markdown
stringlengths 0
37k
| code
stringlengths 1
33.3k
| path
stringlengths 8
215
| repo_name
stringlengths 6
77
| license
stringclasses 15
values |
|---|---|---|---|---|
Some Formal Basics (skip if you just want code examples)
To set the context, we here briefly describe statistical hypothesis testing. Informally, one defines a hypothesis on a certain domain and then uses a statistical test to check whether this hypothesis is true. Formally, the goal is to reject a so-called null-hypothesis $H_0$, which is the complement of an alternative-hypothesis $H_A$.
To distinguish the hypotheses, a test statistic is computed on sample data. Since sample data is finite, this corresponds to sampling the true distribution of the test statistic. There are two different distributions of the test statistic -- one for each hypothesis. The null-distribution corresponds to test statistic samples under the model that $H_0$ holds; the alternative-distribution corresponds to test statistic samples under the model that $H_A$ holds.
In practice, one tries to compute the quantile of the test statistic in the null-distribution. In case the test statistic is in a high quantile, i.e. it is unlikely that the null-distribution has generated the test statistic -- the null-hypothesis $H_0$ is rejected.
There are two different kinds of errors in hypothesis testing:
A type I error is made when $H_0: p=q$ is wrongly rejected. That is, the test says that the samples are from different distributions when they are not.
A type II error is made when $H_A: p\neq q$ is wrongly accepted. That is, the test says that the samples are from the same distribution when they are not.
A so-called consistent test achieves zero type II error for a fixed type I error.
To decide whether to reject $H_0$, one could set a threshold, say at the $95\%$ quantile of the null-distribution, and reject $H_0$ when the test statistic lies below that threshold. This means that the chance that the samples were generated under $H_0$ are $5\%$. We call this number the test power $\alpha$ (in this case $\alpha=0.05$). It is an upper bound on the probability for a type I error. An alternative way is simply to compute the quantile of the test statistic in the null-distribution, the so-called p-value, and to compare the p-value against a desired test power, say $\alpha=0.05$, by hand. The advantage of the second method is that one not only gets a binary answer, but also an upper bound on the type I error.
In order to construct a two-sample test, the null-distribution of the test statistic has to be approximated. One way of doing this for any two-sample test is called bootstrapping, or the permutation test, where samples from both sources are mixed and permuted repeatedly and the test statistic is computed for every of those configurations. While this method works for every statistical hypothesis test, it might be very costly because the test statistic has to be re-computed many times. For many test statistics, there are more sophisticated methods of approximating the null distribution.
Base class for Hypothesis Testing
Shogun implements statistical testing in the abstract class <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CHypothesisTest.html">CHypothesisTest</a>. All implemented methods will work with this interface at their most basic level. This class offers methods to
compute the implemented test statistic,
compute p-values for a given value of the test statistic,
compute a test threshold for a given p-value,
sampling the null distribution, i.e. perform the permutation test or bootstrappig of the null-distribution, and
performing a full two-sample test, and either returning a p-value or a binary rejection decision. This method is most useful in practice. Note that, depending on the used test statistic, it might be faster to call this than to compute threshold and test statistic seperately with the above methods.
There are special subclasses for testing two distributions against each other (<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CTwoSampleTest.html">CTwoSampleTest</a>, <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CIndependenceTest.html">CIndependenceTest</a>), kernel two-sample testing (<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CKernelTwoSampleTest.html">CKernelTwoSampleTest</a>), and kernel independence testing (<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CKernelIndependenceTest.html">CKernelIndependenceTest</a>), which however mostly differ in internals and constructors.
Kernel Two-Sample Testing with the Maximum Mean Discrepancy
$\DeclareMathOperator{\mmd}{MMD}$
An important class of hypothesis tests are the two-sample tests.
In two-sample testing, one tries to find out whether two sets of samples come from different distributions. Given two probability distributions $p,q$ on some arbritary domains $\mathcal{X}, \mathcal{Y}$ respectively, and i.i.d. samples $X={x_i}{i=1}^m\subseteq \mathcal{X}\sim p$ and $Y={y_i}{i=1}^n\subseteq \mathcal{Y}\sim p$, the two sample test distinguishes the hypothesises
\begin{align}
H_0: p=q\
H_A: p\neq q
\end{align}
In order to solve this problem, it is desirable to have a criterion than takes a positive unique value if $p\neq q$, and zero if and only if $p=q$. The so called Maximum Mean Discrepancy (MMD), has this property and allows to distinguish any two probability distributions, if used in a reproducing kernel Hilbert space (RKHS). It is the distance of the mean embeddings $\mu_p, \mu_q$ of the distributions $p,q$ in such a RKHS $\mathcal{F}$ -- which can also be expressed in terms of expectation of kernel functions, i.e.
\begin{align}
\mmd[\mathcal{F},p,q]&=||\mu_p-\mu_q||\mathcal{F}^2\
&=\textbf{E}{x,x'}\left[ k(x,x')\right]-
2\textbf{E}{x,y}\left[ k(x,y)\right]
+\textbf{E}{y,y'}\left[ k(y,y')\right]
\end{align}
Note that this formulation does not assume any form of the input data, we just need a kernel function whose feature space is a RKHS, see [2, Section 2] for details. This has the consequence that in Shogun, we can do tests on any type of data (<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CDenseFeatures.html">CDenseFeatures</a>, <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CSparseFeatures.html">CSparseFeatures</a>, <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CStringFeatures.html">CStringFeatures</a>, etc), as long as we or you provide a positive definite kernel function under the interface of <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CKernel.html">CKernel</a>.
We here only describe how to use the MMD for two-sample testing. Shogun offers two types of test statistic based on the MMD, one with quadratic costs both in time and space, and one with linear time and constant space costs. Both come in different versions and with different methods how to approximate the null-distribution in order to construct a two-sample test.
Running Example Data. Gaussian vs. Laplace
In order to illustrate kernel two-sample testing with Shogun, we use a couple of toy distributions. The first dataset we consider is the 1D Standard Gaussian
$p(x)=\frac{1}{\sqrt{2\pi\sigma^2}}\exp\left(-\frac{(x-\mu)^2}{\sigma^2}\right)$
with mean $\mu$ and variance $\sigma^2$, which is compared against the 1D Laplace distribution
$p(x)=\frac{1}{2b}\exp\left(-\frac{|x-\mu|}{b}\right)$
with the same mean $\mu$ and variance $2b^2$. In order to increase difficulty, we set $b=\sqrt{\frac{1}{2}}$, which means that $2b^2=\sigma^2=1$. | # use scipy for generating samples
from scipy.stats import norm, laplace
def sample_gaussian_vs_laplace(n=220, mu=0.0, sigma2=1, b=sqrt(0.5)):
# sample from both distributions
X=norm.rvs(size=n, loc=mu, scale=sigma2)
Y=laplace.rvs(size=n, loc=mu, scale=b)
return X,Y
mu=0.0
sigma2=1
b=sqrt(0.5)
n=220
X,Y=sample_gaussian_vs_laplace(n, mu, sigma2, b)
# plot both densities and histograms
figure(figsize=(18,5))
suptitle("Gaussian vs. Laplace")
subplot(121)
Xs=linspace(-2, 2, 500)
plot(Xs, norm.pdf(Xs, loc=mu, scale=sigma2))
plot(Xs, laplace.pdf(Xs, loc=mu, scale=b))
title("Densities")
xlabel("$x$")
ylabel("$p(x)$")
_=legend([ 'Gaussian','Laplace'])
subplot(122)
hist(X, alpha=0.5)
xlim([-5,5])
ylim([0,100])
hist(Y,alpha=0.5)
xlim([-5,5])
ylim([0,100])
legend(["Gaussian", "Laplace"])
_=title('Histograms') | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Now how to compare these two sets of samples? Clearly, a t-test would be a bad idea since it basically compares mean and variance of $X$ and $Y$. But we set that to be equal. By chance, the estimates of these statistics might differ, but that is unlikely to be significant. Thus, we have to look at higher order statistics of the samples. In fact, kernel two-sample tests look at all (infinitely many) higher order moments. | print "Gaussian vs. Laplace"
print "Sample means: %.2f vs %.2f" % (mean(X), mean(Y))
print "Samples variances: %.2f vs %.2f" % (var(X), var(Y)) | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Quadratic Time MMD
We now describe the quadratic time MMD, as described in [1, Lemma 6], which is implemented in Shogun. All methods in this section are implemented in <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CQuadraticTimeMMD.html">CQuadraticTimeMMD</a>, which accepts any type of features in Shogun, and use it on the above toy problem.
An unbiased estimate for the MMD expression above can be obtained by estimating expected values with averaging over independent samples
$$
\mmd_u[\mathcal{F},X,Y]^2=\frac{1}{m(m-1)}\sum_{i=1}^m\sum_{j\neq i}^mk(x_i,x_j) + \frac{1}{n(n-1)}\sum_{i=1}^n\sum_{j\neq i}^nk(y_i,y_j)-\frac{2}{mn}\sum_{i=1}^m\sum_{j\neq i}^nk(x_i,y_j)
$$
A biased estimate would be
$$
\mmd_b[\mathcal{F},X,Y]^2=\frac{1}{m^2}\sum_{i=1}^m\sum_{j=1}^mk(x_i,x_j) + \frac{1}{n^ 2}\sum_{i=1}^n\sum_{j=1}^nk(y_i,y_j)-\frac{2}{mn}\sum_{i=1}^m\sum_{j\neq i}^nk(x_i,y_j)
.$$
Computing the test statistic using <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CQuadraticTimeMMD.html">CQuadraticTimeMMD</a> does exactly this, where it is possible to choose between the two above expressions. Note that some methods for approximating the null-distribution only work with one of both types. Both statistics' computational costs are quadratic both in time and space. Note that the method returns $m\mmd_b[\mathcal{F},X,Y]^2$ since null distribution approximations work on $m$ times null distribution. Here is how the test statistic itself is computed. | # turn data into Shogun representation (columns vectors)
feat_p=RealFeatures(X.reshape(1,len(X)))
feat_q=RealFeatures(Y.reshape(1,len(Y)))
# choose kernel for testing. Here: Gaussian
kernel_width=1
kernel=GaussianKernel(10, kernel_width)
# create mmd instance of test-statistic
mmd=QuadraticTimeMMD(kernel, feat_p, feat_q)
# compute biased and unbiased test statistic (default is unbiased)
mmd.set_statistic_type(BIASED)
biased_statistic=mmd.compute_statistic()
mmd.set_statistic_type(UNBIASED)
unbiased_statistic=mmd.compute_statistic()
print "%d x MMD_b[X,Y]^2=%.2f" % (len(X), biased_statistic)
print "%d x MMD_u[X,Y]^2=%.2f" % (len(X), unbiased_statistic) | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Any sub-class of <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CHypothesisTest.html">CHypothesisTest</a> can compute approximate the null distribution using permutation/bootstrapping. This way always is guaranteed to produce consistent results, however, it might take a long time as for each sample of the null distribution, the test statistic has to be computed for a different permutation of the data. Note that each of the below calls samples from the null distribution. It is wise to choose one method in practice. Also note that we set the number of samples from the null distribution to a low value to reduce runtime. Choose larger in practice, it is in fact good to plot the samples. | # this is not necessary as bootstrapping is the default
mmd.set_null_approximation_method(PERMUTATION)
mmd.set_statistic_type(UNBIASED)
# to reduce runtime, should be larger practice
mmd.set_num_null_samples(100)
# now show a couple of ways to compute the test
# compute p-value for computed test statistic
p_value=mmd.compute_p_value(unbiased_statistic)
print "P-value of MMD value %.2f is %.2f" % (unbiased_statistic, p_value)
# compute threshold for rejecting H_0 for a given test power
alpha=0.05
threshold=mmd.compute_threshold(alpha)
print "Threshold for rejecting H0 with a test power of %.2f is %.2f" % (alpha, threshold)
# performing the test by hand given the above results, note that those two are equivalent
if unbiased_statistic>threshold:
print "H0 is rejected with confidence %.2f" % alpha
if p_value<alpha:
print "H0 is rejected with confidence %.2f" % alpha
# or, compute the full two-sample test directly
# fixed test power, binary decision
binary_test_result=mmd.perform_test(alpha)
if binary_test_result:
print "H0 is rejected with confidence %.2f" % alpha
significance_test_result=mmd.perform_test()
print "P-value of MMD test is %.2f" % significance_test_result
if significance_test_result<alpha:
print "H0 is rejected with confidence %.2f" % alpha | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Precomputing Kernel Matrices
Bootstrapping re-computes the test statistic for a bunch of permutations of the test data. For kernel two-sample test methods, in particular those of the MMD class, this means that only the joint kernel matrix of $X$ and $Y$ needs to be permuted. Thus, we can precompute the matrix, which gives a significant performance boost. Note that this is only possible if the matrix can be stored in memory. Below, we use Shogun's <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CCustomKernel.html">CCustomKernel</a> class, which allows to precompute a kernel matrix (multithreaded) of a given kernel and store it in memory. Instances of this class can then be used as if they were standard kernels. | # precompute kernel to be faster for null sampling
p_and_q=mmd.get_p_and_q()
kernel.init(p_and_q, p_and_q);
precomputed_kernel=CustomKernel(kernel);
mmd.set_kernel(precomputed_kernel);
# increase number of iterations since should be faster now
mmd.set_num_null_samples(500);
p_value_boot=mmd.perform_test();
print "P-value of MMD test is %.2f" % p_value_boot | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Now let us visualise distribution of MMD statistic under $H_0:p=q$ and $H_A:p\neq q$. Sample both null and alternative distribution for that. Use the interface of <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CTwoSampleTest.html">CTwoSampleTest</a> to sample from the null distribution (permutations, re-computing of test statistic is done internally). For the alternative distribution, compute the test statistic for a new sample set of $X$ and $Y$ in a loop. Note that the latter is expensive, as the kernel cannot be precomputed, and infinite data is needed. Though it is not needed in practice but only for illustrational purposes here. | num_samples=500
# sample null distribution
mmd.set_num_null_samples(num_samples)
null_samples=mmd.sample_null()
# sample alternative distribution, generate new data for that
alt_samples=zeros(num_samples)
for i in range(num_samples):
X=norm.rvs(size=n, loc=mu, scale=sigma2)
Y=laplace.rvs(size=n, loc=mu, scale=b)
feat_p=RealFeatures(reshape(X, (1,len(X))))
feat_q=RealFeatures(reshape(Y, (1,len(Y))))
mmd=QuadraticTimeMMD(kernel, feat_p, feat_q)
alt_samples[i]=mmd.compute_statistic() | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Null and Alternative Distribution Illustrated
Visualise both distributions, $H_0:p=q$ is rejected if a sample from the alternative distribution is larger than the $(1-\alpha)$-quantil of the null distribution. See [1] for more details on their forms. From the visualisations, we can read off the test's type I and type II error:
type I error is the area of the null distribution being right of the threshold
type II error is the area of the alternative distribution being left from the threshold | def plot_alt_vs_null(alt_samples, null_samples, alpha):
figure(figsize=(18,5))
subplot(131)
hist(null_samples, 50, color='blue')
title('Null distribution')
subplot(132)
title('Alternative distribution')
hist(alt_samples, 50, color='green')
subplot(133)
hist(null_samples, 50, color='blue')
hist(alt_samples, 50, color='green', alpha=0.5)
title('Null and alternative distriution')
# find (1-alpha) element of null distribution
null_samples_sorted=sort(null_samples)
quantile_idx=int(num_samples*(1-alpha))
quantile=null_samples_sorted[quantile_idx]
axvline(x=quantile, ymin=0, ymax=100, color='red', label=str(int(round((1-alpha)*100))) + '% quantile of null')
_=legend()
plot_alt_vs_null(alt_samples, null_samples, alpha) | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Different Ways to Approximate the Null Distribution for the Quadratic Time MMD
As already mentioned, bootstrapping the null distribution is expensive business. There exist a couple of methods that are more sophisticated and either allow very fast approximations without guarantees or reasonably fast approximations that are consistent. We present a selection from [2], which are implemented in Shogun.
The first one is a spectral method that is based around the Eigenspectrum of the kernel matrix of the joint samples. It is faster than bootstrapping while being a consistent test. Effectively, the null-distribution of the biased statistic is sampled, but in a more efficient way than the bootstrapping approach. The converges as
$$
m\mmd^2_b \rightarrow \sum_{l=1}^\infty \lambda_l z_l^2
$$
where $z_l\sim \mathcal{N}(0,2)$ are i.i.d. normal samples and $\lambda_l$ are Eigenvalues of expression 2 in [2], which can be empirically estimated by $\hat\lambda_l=\frac{1}{m}\nu_l$ where $\nu_l$ are the Eigenvalues of the centred kernel matrix of the joint samples $X$ and $Y$. The distribution above can be easily sampled. Shogun's implementation has two parameters:
Number of samples from null-distribution. The more, the more accurate. As a rule of thumb, use 250.
Number of Eigenvalues of the Eigen-decomposition of the kernel matrix to use. The more, the better the results get. However, the Eigen-spectrum of the joint gram matrix usually decreases very fast. Plotting the Spectrum can help. See [2] for details.
If the kernel matrices are diagonal dominant, this method is likely to fail. For that and more details, see the original paper. Computational costs are much lower than bootstrapping, which is the only consistent alternative. Since Eigenvalues of the gram matrix has to be computed, costs are in $\mathcal{O}(m^3)$.
Below, we illustrate how to sample the null distribution and perform two-sample testing with the Spectrum approximation in the class <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CQuadraticTimeMMD.html">CQuadraticTimeMMD</a>. This method only works with the biased statistic. | # optional: plot spectrum of joint kernel matrix
from numpy.linalg import eig
# get joint feature object and compute kernel matrix and its spectrum
feats_p_q=mmd.get_p_and_q()
mmd.get_kernel().init(feats_p_q, feats_p_q)
K=mmd.get_kernel().get_kernel_matrix()
w,_=eig(K)
# visualise K and its spectrum (only up to threshold)
figure(figsize=(18,5))
subplot(121)
imshow(K, interpolation="nearest")
title("Kernel matrix K of joint data $X$ and $Y$")
subplot(122)
thresh=0.1
plot(w[:len(w[w>thresh])])
_=title("Eigenspectrum of K until component %d" % len(w[w>thresh])) | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
The above plot of the Eigenspectrum shows that the Eigenvalues are decaying extremely fast. We choose the number for the approximation such that all Eigenvalues bigger than some threshold are used. In this case, we will not loose a lot of accuracy while gaining a significant speedup. For slower decaying Eigenspectrums, this approximation might be more expensive. | # threshold for eigenspectrum
thresh=0.1
# compute number of eigenvalues to use
num_eigen=len(w[w>thresh])
# finally, do the test, use biased statistic
mmd.set_statistic_type(BIASED)
#tell Shogun to use spectrum approximation
mmd.set_null_approximation_method(MMD2_SPECTRUM)
mmd.set_num_eigenvalues_spectrum(num_eigen)
mmd.set_num_samples_spectrum(num_samples)
# the usual test interface
p_value_spectrum=mmd.perform_test()
print "Spectrum: P-value of MMD test is %.2f" % p_value_spectrum
# compare with ground truth bootstrapping
mmd.set_null_approximation_method(PERMUTATION)
mmd.set_num_null_samples(num_samples)
p_value_boot=mmd.perform_test()
print "Bootstrapping: P-value of MMD test is %.2f" % p_value_spectrum | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
The Gamma Moment Matching Approximation and Type I errors
$\DeclareMathOperator{\var}{var}$
Another method for approximating the null-distribution is by matching the first two moments of a <a href="http://en.wikipedia.org/wiki/Gamma_distribution">Gamma distribution</a> and then compute the quantiles of that. This does not result in a consistent test, but usually also gives good results while being very fast. However, there are distributions where the method fail. Therefore, the type I error should always be monitored. Described in [2]. It uses
$$
m\mmd_b(Z) \sim \frac{x^{\alpha-1}\exp(-\frac{x}{\beta})}{\beta^\alpha \Gamma(\alpha)}
$$
where
$$
\alpha=\frac{(\textbf{E}(\text{MMD}_b(Z)))^2}{\var(\text{MMD}_b(Z))} \qquad \text{and} \qquad
\beta=\frac{m \var(\text{MMD}_b(Z))}{(\textbf{E}(\text{MMD}_b(Z)))^2}
$$
Then, any threshold and p-value can be computed using the gamma distribution in the above expression. Computational costs are in $\mathcal{O}(m^2)$. Note that the test is parameter free. It only works with the biased statistic. | # tell Shogun to use gamma approximation
mmd.set_null_approximation_method(MMD2_GAMMA)
# the usual test interface
p_value_gamma=mmd.perform_test()
print "Gamma: P-value of MMD test is %.2f" % p_value_gamma
# compare with ground truth bootstrapping
mmd.set_null_approximation_method(PERMUTATION)
p_value_boot=mmd.perform_test()
print "Bootstrapping: P-value of MMD test is %.2f" % p_value_spectrum | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
As we can see, the above example was kind of unfortunate, as the approximation fails badly. We check the type I error to verify that. This works similar to sampling the alternative distribution: re-sample data (assuming infinite amounts), perform the test and average results. Below we compare type I errors or all methods for approximating the null distribution. This will take a while. | # type I error is false alarm, therefore sample data under H0
num_trials=50
rejections_gamma=zeros(num_trials)
rejections_spectrum=zeros(num_trials)
rejections_bootstrap=zeros(num_trials)
num_samples=50
alpha=0.05
for i in range(num_trials):
X=norm.rvs(size=n, loc=mu, scale=sigma2)
Y=laplace.rvs(size=n, loc=mu, scale=b)
# simulate H0 via merging samples before computing the
Z=hstack((X,Y))
X=Z[:len(X)]
Y=Z[len(X):]
feat_p=RealFeatures(reshape(X, (1,len(X))))
feat_q=RealFeatures(reshape(Y, (1,len(Y))))
# gamma
mmd=QuadraticTimeMMD(kernel, feat_p, feat_q)
mmd.set_null_approximation_method(MMD2_GAMMA)
mmd.set_statistic_type(BIASED)
rejections_gamma[i]=mmd.perform_test(alpha)
# spectrum
mmd=QuadraticTimeMMD(kernel, feat_p, feat_q)
mmd.set_null_approximation_method(MMD2_SPECTRUM)
mmd.set_num_eigenvalues_spectrum(num_eigen)
mmd.set_num_samples_spectrum(num_samples)
mmd.set_statistic_type(BIASED)
rejections_spectrum[i]=mmd.perform_test(alpha)
# bootstrap (precompute kernel)
mmd=QuadraticTimeMMD(kernel, feat_p, feat_q)
p_and_q=mmd.get_p_and_q()
kernel.init(p_and_q, p_and_q)
precomputed_kernel=CustomKernel(kernel)
mmd.set_kernel(precomputed_kernel)
mmd.set_null_approximation_method(PERMUTATION)
mmd.set_num_null_samples(num_samples)
mmd.set_statistic_type(BIASED)
rejections_bootstrap[i]=mmd.perform_test(alpha)
convergence_gamma=cumsum(rejections_gamma)/(arange(num_trials)+1)
convergence_spectrum=cumsum(rejections_spectrum)/(arange(num_trials)+1)
convergence_bootstrap=cumsum(rejections_bootstrap)/(arange(num_trials)+1)
print "Average rejection rate of H0 for Gamma is %.2f" % mean(convergence_gamma)
print "Average rejection rate of H0 for Spectrum is %.2f" % mean(convergence_spectrum)
print "Average rejection rate of H0 for Bootstrapping is %.2f" % mean(rejections_bootstrap) | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
We see that Gamma basically never rejects, which is inline with the fact that the p-value was massively overestimated above. Note that for the other tests, the p-value is also not at its desired value, but this is due to the low number of samples/repetitions in the above code. Increasing them leads to consistent type I errors.
Linear Time MMD on Gaussian Blobs
So far, we basically had to precompute the kernel matrix for reasonable runtimes. This is not possible for more than a few thousand points. The linear time MMD statistic, implemented in <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CLinearTimeMMD.html">CLinearTimeMMD</a> can help here, as it accepts data under the streaming interface <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CStreamingFeatures.html">CStreamingFeatures</a>, which deliver data one-by-one.
And it can do more cool things, for example choose the best single (or combined) kernel for you. But we need a more fancy dataset for that to show its power. We will use one of Shogun's streaming based data generator, <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CGaussianBlobsDataGenerator.html">CGaussianBlobsDataGenerator</a> for that. This dataset consists of two distributions which are a grid of Gaussians where in one of them, the Gaussians are stretched and rotated. This dataset is regarded as challenging for two-sample testing. | # paramters of dataset
m=20000
distance=10
stretch=5
num_blobs=3
angle=pi/4
# these are streaming features
gen_p=GaussianBlobsDataGenerator(num_blobs, distance, 1, 0)
gen_q=GaussianBlobsDataGenerator(num_blobs, distance, stretch, angle)
# stream some data and plot
num_plot=1000
features=gen_p.get_streamed_features(num_plot)
features=features.create_merged_copy(gen_q.get_streamed_features(num_plot))
data=features.get_feature_matrix()
figure(figsize=(18,5))
subplot(121)
grid(True)
plot(data[0][0:num_plot], data[1][0:num_plot], 'r.', label='$x$')
title('$X\sim p$')
subplot(122)
grid(True)
plot(data[0][num_plot+1:2*num_plot], data[1][num_plot+1:2*num_plot], 'b.', label='$x$', alpha=0.5)
_=title('$Y\sim q$') | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
We now describe the linear time MMD, as described in [1, Section 6], which is implemented in Shogun. A fast, unbiased estimate for the original MMD expression which still uses all available data can be obtained by dividing data into two parts and then compute
$$
\mmd_l^2[\mathcal{F},X,Y]=\frac{1}{m_2}\sum_{i=1}^{m_2} k(x_{2i},x_{2i+1})+k(y_{2i},y_{2i+1})-k(x_{2i},y_{2i+1})-
k(x_{2i+1},y_{2i})
$$
where $ m_2=\lfloor\frac{m}{2} \rfloor$. While the above expression assumes that $m$ data are available from each distribution, the statistic in general works in an online setting where features are obtained one by one. Since only pairs of four points are considered at once, this allows to compute it on data streams. In addition, the computational costs are linear in the number of samples that are considered from each distribution. These two properties make the linear time MMD very applicable for large scale two-sample tests. In theory, any number of samples can be processed -- time is the only limiting factor.
We begin by illustrating how to pass data to <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CLinearTimeMMD.html">CLinearTimeMMD</a>. In order not to loose performance due to overhead, it is possible to specify a block size for the data stream. | block_size=100
# if features are already under the streaming interface, just pass them
mmd=LinearTimeMMD(kernel, gen_p, gen_q, m, block_size)
# compute an unbiased estimate in linear time
statistic=mmd.compute_statistic()
print "MMD_l[X,Y]^2=%.2f" % statistic
# note: due to the streaming nature, successive calls of compute statistic use different data
# and produce different results. Data cannot be stored in memory
for _ in range(5):
print "MMD_l[X,Y]^2=%.2f" % mmd.compute_statistic() | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Sometimes, one might want to use <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CLinearTimeMMD.html">CLinearTimeMMD</a> with data that is stored in memory. In that case, it is easy to data in the form of for example <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CStreamingDenseFeatures.html">CStreamingDenseFeatures</a> into <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CDenseFeatures.html">CDenseFeatures</a>. | # data source
gen_p=GaussianBlobsDataGenerator(num_blobs, distance, 1, 0)
gen_q=GaussianBlobsDataGenerator(num_blobs, distance, stretch, angle)
# retreive some points, store them as non-streaming data in memory
data_p=gen_p.get_streamed_features(100)
data_q=gen_q.get_streamed_features(data_p.get_num_vectors())
print "Number of data is %d" % data_p.get_num_vectors()
# cast data in memory as streaming features again (which now stream from the in-memory data)
streaming_p=StreamingRealFeatures(data_p)
streaming_q=StreamingRealFeatures(data_q)
# it is important to start the internal parser to avoid deadlocks
streaming_p.start_parser()
streaming_q.start_parser()
# example to create mmd (note that m can be maximum the number of data in memory)
mmd=LinearTimeMMD(GaussianKernel(10,1), streaming_p, streaming_q, data_p.get_num_vectors(), 1)
print "Linear time MMD statistic: %.2f" % mmd.compute_statistic() | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
The Gaussian Approximation to the Null Distribution
As for any two-sample test in Shogun, bootstrapping can be used to approximate the null distribution. This results in a consistent, but slow test. The number of samples to take is the only parameter. Note that since <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CLinearTimeMMD.html">CLinearTimeMMD</a> operates on streaming features, new data is taken from the stream in every iteration.
Bootstrapping is not really necessary since there exists a fast and consistent estimate of the null-distribution. However, to ensure that any approximation is accurate, it should always be checked against bootstrapping at least once.
Since both the null- and the alternative distribution of the linear time MMD are Gaussian with equal variance (and different mean), it is possible to approximate the null-distribution by using a linear time estimate for this variance. An unbiased, linear time estimator for
$$
\var[\mmd_l^2[\mathcal{F},X,Y]]
$$
can simply be computed by computing the empirical variance of
$$
k(x_{2i},x_{2i+1})+k(y_{2i},y_{2i+1})-k(x_{2i},y_{2i+1})-k(x_{2i+1},y_{2i}) \qquad (1\leq i\leq m_2)
$$
A normal distribution with this variance and zero mean can then be used as an approximation for the null-distribution. This results in a consistent test and is very fast. However, note that it is an approximation and its accuracy depends on the underlying data distributions. It is a good idea to compare to the bootstrapping approach first to determine an appropriate number of samples to use. This number is usually in the tens of thousands.
<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CLinearTimeMMD.html">CLinearTimeMMD</a> allows to approximate the null distribution in the same pass as computing the statistic itself (in linear time). This should always be used in practice since seperate calls of computing statistic and p-value will operator on different data from the stream. Below, we compute the test on a large amount of data (impossible to perform quadratic time MMD for this one as the kernel matrices cannot be stored in memory) | mmd=LinearTimeMMD(kernel, gen_p, gen_q, m, block_size)
print "m=%d samples from p and q" % m
print "Binary test result is: " + ("Rejection" if mmd.perform_test(alpha) else "No rejection")
print "P-value test result is %.2f" % mmd.perform_test() | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Kernel Selection for the MMD -- Overview
$\DeclareMathOperator{\argmin}{arg\,min}
\DeclareMathOperator{\argmax}{arg\,max}$
Now which kernel do we actually use for our tests? So far, we just plugged in arbritary ones. However, for kernel two-sample testing, it is possible to do something more clever.
Shogun's kernel selection methods for MMD based two-sample tests are all based around [3, 4]. For the <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CLinearTimeMMD.html">CLinearTimeMMD</a>, [3] describes a way of selecting the optimal kernel in the sense that the test's type II error is minimised. For the linear time MMD, this is the method of choice. It is done via maximising the MMD statistic divided by its standard deviation and it is possible for single kernels and also for convex combinations of them. For the <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CQuadraticTimeMMD.html">CQuadraticTimeMMD</a>, the best method in literature is choosing the kernel that maximised the MMD statistic [4]. For convex combinations of kernels, this can be achieved via a $L2$ norm constraint. A detailed comparison of all methods on numerous datasets can be found in [5].
MMD Kernel selection in Shogun always involves an implementation of the base class <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelection.html">CMMDKernelSelection</a>, which defines the interface for kernel selection. If combinations of kernel should be considered, there is a sub-class <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionComb.html">CMMDKernelSelectionComb</a>. In addition, it involves setting up a number of baseline kernels $\mathcal{K}$ to choose from/combine in the form of a <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CCombinedKernel.html">CCombinedKernel</a>. All methods compute their results for a fixed set of these baseline kernels. We later give an example how to use these classes after providing a list of available methods.
<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionMedian.html">CMMDKernelSelectionMedian</a> Selects from a set <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CGaussianKernel.html">CGaussianKernel</a> instances the one whose width parameter is closest to the median of the pairwise distances in the data. The median is computed on a certain number of points from each distribution that can be specified as a parameter. Since the median is a stable statistic, one does not have to compute all pairwise distances but rather just a few thousands. This method a useful (and fast) heuristic that in many cases gives a good hint on where to start looking for Gaussian kernel widths. It is for example described in [1]. Note that it may fail badly in selecting a good kernel for certain problems.
<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionMax.html">CMMDKernelSelectionMax</a> Selects from a set of arbitrary baseline kernels a single one that maximises the used MMD statistic -- more specific its estimate.
$$
k^*=\argmax_{k\in\mathcal{K}} \hat \eta_k,
$$
where $\eta_k$ is an empirical MMD estimate for using a kernel $k$.
This was first described in [4] and was empirically shown to perform better than the median heuristic above. However, it remains a heuristic that comes with no guarantees. Since MMD estimates can be computed in linear and quadratic time, this method works for both methods. However, for the linear time statistic, there exists a better method.
<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionOpt.html">CMMDKernelSelectionOpt</a> Selects the optimal single kernel from a set of baseline kernels. This is done via maximising the ratio of the linear MMD statistic and its standard deviation.
$$
k^=\argmax_{k\in\mathcal{K}} \frac{\hat \eta_k}{\hat\sigma_k+\lambda},
$$
where $\eta_k$ is a linear time MMD estimate for using a kernel $k$ and $\hat\sigma_k$ is a linear time variance estimate of $\eta_k$ to which a small number $\lambda$ is added to prevent division by zero.
These are estimated in a linear time way with the streaming framework that was described earlier. Therefore, this method is only available for <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CLinearTimeMMD.html">CLinearTimeMMD</a>. Optimal here means that the resulting test's type II error is minimised for a fixed type I error. Important: For this method to work, the kernel needs to be selected on different* data than the test is performed on. Otherwise, the method will produce wrong results.
<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionCombMaxL2.html">CMMDKernelSelectionCombMaxL2</a> Selects a convex combination of kernels that maximises the MMD statistic. This is the multiple kernel analogous to <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionMax.html">CMMDKernelSelectionMax</a>. This is done via solving the convex program
$$
\boldsymbol{\beta}^*=\min_{\boldsymbol{\beta}} {\boldsymbol{\beta}^T\boldsymbol{\beta} : \boldsymbol{\beta}^T\boldsymbol{\eta}=\mathbf{1}, \boldsymbol{\beta}\succeq 0},
$$
where $\boldsymbol{\beta}$ is a vector of the resulting kernel weights and $\boldsymbol{\eta}$ is a vector of which each component contains a MMD estimate for a baseline kernel. See [3] for details. Note that this method is unable to select a single kernel -- even when this would be optimal.
Again, when using the linear time MMD, there are better methods available.
<a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionCombOpt.html">CMMDKernelSelectionCombOpt</a> Selects a convex combination of kernels that maximises the MMD statistic divided by its covariance. This corresponds to \emph{optimal} kernel selection in the same sense as in class <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionOpt.html">CMMDKernelSelectionOpt</a> and is its multiple kernel analogous. The convex program to solve is
$$
\boldsymbol{\beta}^*=\min_{\boldsymbol{\beta}} (\hat Q+\lambda I) {\boldsymbol{\beta}^T\boldsymbol{\beta} : \boldsymbol{\beta}^T\boldsymbol{\eta}=\mathbf{1}, \boldsymbol{\beta}\succeq 0},
$$
where again $\boldsymbol{\beta}$ is a vector of the resulting kernel weights and $\boldsymbol{\eta}$ is a vector of which each component contains a MMD estimate for a baseline kernel. The matrix $\hat Q$ is a linear time estimate of the covariance matrix of the vector $\boldsymbol{\eta}$ to whose diagonal a small number $\lambda$ is added to prevent division by zero. See [3] for details. In contrast to <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionCombMaxL2.html">CMMDKernelSelectionCombMaxL2</a>, this method is able to select a single kernel when this gives a lower type II error than a combination. In this sense, it contains <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CMMDKernelSelectionOpt.html">CMMDKernelSelectionOpt</a>.
MMD Kernel Selection in Shogun
In order to use one of the above methods for kernel selection, one has to create a new instance of <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CCombinedKernel.html">CCombinedKernel</a> append all desired baseline kernels to it. This combined kernel is then passed to the MMD class. Then, an object of any of the above kernel selection methods is created and the MMD instance is passed to it in the constructor. There are then multiple methods to call
compute_measures to compute a vector kernel selection criteria if a single kernel selection method is used. It will return a vector of selected kernel weights if a combined kernel selection method is used. For \shogunclass{CMMDKernelSelectionMedian}, the method does throw an error.
select_kernel returns the selected kernel of the method. For single kernels this will be one of the baseline kernel instances. For the combined kernel case, this will be the underlying <a href="http://www.shogun-toolbox.org/doc/en/latest/classshogun_1_1CCombinedKernel.html">CCombinedKernel</a> instance where the subkernel weights are set to the weights that were selected by the method.
In order to utilise the selected kernel, it has to be passed to an MMD instance. We now give an example how to select the optimal single and combined kernel for the Gaussian Blobs dataset.
What is the best kernel to use here? This is tricky since the distinguishing characteristics are hidden at a small length-scale. Create some kernels to select the best from | sigmas=[2**x for x in linspace(-5,5, 10)]
print "Choosing kernel width from", ["{0:.2f}".format(sigma) for sigma in sigmas]
combined=CombinedKernel()
for i in range(len(sigmas)):
combined.append_kernel(GaussianKernel(10, sigmas[i]))
# mmd instance using streaming features
block_size=1000
mmd=LinearTimeMMD(combined, gen_p, gen_q, m, block_size)
# optmal kernel choice is possible for linear time MMD
selection=MMDKernelSelectionOpt(mmd)
# select best kernel
best_kernel=selection.select_kernel()
best_kernel=GaussianKernel.obtain_from_generic(best_kernel)
print "Best single kernel has bandwidth %.2f" % best_kernel.get_width() | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
Now perform two-sample test with that kernel | alpha=0.05
mmd=LinearTimeMMD(best_kernel, gen_p, gen_q, m, block_size)
mmd.set_null_approximation_method(MMD1_GAUSSIAN);
p_value_best=mmd.perform_test();
print "Bootstrapping: P-value of MMD test with optimal kernel is %.2f" % p_value_best | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
For the linear time MMD, the null and alternative distributions look different than for the quadratic time MMD as plotted above. Let's sample them (takes longer, reduce number of samples a bit). Note how we can tell the linear time MMD to smulate the null hypothesis, which is necessary since we cannot permute by hand as samples are not in memory) | mmd=LinearTimeMMD(best_kernel, gen_p, gen_q, 5000, block_size)
num_samples=500
# sample null and alternative distribution, implicitly generate new data for that
null_samples=zeros(num_samples)
alt_samples=zeros(num_samples)
for i in range(num_samples):
alt_samples[i]=mmd.compute_statistic()
# tell MMD to merge data internally while streaming
mmd.set_simulate_h0(True)
null_samples[i]=mmd.compute_statistic()
mmd.set_simulate_h0(False) | doc/ipython-notebooks/statistics/mmd_two_sample_testing.ipynb | hongguangguo/shogun | gpl-3.0 |
ICING tutorial
<hr>
ICING is a IG clonotype inference library developed in Python.
<font color="red"><b>NB:</b></font> This is <font color="red"><b>NOT</b></font> a quickstart guide for ICING. This intended as a detailed tutorial on how ICING works internally. If you're only interested into using ICING, please refer to the Quickstart Manual on github, or the <font color="blue">Quickstart section at the end of this notebook</font>.
ICING needs as input a file (TAB-delimited or CSV) which contains, in each row, a particular sequence.
The format used is the same as returned by Change-O's MakeDb.py script, which, starting from a IMGT results, it builds a single file with all the information extracted from IMGT starting from the RAW fasta sequences.
0. Data loading
Load the dataset into a single pandas dataframe called 'df'.
The dataset MUST CONTAIN at least the following columns (NOT case-sensitive):
- SEQUENCE_ID
- V_CALL
- J_CALL
- JUNCTION
- MUT (only if correct is True) | db_file = '../examples/data/clones_100.100.tab'
# dialect="excel" for CSV or XLS files
# for computational reasons, let's limit the dataset to the first 1000 sequences
X = io.load_dataframe(db_file, dialect="excel-tab")[:1000]
# turn the following off if data are real
# otherwise, assume that the "SEQUENCE_ID" field is composed as
# "[db]_[extension]_[id]_[id-true-clonotype]_[other-info]"
# See the example file for the format of the input.
X['true_clone'] = [x[3] for x in X.sequence_id.str.split('_')] | notebooks/icing_tutorial.ipynb | slipguru/ignet | bsd-2-clause |
1. Preprocessing step: data shrinking
Specially in CLL patients, most of the input sequences have the same V genes AND junction. In this case, it is possible to remove such sequences from the analysis (we just need to remember them after.)
In other words, we can collapse repeated sequences into a single one, which will weight as high as the number of sequences it represents. | # group by junction and v genes
groups = X.groupby(["v_gene_set_str", "junc"]).groups.values()
idxs = np.array([elem[0] for elem in groups]) # take one of them
weights = np.array([len(elem) for elem in groups]) # assign its weight | notebooks/icing_tutorial.ipynb | slipguru/ignet | bsd-2-clause |
2. High-level group inference
The number of sequences at this point may be still very high, in particular when IGs are mutated and there is not much replication. However, we rely on the fact that IG similarity is mainly constrained on their junction length. Therefore, we infer high-level groups based on their junction lengths.
This is a fast and efficient step. Also, by exploiting MiniBatchKMeans, we can specify an upperbound on the number of clusters we want to obtain. However, contrary to the standard KMeans algorithm, in this case some clusters may vanish. If one is expected to have related IGs with very different junction lengths, however, it is reasonable to specify a low value of clusters.
Keep in mind, however, that a low number of clusters correspond to higher computational workload of the method in the next phases. | n_clusters = 50
X_all = idxs.reshape(-1,1)
kmeans = MiniBatchKMeans(n_init=100, n_clusters=min(n_clusters, X_all.shape[0]))
lengths = X['junction_length'].values
kmeans.fit(lengths[idxs].reshape(-1,1)) | notebooks/icing_tutorial.ipynb | slipguru/ignet | bsd-2-clause |
3. Fine-grained group inference
Now we have higih-level groups of IGs we have to extract clonotypes from.
Divide the dataset based on the labels extracted from MiniBatchKMeans.
For each one of the cluster, find clonotypes contained in it using DBSCAN.
This algorithm allows us to use a custom metric between IGs.
[<font color='blue'><b>ADVANCED</b></font>] To develop a custom metric, see the format of icing.core.distances.distance_dataframe. If you use a custom function, then you only need to put it as parameter of DBSCAN metric. Note that partial is required if the metric has more than 2 parameters. To be a valid metric for DBSCAN, the function must take ONLY two params (the two elements to compare). For this reason, the other arguments are pre-computed with partial in the following example. | dbscan = DBSCAN(min_samples=20, n_jobs=-1, algorithm='brute', eps=0.2,
metric=partial(distance_dataframe, X,
junction_dist=distances.StringDistance(model='ham'),
correct=True, tol=0))
dbscan_labels = np.zeros_like(kmeans.labels_).ravel()
for label in np.unique(kmeans.labels_):
idx_row = np.where(kmeans.labels_ == label)[0]
X_idx = idxs[idx_row].reshape(-1,1).astype('float64')
weights_idx = weights[idx_row]
if idx_row.size == 1:
db_labels = np.array([0])
db_labels = dbscan.fit_predict(X_idx, sample_weight=weights_idx)
if len(dbscan.core_sample_indices_) < 1:
db_labels[:] = 0
if -1 in db_labels:
# this means that DBSCAN found some IG as noise. We choose to assign to the nearest cluster
balltree = BallTree(
X_idx[dbscan.core_sample_indices_],
metric=dbscan.metric)
noise_labels = balltree.query(
X_idx[db_labels == -1], k=1, return_distance=False).ravel()
# get labels for core points, then assign to noise points based
# on balltree
dbscan_noise_labels = db_labels[
dbscan.core_sample_indices_][noise_labels]
db_labels[db_labels == -1] = dbscan_noise_labels
# hopefully, there are no noisy samples at this time
db_labels[db_labels > -1] = db_labels[db_labels > -1] + np.max(dbscan_labels) + 1
dbscan_labels[idx_row] = db_labels # + np.max(dbscan_labels) + 1
labels = dbscan_labels
# new part: put together the labels
labels_ext = np.zeros(X.shape[0], dtype=int)
labels_ext[idxs] = labels
for i, list_ in enumerate(groups):
labels_ext[list_] = labels[i]
labels = labels_ext | notebooks/icing_tutorial.ipynb | slipguru/ignet | bsd-2-clause |
Quickstart
<hr>
All of the above-mentioned steps are integrated in ICING with a simple call to the class inference.ICINGTwoStep.
The following is an example of a working script. | db_file = '../examples/data/clones_100.100.tab'
correct = True
tolerance = 0
X = io.load_dataframe(db_file)[:1000]
# turn the following off if data are real
X['true_clone'] = [x[3] for x in X.sequence_id.str.split('_')]
true_clones = LabelEncoder().fit_transform(X.true_clone.values)
ii = inference.ICINGTwoStep(
model='nt', eps=0.2, method='dbscan', verbose=True,
kmeans_params=dict(n_init=100, n_clusters=20),
dbscan_params=dict(min_samples=20, n_jobs=-1, algorithm='brute',
metric=partial(distance_dataframe, X, **dict(
junction_dist=StringDistance(model='ham'),
correct=correct, tol=tolerance))))
tic = time.time()
labels = ii.fit_predict(X)
tac = time.time() - tic
print("\nElapsed time: %.1fs" % tac) | notebooks/icing_tutorial.ipynb | slipguru/ignet | bsd-2-clause |
If you want to save the results: | X['icing_clones (%s)' % ('_'.join(('StringDistance', str(eps), '0', 'corr' if correct else 'nocorr',
"%.4f" % tac)))] = labels
X.to_csv(db_file.split('/')[-1] + '_icing.csv') | notebooks/icing_tutorial.ipynb | slipguru/ignet | bsd-2-clause |
How is the result? | from sklearn import metrics
true_clones = LabelEncoder().fit_transform(X.true_clone.values)
print "FMI: %.5f" % (metrics.fowlkes_mallows_score(true_clones, labels))
print "ARI: %.5f" % (metrics.adjusted_rand_score(true_clones, labels))
print "AMI: %.5f" % (metrics.adjusted_mutual_info_score(true_clones, labels))
print "NMI: %.5f" % (metrics.normalized_mutual_info_score(true_clones, labels))
print "Hom: %.5f" % (metrics.homogeneity_score(true_clones, labels))
print "Com: %.5f" % (metrics.completeness_score(true_clones, labels))
print "Vsc: %.5f" % (metrics.v_measure_score(true_clones, labels)) | notebooks/icing_tutorial.ipynb | slipguru/ignet | bsd-2-clause |
Is it better or worse than the result with everyone at the same time? | labels = dbscan.fit_predict(np.arange(X.shape[0]).reshape(-1, 1))
print "FMI: %.5f" % metrics.fowlkes_mallows_score(true_clones, labels)
print "ARI: %.5f" % (metrics.adjusted_rand_score(true_clones, labels))
print "AMI: %.5f" % (metrics.adjusted_mutual_info_score(true_clones, labels))
print "NMI: %.5f" % (metrics.normalized_mutual_info_score(true_clones, labels))
print "Hom: %.5f" % (metrics.homogeneity_score(true_clones, labels))
print "Com: %.5f" % (metrics.completeness_score(true_clones, labels))
print "Vsc: %.5f" % (metrics.v_measure_score(true_clones, labels)) | notebooks/icing_tutorial.ipynb | slipguru/ignet | bsd-2-clause |
Now fit using the XID+ interface to pystan | %%time
from xidplus.stan_fit import SPIRE
fit=SPIRE.all_bands(prior250,prior350,prior500,iter=1000)
| docs/notebooks/examples/XID+example_run_script.ipynb | pdh21/XID_plus | mit |
Initialise the posterior class with the fit object from pystan, and save alongside the prior classes | posterior=xidplus.posterior_stan(fit,[prior250,prior350,prior500])
xidplus.save([prior250,prior350,prior500],posterior,'test') | docs/notebooks/examples/XID+example_run_script.ipynb | pdh21/XID_plus | mit |
Alternatively, you can fit with the pyro backend. | %%time
from xidplus.pyro_fit import SPIRE
fit_pyro=SPIRE.all_bands([prior250,prior350,prior500],n_steps=10000,lr=0.001,sub=0.1)
posterior_pyro=xidplus.posterior_pyro(fit_pyro,[prior250,prior350,prior500])
xidplus.save([prior250,prior350,prior500],posterior_pyro,'test_pyro')
plt.semilogy(posterior_pyro.loss_history) | docs/notebooks/examples/XID+example_run_script.ipynb | pdh21/XID_plus | mit |
You can fit with the numpyro backend. | %%time
from xidplus.numpyro_fit import SPIRE
fit_numpyro=SPIRE.all_bands([prior250,prior350,prior500])
posterior_numpyro=xidplus.posterior_numpyro(fit_numpyro,[prior250,prior350,prior500])
xidplus.save([prior250,prior350,prior500],posterior_numpyro,'test_numpyro')
prior250.bkg | docs/notebooks/examples/XID+example_run_script.ipynb | pdh21/XID_plus | mit |
We will want to run the notebook in the future with updated values. How can we do this? Make the dates automatically updated. | start = datetime.datetime(2017, 3, 2) # the day Snap went public
end = datetime.date.today() # datetime.date.today
snap = web.DataReader("SNAP", 'google', start, end)
snap
snap.index.tolist() | Code/notebooks/bootcamp_format_plotting.ipynb | NYUDataBootcamp/Materials | mit |
.format()
We want to print something with systematic changes in the text.
Suppose we want to print out the following information:
'On day X Snap closed at VALUE Y and the volume was Z.' | # How did we do this before?
for index in snap.index:
print('On day', index, 'Snap closed at', snap['Close'][index], 'and the volume was', snap['Volume'][index], '.')
| Code/notebooks/bootcamp_format_plotting.ipynb | NYUDataBootcamp/Materials | mit |
This looks aweful. We want to cut the day and express the volume in millions. | # express Volume in millions
snap['Volume'] = snap['Volume']/10**6
snap | Code/notebooks/bootcamp_format_plotting.ipynb | NYUDataBootcamp/Materials | mit |
The .format() method
what is format and how does it work? Google and find a good link | print('Today is {}.'.format(datetime.date.today()))
for index in snap.index:
print('On {} Snap closed at ${} and the volume was {} million.'.format(index, snap['Close'][index], snap['Volume'][index]))
for index in snap.index:
print('On {:.10} Snap closed at ${} and the volume was {:.1f} million.'.format(str(index), snap['Close'][index], snap['Volume'][index]))
| Code/notebooks/bootcamp_format_plotting.ipynb | NYUDataBootcamp/Materials | mit |
Check Olson's blog and style recommendation | fig, ax = plt.subplots() #figsize=(8,5))
snap['Close'].plot(ax=ax, grid=True, style='o', alpha=.6)
ax.set_xlim([snap.index[0]-datetime.timedelta(days=1), snap.index[-1]+datetime.timedelta(days=1)])
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.yaxis.set_ticks_position('left')
ax.xaxis.set_ticks_position('bottom')
ax.vlines(snap.index, snap['Low'], snap['High'], alpha=.2, lw=.9)
ax.set_ylabel('SNAP share price', fontsize=14)
ax.set_xlabel('Date', fontsize=14)
plt.show()
start_w = datetime.datetime(2008, 6, 8)
oilwater = web.DataReader(['BP', 'AWK'], 'google', start_w, end)
oilwater.describe
type(oilwater[:,:,'AWK'])
water = oilwater[:, :, 'AWK']
oil = oilwater[:, :, 'BP']
#import seaborn as sns
#import matplotlib as mpl
#mpl.rcParams.update(mpl.rcParamsDefault)
plt.style.use('seaborn-notebook')
plt.rc('font', family='serif')
deepwater = datetime.datetime(2010, 4, 20)
fig, ax = plt.subplots(figsize=(8, 5))
water['Close'].plot(ax=ax, label='AWK', lw=.7) #grid=True,
oil['Close'].plot(ax=ax, label='BP', lw=.7) #grid=True,
ax.yaxis.grid(True)
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.yaxis.set_ticks_position('left')
ax.xaxis.set_ticks_position('bottom')
ax.vlines(deepwater, 0, 100, linestyles='dashed', alpha=.6)
ax.text(deepwater, 70, 'Deepwater catastrophe', horizontalalignment='center')
ax.set_ylim([0, 100])
ax.legend(bbox_to_anchor=(1.2, .9), frameon=False)
plt.show()
print(plt.style.available)
fig, ax = plt.subplots(figsize=(8, 5))
water['AWK_pct_ch'] = water['Close'].diff().cumsum()/water['Close'].iloc[0]
oil['BP_pct_ch'] = oil['Close'].diff().cumsum()/oil['Close'].iloc[0]
#water['Close'].pct_change().cumsum().plot(ax=ax, label='AWK')
water['AWK_pct_ch'].plot(ax=ax, label='AWK', lw=.7)
#oil['Close'].pct_change().cumsum().plot(ax=ax, label='BP')
oil['BP_pct_ch'].plot(ax=ax, label='BP', lw=.7)
ax.yaxis.grid(True)
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.yaxis.set_ticks_position('left')
ax.xaxis.set_ticks_position('bottom')
ax.vlines(deepwater, -1, 3, linestyles='dashed', alpha=.6)
ax.text(deepwater, 1.2, 'Deepwater catastrophe', horizontalalignment='center')
ax.set_ylim([-1, 3])
ax.legend(bbox_to_anchor=(1.2, .9), frameon=False)
ax.set_title('Percentage change relative to {:.10}\n'.format(str(start_w)), fontsize=14, loc='left')
plt.show() | Code/notebooks/bootcamp_format_plotting.ipynb | NYUDataBootcamp/Materials | mit |
Machine Dependent Options
Each installation of GPy also creates an installation.cfg file. This file should include any installation specific settings for your GPy installation. For example, if a particular machine is set up to run OpenMP then the installation.cfg file should contain | # This is the local installation configuration file for GPy
[parallel]
openmp=True | GPy/config.ipynb | SheffieldML/notebook | bsd-3-clause |
在可变的集合类型中（list和dictionary）中，如果默认参数为该类型，那么所有的操作调用该函数的操作将会发生变化 | def foo(values, x=[]):
for value in values:
x.append(value)
return x
foo([0,1,2])
foo([4,5])
def foo_fix(values, x=[]):
if len(x) != 0:
x = []
for value in values:
x.append(value)
return x
foo_fix([0,1,2])
foo_fix([4,5]) | python-statatics-tutorial/basic-theme/python-language/Function.ipynb | gaufung/Data_Analytics_Learning_Note | mit |
2 global 参数 | x = 5
def set_x(y):
x = y
print 'inner x is {}'.format(x)
set_x(10)
print 'global x is {}'.format(x) | python-statatics-tutorial/basic-theme/python-language/Function.ipynb | gaufung/Data_Analytics_Learning_Note | mit |
x = 5 表明为global变量，但是在set_x函数内部中，出现了x,但是其为局部变量,因此全局变量x并没有发生改变。 | def set_global_x(y):
global x
x = y
print 'global x is {}'.format(x)
set_global_x(10)
print 'global x now is {}'.format(x) | python-statatics-tutorial/basic-theme/python-language/Function.ipynb | gaufung/Data_Analytics_Learning_Note | mit |
通过添加global关键字，使得global变量x发生了改变。
3 Exercise
Fibonacci sequence
$F_{n+1}=F_{n}+F_{n-1}$ 其中 $F_{0}=0,F_{1}=1,F_{2}=1,F_{3}=2 \cdots$
递归版本
算法时间时间复杂度高达 $T(n)=n^2$ | def fib_recursive(n):
if n == 0 or n == 1:
return n
else:
return fib_recursive(n-1) + fib_recursive(n-2)
fib_recursive(10) | python-statatics-tutorial/basic-theme/python-language/Function.ipynb | gaufung/Data_Analytics_Learning_Note | mit |
迭代版本
算法时间复杂度为$T(n)=n$ | def fib_iterator(n):
g = 0
h = 1
i = 0
while i < n:
h = g + h
g = h - g
i += 1
return g
fib_iterator(10) | python-statatics-tutorial/basic-theme/python-language/Function.ipynb | gaufung/Data_Analytics_Learning_Note | mit |
迭代器版本
使用 yield 关键字可以实现迭代器 | def fib_iter(n):
g = 0
h = 1
i = 0
while i < n:
h = g + h
g = h -g
i += 1
yield g
for value in fib_iter(10):
print value, | python-statatics-tutorial/basic-theme/python-language/Function.ipynb | gaufung/Data_Analytics_Learning_Note | mit |
矩阵求解法
$$\begin{bmatrix}F_{n+1}\F_{n}\end{bmatrix}=\begin{bmatrix}1&1\1&0\end{bmatrix}\begin{bmatrix}F_{n}\F_{n-1}\end{bmatrix}$$
令$u_{n+1}=Au_{n}$ 其中 $u_{n+1}=\begin{bmatrix}F_{n+1}\F_{n}\end{bmatrix}$
通过矩阵的迭代求解
$u_{n+1}=A^{n}u_{0}$，其中 $u_{0}=\begin{bmatrix}1 \0 \end{bmatrix}$，对于$A^n$ 可以通过 $(A^{n/2})^{2}$ 方式求解，使得算法时间复杂度达到 $log(n)$ | import numpy as np
a = np.array([[1,1],[1,0]])
def pow_n(n):
if n == 1:
return a
elif n % 2 == 0:
half = pow_n(n/2)
return half.dot(half)
else:
half = pow_n((n-1)/2)
return a.dot(half).dot(half)
def fib_pow(n):
a_n = pow_n(n)
u_0 = np.array([1,0])
return a_n.dot(u_0)[1]
fib_pow(10) | python-statatics-tutorial/basic-theme/python-language/Function.ipynb | gaufung/Data_Analytics_Learning_Note | mit |
Quick Sort | def quick_sort(array):
if len(array) < 2:
return array
else:
pivot = array[0]
left = [item for item in array[1:] if item < pivot]
right = [item for item in array[1:] if item >= pivot]
return quick_sort(left)+[pivot]+quick_sort(right)
quick_sort([10,11,3,21,9,22]) | python-statatics-tutorial/basic-theme/python-language/Function.ipynb | gaufung/Data_Analytics_Learning_Note | mit |
Y yo para que quiero eso? De que sirve pandas?
Pandas te sirve si quieres:
Trabajar con datos de manera facil.
Explorar un conjunto de datos de manera rapida, enterder los datos que tienes.
Facilmente manipular informacion, por ejemplo sacar estadisticas.
Graficas patrones y distribuciones de datos.
Trabajar con Exceles, base de datos, sin tener que suar esas herramientas.
Y mucho mas...
El DataFrame en Pandas
Una estructura de datos en Pandas se llama un DataFrame, con el manejamos todos los datos y aplicamos tranformaciones.
Asi creamos un DataFrame vacio: | df.head() | Dia1/.ipynb_checkpoints/2_PandasIntro-checkpoint.ipynb | beangoben/HistoriaDatos_Higgs | gpl-2.0 |
No nos sirve nada vacio, entonces agreguemos le informacion!
LLenando informacion con un Dataframe
Situacion:
Suponte que eres un taquero y quieres hacer un dataframe de cuantos tacos vendes en una semana igual y para ver que tacos son mas populares y echarle mas ganas en ellos,
Asumiremos:
Que vende tacos de Pastor, Tripa y Chorizo
Hay 7 dias en una semana de Lunes a Domingo (obvio)
Crearemos el numero de tacos como numeros enteros aleatorios (np.random.randint)
Ojo! Si ponemos la variable de un dataframe al final de una celda no saldra una tabla con los datos, eah! | df['Pastor']=np.random.randint(100, size=7)
df['Tripas']=np.random.randint(100, size=7)
df['Chorizo']=np.random.randint(100, size=7)
df.index=['Lunes','Martes','Miercoles','Jueves','Viernes','Sabado','Domingo']
df. | Dia1/.ipynb_checkpoints/2_PandasIntro-checkpoint.ipynb | beangoben/HistoriaDatos_Higgs | gpl-2.0 |
Lesson 1
Create Data - We begin by creating our own data set for analysis. This prevents the end user reading this tutorial from having to download any files to replicate the results below. We will export this data set to a text file so that you can get some experience pulling data from a text file.
Get Data - We will learn how to read in the text file. The data consist of baby names and the number of baby names born in the year 1880.
Prepare Data - Here we will simply take a look at the data and make sure it is clean. By clean I mean we will take a look inside the contents of the text file and look for any anomalities. These can include missing data, inconsistencies in the data, or any other data that seems out of place. If any are found we will then have to make decisions on what to do with these records.
Analyze Data - We will simply find the most popular name in a specific year.
Present Data - Through tabular data and a graph, clearly show the end user what is the most popular name in a specific year.
The pandas library is used for all the data analysis excluding a small piece of the data presentation section. The matplotlib library will only be needed for the data presentation section. Importing the libraries is the first step we will take in the lesson. | # Import all libraries needed for the tutorial
# General syntax to import specific functions in a library:
##from (library) import (specific library function)
from pandas import DataFrame, read_csv
# General syntax to import a library but no functions:
##import (library) as (give the library a nickname/alias)
import matplotlib.pyplot as plt
import pandas as pd #this is how I usually import pandas
import sys #only needed to determine Python version number
# Enable inline plotting
%matplotlib inline
print 'Python version ' + sys.version
print 'Pandas version ' + pd.__version__ | notebooks/pandas_tutorial.ipynb | babraham123/script-runner | mit |
Single computer
get data (5) # usdgbp | demo.get_price('GBPUSD')
process.processSinglePrice()
demo.get_price('USDEUR')
process.processSinglePrice()
demo.get_price('EURGBP')
process.processSinglePrice()
demo.get_prices(1)
process.processPrices(3)
| demos/demo1/00_DEMO_01.ipynb | mhallett/MeDaReDa | mit |
limitations
Multi computer (cloud)
set workers to work | while True:
process.processSinglePrice()
#break | demos/demo1/00_DEMO_01.ipynb | mhallett/MeDaReDa | mit |
Trying reduce without and with an initializer. | from operator import add
for result in (reduce(add, [42]), reduce(add, [42], 10)):
print(result) | content/posts/coding/recursion_looping_relationship.ipynb | dm-wyncode/zipped-code | mit |
My rewrite of functools.reduce using recursion.
For the sake of demonstration only. | def first(value_list):
return value_list[0]
def rest(value_list):
return value_list[1:]
def is_undefined(value):
return value is None
def recursive_reduce(function, iterable, initializer=None):
if is_undefined(initializer):
initializer = accum_value = first(iterable)
else:
accum_value = function(initializer, first(iterable))
if len(iterable) == 1: # base case
return accum_value
return recursive_reduce(function, rest(iterable), accum_value) | content/posts/coding/recursion_looping_relationship.ipynb | dm-wyncode/zipped-code | mit |
Test.
Test if the two functions return the sum of a list of random numbers. | from random import choice
from operator import add
LINE = ''.join(('-', ) * 20)
print(LINE)
for _ in range(5):
# create a tuple of random numbers of length 2 to 10
test_values = tuple(choice(range(101)) for _ in range(choice(range(2, 11))))
print('Testing these values: {}'.format(test_values))
# use sum for canonical value
expected = sum(test_values)
print('The expected result: {}\n'.format(expected))
test_answers = ((f.__name__, f(add, test_values))
for f
in (reduce, recursive_reduce))
test_results = ((f_name, test_answer == expected, )
for f_name, test_answer in test_answers)
for f_name, answer in test_results:
try:
assert answer
print('`{}` passed: {}'.format(f_name, answer))
except AssertionError:
print('`{}` failed: {}'.format(f_name, not answer))
print(LINE)
from recursion_looping_relationship_meta import tweets | content/posts/coding/recursion_looping_relationship.ipynb | dm-wyncode/zipped-code | mit |
Then let us generate some points in 2-D that will form our dataset: | # Create some data points | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Let's visualise these points in a scatterplot using the plot function from matplotlib | # Visualise the points in a scatterplot | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Here, imagine that the purpose is to build a classifier that for a given new point will return whether it belongs to the crosses (class 1) or circles (class 0).
Learning Activity 2: Computing the output of a Perceptron
Let’s now define a function which returns the output of a Perceptron for a single input point. | # Now let's build a perceptron for our points
def outPerceptron(x,w,b):
innerProd = np.dot(x,w) # computes the weighted sum of input
output = 0
if innerProd > b:
output = 1
return output | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
It’s useful to define a function which returns the sequence of outputs of the Perceptron for a sequence
of input points: | # Define a function which returns the sequence of outputs for a sequence of input points
def multiOutPerceptron(X,w,b):
nInstances = X.shape[0]
outputs = np.zeros(nInstances)
for i in range(0,nInstances):
outputs[i] = outPerceptron(X[i,:],w,b)
return outputs | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Bonus Activity: Efficient coding of multiOutPerceptron
In the above implementation, the simple outPerceptron function is called for every single instance. It
is cleaner and more efficient to code everything in one function using matrices: | # Optimise the multiOutPerceptron function | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
In the above implementation, the simple outPerceptron function is called for every single instance. It is cleaner and more efficient to code everything in one function using matrices.
Learning Activity 4: Playing with weights and thresholds
Let’s try some weights and thresholds, and see what happens: | # Try some initial weights and thresholds | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
So this is clearly not great! it classifies the first point as in one category and all the others in the other one. Let's try something else (an educated guess this time). | # Try an "educated guess" | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
This is much better! To obtain these values, we found a separating hyperplane (here a line) between the points. The equation of the line is
y = 0.5x-0.2
Quiz
- Can you explain why this line corresponds to the weights and bias we used?
- Is this separating line unique? what does it mean?
Can you check that the perceptron will indeed classify any point above the red line as a 1 (cross) and every point below as a 0 (circle)?
Learning Activity 5: Illustration of the output of the Perceptron and the separating line | # Visualise the separating line | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Now try adding new points to see how they are classified: | # Add new points and test | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Visualise the new test points in the graph and plot the separating lines. | # Visualise the new points and line | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Note here that the two sets of parameters classify the squares identically but not the triangle. You can now ask yourself, which one of the two sets of parameters makes more sense? How would you classify that triangle? These type of points are frequent in realistic datasets and the question of how to classify them "accurately" is often very hard to answer...
Gradient Descent
Learning Activity 6: Coding a simple gradient descent
Definition of a function and it's gradient
$f(x) = \exp(-\sin(x))x^2$
$f'(x) = -x \exp(-\sin(x)) (x\cos(x)-2)$
It is convenient to define python functions which return the value of the function and its gradient at an arbitrary point $x$ | def function(x):
return np.exp(-np.sin(x))*(x**2)
def gradient(x):
return -x*np.exp(-np.sin(x))*(x*np.cos(x)-2) # use wolfram alpha! | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Let's see what the function looks like | # Visualise the function | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Now let us implement a simple Gradient Descent that uses constant stepsizes. We define two functions, the first one is the most simple version which doesn't store the intermediate steps that are taken. The second one does store the steps which is useful to visualize what is going on and explain some of the typical behaviour of GD. | def simpleGD(x0,stepsize,nsteps):
x = x0
for k in range(0,nsteps):
x -= stepsize*gradient(x)
return x
def simpleGD2(x0,stepsize,nsteps):
x = np.zeros(nsteps+1)
x[0] = x0
for k in range(0,nsteps):
x[k+1] = x[k]-stepsize*gradient(x[k])
return x | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Let's see what it looks like. Let's start from $x_0 = 3$, use a (constant) stepsize of $\delta=0.1$ and let's go for 100 steps. | # Try the first given values | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Simple inspection of the figure above shows that that is close enough to the actual true minimum ($x^\star=0$)
A few standard situations: | # Try the second given values | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Ok! so that's still alright | # Try the third given values | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
That's not... Visual inspection of the figure above shows that we got stuck in a local optimum.
Below we define a simple visualization function to show where the GD algorithm brings us. It can be overlooked. | def viz(x,a=-10,b=10):
xx = np.linspace(a,b,100)
yy = function(xx)
ygd = function(x)
plt.plot(xx,yy)
plt.plot(x,ygd,color='red')
plt.plot(x[0],ygd[0],marker='o',color='green',markersize=10)
plt.plot(x[len(x)-1],ygd[len(x)-1],marker='o',color='red',markersize=10)
plt.show()
| misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Let's show the steps that were taken in the various cases that we considered above | # Visualise the steps taken in the previous cases | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
To summarise these three cases:
- In the first case, we start from a sensible point (not far from the optimal value $x^\star = 0$ and on a slope that leads directly to it) and we get to a very satisfactory point.
- In the second case, we start from a less sensible point (on a slope that does not lead directly to it) and yet the algorithm still gets us to a very satisfactory point.
- In the third case, we also start from a bad location but this time the algorithm gets stuck in a local minima.
Attacking MNIST
Learning Activity 7: Loading the Python libraries
Import statements for KERAS library | from keras.datasets import mnist
from keras.models import Sequential
from keras.layers.core import Dense, Activation
from keras.optimizers import SGD, RMSprop
from keras.utils import np_utils
# Some generic parameters for the learning process
batch_size = 100 # number of instances each noisy gradient will be evaluated upon
nb_classes = 10 # 10 classes 0-1-...-9
nb_epoch = 10 # computational budget: 10 passes through the whole dataset | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Learning Activity 8: Loading the MNIST dataset
Keras does the loading of the data itself and shuffles the data randomly. This is useful since the difficulty
of the examples in the dataset is not uniform (the last examples are harder than the first ones) | # Load the MNIST data | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
You can also depict a sample from either the training or the test set using the imshow() function: | # Display the first image | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Ok the label 5 does indeed seem to correspond to that number!
Let's check the dimension of the dataset
Learning Activity 9: Reshaping the dataset
Each image in MNIST has 28 by 28 pixels, which results in a $28\times 28$ array. As a next step, and prior to feeding the data into our NN classifier, we needd to flatten each array into a $28\times 28$=784 dimensional vector. Each component of the vector holds an integer value between 0 (black) and 255 (white), which we need to normalise to the range 0 and 1. | # Reshaping of vectors in a format that works with the way the layers are coded | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Remember, it is always good practice to check the dimensionality of your train and test data using the shape command prior to constructing any classification model: | # Check the dimensionality of train and test | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
So we have 60,000 training samples, 10,000 test samples and the dimension of the samples (instances) are 28x28 arrays. We need to reshape these instances as vectors (of 784=28x28 components). For storage efficiency, the values of the components are stored as Uint8, we need to cast that as float32 so that Keras can deal with them. Finally we normalize the values to the 0-1 range.
The labels are stored as integer values from 0 to 9. We need to tell Keras that these form the output categories via the function to_categorical. | # Set y categorical | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Learning Activity 10: Building a NN classifier
A neural network model consists of artificial neurons arranged in a sequence of layers. Each layer receives a vector of inputs and converts these into some output. The interconnection pattern is "dense" meaning it is fully connected to the previous layer. Note that the first hidden layer needs to specify the size of the input which amounts to implicitly having an input layer. | # First, declare a model with a sequential architecture
# Then add a first layer with 500 nodes and 784 inputs (the pixels of the image)
# Define the activation function to use on the nodes of that first layer
# Second hidden layer with 300 nodes
# Output layer with 10 categories (+using softmax) | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Learning Activity 11: Training and testing of the model
Here we define a somewhat standard optimizer for NN. It is based on Stochastic Gradient Descent with some standard choice for the annealing. | # Definition of the optimizer. | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Finding the right arguments here is non trivial but the choice suggested here will work well. The only parameter we can explain here is the first one which can be understood as an initial scaling of the gradients.
At this stage, launch the learning (fit the model). The model.fit function takes all the necessary arguments and trains the model. We describe below what these arguments are:
the training set (points and labels)
global parameters for the learning (batch size and number of epochs)
whether or not we want to show output during the learning
the test set (points and labels) | # Fit the model | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Obviously we care far more about the results on the validation set since it is the data that the NN has not used for its training. Good results on the test set means the model is robust. | # Display the results, the accuracy (over the test set) should be in the 98% | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Bonus: Does it work? | def whatAmI(img):
score = model.predict(img,batch_size=1,verbose=0)
for s in range(0,10):
print ('Am I a ', s, '? -- score: ', np.around(score[0][s]*100,3))
index = 1004 # here use anything between 0 and 9999
test = np.reshape(images_train[index,],(1,784))
plt.imshow(np.reshape(test,(28,28)), cmap="gray")
whatAmI(test)
| misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
Does it work? (experimental Pt2) | from scipy import misc
test = misc.imread('data/ex7.jpg')
test = np.reshape(test,(1,784))
test = test.astype('float32')
test /= 255.
plt.imshow(np.reshape(test,(28,28)), cmap="gray")
whatAmI(test) | misc/machinelearningbootcamp/day2/neural_nets.ipynb | kinshuk4/MoocX | mit |
To keep the calculations below manageable we specify a single nside=64 healpixel in an arbitrary location of the DESI footprint. | healpixel = 26030
nside = 64 | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Specifying the random seed makes our calculations reproducible. | seed = 555
rand = np.random.RandomState(seed) | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Define a couple wrapper routines we will use below several times. | def plot_subset(wave, flux, truth, objtruth, nplot=16, ncol=4, these=None,
xlim=None, loc='right', targname='', objtype=''):
"""Plot a random sampling of spectra."""
nspec, npix = flux.shape
if nspec < nplot:
nplot = nspec
nrow = np.ceil(nplot / ncol).astype('int')
if loc == 'left':
xtxt, ytxt, ha = 0.05, 0.93, 'left'
else:
xtxt, ytxt, ha = 0.93, 0.93, 'right'
if these is None:
these = rand.choice(nspec, nplot, replace=False)
these = np.sort(these)
ww = (wave > 5500) * (wave < 5550)
fig, ax = plt.subplots(nrow, ncol, figsize=(2.5*ncol, 2*nrow), sharey=False, sharex=True)
for thisax, indx in zip(ax.flat, these):
thisax.plot(wave, flux[indx, :] / np.median(flux[indx, ww]))
if objtype == 'STAR' or objtype == 'WD':
thisax.text(xtxt, ytxt, r'$T_{{eff}}$={:.0f} K'.format(objtruth['TEFF'][indx]),
ha=ha, va='top', transform=thisax.transAxes, fontsize=13)
else:
thisax.text(xtxt, ytxt, 'z={:.3f}'.format(truth['TRUEZ'][indx]),
ha=ha, va='top', transform=thisax.transAxes, fontsize=13)
thisax.xaxis.set_major_locator(plt.MaxNLocator(3))
if xlim:
thisax.set_xlim(xlim)
for thisax in ax.flat:
thisax.yaxis.set_ticks([])
thisax.margins(0.2)
fig.suptitle(targname)
fig.subplots_adjust(wspace=0.05, hspace=0.05, top=0.93) | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Tracer QSOs
Both tracer and Lya QSO spectra contain an underlying QSO spectrum, but the Lya QSOs (which we demonstrate below) also include the Lya forest (here, based on the v2.0 of the "London" mocks).
Every target class has its own dedicated "Maker" class. | from desitarget.mock.mockmaker import QSOMaker
QSO = QSOMaker(seed=seed) | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
The various read methods return a dictionary with (hopefully self-explanatory) target- and mock-specific quantities.
Because most mock catalogs only come with (cosmologically accurate) 3D positions (RA, Dec, redshift), we use Gaussian mixture models trained on real data to assign other quantities like shapes, magnitudes, and colors, depending on the target class. For more details see the gmm-dr7.pynb Python notebook. | dir(QSOMaker)
data = QSO.read(healpixels=healpixel, nside=nside)
for key in sorted(list(data.keys())):
print('{:>20}'.format(key)) | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Now we can generate the spectra as well as the targeting catalogs (targets) and corresponding truth table. | %time flux, wave, targets, truth, objtruth = QSO.make_spectra(data)
print(flux.shape, wave.shape) | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
The truth catalog contains the target-type-agnostic, known properties of each object (including the noiseless photometry), while the objtruth catalog contains different information depending on the type of target. | truth
objtruth | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Next, let's run target selection, after which point the targets catalog should look just like an imaging targeting catalog (here, using the DR7 data model). | QSO.select_targets(targets, truth)
targets | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
And indeed, we can see that only a subset of the QSOs were identified as targets (the rest scattered out of the QSO color selection boxes). | from desitarget.targetmask import desi_mask
isqso = (targets['DESI_TARGET'] & desi_mask.QSO) != 0
print('Identified {} / {} QSO targets.'.format(np.count_nonzero(isqso), len(targets))) | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Finally, let's plot some example spectra. | plot_subset(wave, flux, truth, objtruth, targname='QSO') | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Generating QSO spectra with cosmological Lya skewers proceeds along similar lines.
Here, we also include BALs with 25% probability. | from desitarget.mock.mockmaker import LYAMaker
mockfile='/project/projectdirs/desi/mocks/lya_forest/london/v9.0/v9.0.0/master.fits'
LYA = LYAMaker(seed=seed, balprob=0.25)
lyadata = LYA.read(mockfile=mockfile,healpixels=healpixel, nside=nside)
%time lyaflux, lyawave, lyatargets, lyatruth, lyaobjtruth = LYA.make_spectra(lyadata)
lyaobjtruth
plot_subset(lyawave, lyaflux, lyatruth, lyaobjtruth, xlim=(3500, 5500), targname='LYA')
#Now lets generate the same spectra but including the different features and the new continum model.
#For this we need to reload the desitarget module, for some reason it seems not be enough with defining a diferen variable for the LYAMaker
del sys.modules['desitarget.mock.mockmaker']
from desitarget.mock.mockmaker import LYAMaker
LYA = LYAMaker(seed=seed,sqmodel='lya_simqso_model_develop',balprob=0.25)
lyadata_continum = LYA.read(mockfile=mockfile,healpixels=healpixel, nside=nside)
%time lyaflux_cont, lyawave_cont, lyatargets_cont, lyatruth_cont, lyaobjtruth_cont = LYA.make_spectra(lyadata_continum) | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Lets plot together some of the spectra with the old and new continum model | plt.figure(figsize=(20, 10))
indx=rand.choice(len(lyaflux),9)
for i in range(9):
plt.subplot(3, 3, i+1)
plt.plot(lyawave,lyaflux[indx[i]],label="Old Continum")
plt.plot(lyawave_cont,lyaflux_cont[indx[i]],label="New Continum")
plt.legend()
| doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
And finally we compare the colors, for the two runs with the new and old continum | plt.plot(lyatruth["FLUX_W1"],lyatruth_cont["FLUX_W1"]/lyatruth["FLUX_W1"]-1,'.')
plt.xlabel("FLUX_W1")
plt.ylabel(r"FLUX_W1$^{new}$/FLUX_W1-1")
plt.plot(lyatruth["FLUX_W2"],lyatruth_cont["FLUX_W2"]/lyatruth["FLUX_W2"]-1,'.')
plt.xlabel("FLUX_W2")
plt.ylabel(r"(FLUX_W2$^{new}$/FLUX_W2)-1")
plt.hist(lyatruth["FLUX_W1"],bins=100,label="Old Continum",alpha=0.7)
plt.hist(lyatruth_cont["FLUX_W1"],bins=100,label="New Continum",histtype='step',linestyle='--')
plt.xlim(0,100) #Limiting to 100 to see it better.
plt.xlabel("FLUX_W1")
plt.legend()
plt.hist(lyatruth["FLUX_W2"],bins=100,label="Old Continum",alpha=0.7)
plt.hist(lyatruth_cont["FLUX_W2"],bins=100,label="New Continum",histtype='step',linestyle='--')
plt.xlim(0,100) #Limiting to 100 to see it better.
plt.xlabel("FLUX_W2")
plt.legend() | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Conclusion: Colors are slightly affected by changing the continum model.
To Finalize the LYA section, lets generate another set of spectra now including DLAs, metals, LYB, etc. | del sys.modules['desitarget.mock.mockmaker']
from desitarget.mock.mockmaker import LYAMaker ##Done in order to reload the desitarget, it doesn't seem to be enough with initiating a diferent variable for the LYAMaker class.
LYA = LYAMaker(seed=seed,sqmodel='lya_simqso_model',balprob=0.25,add_dla=True,add_metals="all",add_lyb=True)
lyadata_all= LYA.read(mockfile=mockfile,healpixels=healpixel, nside=nside)
%time lyaflux_all, lyawave_all, lyatargets_all, lyatruth_all, lyaobjtruth_all = LYA.make_spectra(lyadata_all)
plot_subset(lyawave_all, lyaflux_all, lyatruth_all, lyaobjtruth_all, xlim=(3500, 5500), targname='LYA') | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
Demonstrate the other extragalactic target classes: LRG, ELG, and BGS.
For simplicity let's write a little wrapper script that does all the key steps. | def demo_mockmaker(Maker, seed=None, nrand=16, loc='right'):
TARGET = Maker(seed=seed)
log.info('Reading the mock catalog for {}s'.format(TARGET.objtype))
tdata = TARGET.read(healpixels=healpixel, nside=nside)
log.info('Generating {} random spectra.'.format(nrand))
indx = rand.choice(len(tdata['RA']), np.min( (nrand, len(tdata['RA'])) ) )
tflux, twave, ttargets, ttruth, tobjtruth = TARGET.make_spectra(tdata, indx=indx)
log.info('Selecting targets')
TARGET.select_targets(ttargets, ttruth)
plot_subset(twave, tflux, ttruth, tobjtruth, loc=loc,
targname=tdata['TARGET_NAME'], objtype=TARGET.objtype) | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
LRGs | from desitarget.mock.mockmaker import LRGMaker
%time demo_mockmaker(LRGMaker, seed=seed, loc='left') | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
ELGs | from desitarget.mock.mockmaker import ELGMaker
%time demo_mockmaker(ELGMaker, seed=seed, loc='left') | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |
BGS | from desitarget.mock.mockmaker import BGSMaker
%time demo_mockmaker(BGSMaker, seed=seed) | doc/nb/connecting-spectra-to-mocks.ipynb | desihub/desitarget | bsd-3-clause |