a vector f with a local maximum at index 3 (= 6), find local
maximum if points 2, 3, and 4 actually defined a parabola.
In [3]: f = [2, 3, 1, 6, 4, 2, 3, 1]
In [4]: parabolic(f, argmax(f))
Out[4]: (3.2142857142857144, 6.1607142857142856)
"""
xv = 1/2. * (f[x-1] - f[x+1]) / (f[x-1] - 2 * f[x] + f[x+1]) + x
yv = f[x] - 1/4. * (f[x-1] - f[x+1]) * (xv - x)
return (xv, yv)
# Time window after initial onset (in units of seconds)
window = 0.1
# Open the file and obtain the sampling rate
y, sr = librosa.core.load("./Vocaroo_s1A26VqpKgT0.mp3")
idx = np.arange(len(y))
# Set the window size in terms of number of samples
winsamp = int(window * sr)
# Calcualte the onset frames in the usual way
onset_frames = librosa.onset.onset_detect(y=y, sr=sr)
onstm = librosa.frames_to_time(onset_frames, sr=sr)
fqlist = [] # List of estimated frequencies, one per note
crlist = [] # List of autocorrelation arrays, one array per note
iplist = [] # List of peak interpolated peak indices, one per note
for tm in onstm:
startidx = int(tm * sr)
freq, corr, ip = freq_from_autocorr(y[startidx:startidx+winsamp], sr)
fqlist.append(freq)
crlist.append(corr)
iplist.append(ip)
pprint(fqlist)
# Choose which notes to plot (it's set to show all 8 notes in this case)
plidx = [0, 1, 2, 3, 4, 5, 6, 7]
# Plot amplitude curves of all notes in the plidx list
fgwin = plt.figure(figsize=[8, 10])
fgwin.subplots_adjust(bottom=0.0, top=0.98, hspace=0.3)
axwin = []
ii = 1
for tm in onstm[plidx]:
axwin.append(fgwin.add_subplot(len(plidx)+1, 1, ii))
startidx = int(tm * sr)
axwin[-1].plot(np.arange(startidx, startidx+winsamp), y[startidx:startidx+winsamp])
ii += 1
axwin[-1].set_xlabel('Sample ID Number', fontsize=18)
fgwin.show()
# Plot autocorrelation function of all notes in the plidx list
fgcorr = plt.figure(figsize=[8,10])
fgcorr.subplots_adjust(bottom=0.0, top=0.98, hspace=0.3)
axcorr = []
ii = 1
for cr, ip in zip([crlist[ii] for ii in plidx], [iplist[ij] for ij in plidx]):
if ii == 1:
shax = None
else:
shax = axcorr[0]
axcorr.append(fgcorr.add_subplot(len(plidx)+1, 1, ii, sharex=shax))
axcorr[-1].plot(np.arange(500), cr[0:500])
# Plot the location of the leftmost peak
axcorr[-1].axvline(ip, color='r')
ii += 1
axcorr[-1].set_xlabel('Time Lag Index (Zoomed)', fontsize=18)
fgcorr.show()
The printed output looks like:
In [1]: %run autocorr.py
[325.81996740236065,
346.43374761017725,
367.12435233192753,
390.17291696559079,
412.9358117076161,
436.04054933498134,
465.38986619237039,
490.34120132405866]
The first figure produced by my code sample depicts the amplitude curves for the next 0.1 seconds following each detected onset time:
The second figure produced by the code shows the autocorrelation curves, as computed inside of the freq_from_autocorr() function. The vertical red lines depict the location of the first peak on the left for each curve, as estimated by the peakutils package. The method used by the other developer was getting incorrect results for some of these red lines; that's why his version of that function was occasionally returning the wrong frequencies.
My suggestion would be to test the revised version of the freq_from_autocorr() function on other recordings, see if you can find more challenging examples where even the improved version still gives incorrect results, and then get creative and try to develop an even more robust peak finding algorithm that never, ever mis-fires.
The autocorrelation method is not always right. You may want to implement a more sophisticated method like YIN:
http://audition.ens.fr/adc/pdf/2002_JASA_YIN.pdf
or MPM:
http://www.cs.otago.ac.nz/tartini/papers/A_Smarter_Way_to_Find_Pitch.pdf
Both of the above papers are good reads.