I am working with bidimensional arrays on Numpy for Extreme Learning Machines. One of my arrays, H, is random, and I want to compute its pseudoinverse.
If I use scipy.linalg.pinv2 everything runs smoothly. However, if I use scipy.linalg.pinv, sometimes (30-40% of the times) problems arise.
The reason why I am using pinv2 is because I read (here: http://vene.ro/blog/inverses-pseudoinverses-numerical-issues-speed-symmetry.html ) that pinv2 performs better on "tall" and on "wide" arrays.
The problem is that, if H has a column j of all 1, pinv(H) has huge coefficients at row j.
This is in turn a problem because, in such cases, np.dot(pinv(H), Y) contains some nan values (Y is an array of small integers).
Now, I am not into linear algebra and numeric computation enough to understand if this is a bug or some precision related property of the two functions. I would like you to answer this question so that, if it's the case, I can file a bug report (honestly, at the moment I would not even know what to write).
I saved the arrays with np.savetxt(fn, a, '%.2e', ';'): please, see https://dl.dropboxusercontent.com/u/48242012/example.tar.gz to find them.
Any help is appreciated. In the provided file, you can see in pinv(H).csv that rows 14, 33, 55, 56 and 99 have huge values, while in pinv2(H) the same rows have more decent values.
Your help is appreciated.
In short, the two functions implement two different ways to calculate the pseudoinverse matrix:
scipy.linalg.pinv uses least squares, which may be quite compute intensive and take up a lot of memory.
https://docs.scipy.org/doc/scipy/reference/generated/scipy.linalg.pinv.html#scipy.linalg.pinv
scipy.linalg.pinv2 uses SVD (singular value decomposition), which should run with a smaller memory footprint in most cases.
https://docs.scipy.org/doc/scipy/reference/generated/scipy.linalg.pinv2.html#scipy.linalg.pinv2
numpy.linalg.pinv also implements this method.
As these are two different evaluation methods, the resulting matrices will not be the same. Each method has its own advantages and disadvantages, and it is not always easy to determine which one should be used without deeply understanding the data and what the pseudoinverse will be used for. I'd simply suggest some trial-and-error and use the one which gives you the best results for your classifier.
Note that in some cases these functions cannot converge to a solution, and will then raise a scipy.stats.LinAlgError. In that case you may try to use the second pinv implementation, which will greatly reduce the amount of errors you receive.
Starting from scipy 1.7.0 , pinv2 is deprecated and also replaced by a SVD solution.
DeprecationWarning: scipy.linalg.pinv2 is deprecated since SciPy 1.7.0, use scipy.linalg.pinv instead
That means, numpy.pinv, scipy.pinv and scipy.pinv2 now compute all equivalent solutions. They are also equally fast in their computation, with scipy being slightly faster.
import numpy as np
import scipy
arr = np.random.rand(1000, 2000)
res1 = np.linalg.pinv(arr)
res2 = scipy.linalg.pinv(arr)
res3 = scipy.linalg.pinv2(arr)
np.testing.assert_array_almost_equal(res1, res2, decimal=10)
np.testing.assert_array_almost_equal(res1, res3, decimal=10)
Related
IEEE floating point operations are deterministic, but see How can floating point calculations be made deterministic? for one way that an overall floating point computation can be non-deterministic:
... parallel computations are non-deterministic in terms of the order in which floating-point computations are performed, which can result in non-bit-exact results across runs.
Two-part question:
How else can an overall floating point computation be non-deterministic, yielding results that are not exactly equal?
Consider a single-threaded Python program that calls NumPy, CVXOPT, and SciPy subroutines such as scipy.optimize.fsolve(), which in turn call native libraries like MINPACK and GLPK and optimized linear algebra subroutines like BLAS, ATLAS, and MKL. “If your numpy/scipy is compiled using one of these, then dot() will be computed in parallel (if this is faster) without you doing anything.”
Do these native libraries ever parallelize in a way that introduces non-deterministic results?
Assumptions:
The same software, with the same inputs, on the same hardware. The output of multiple runs should be equal.
If that works, it's highly desirable to test that the output after doing a code refactoring is equal. (Yes, some changes in order of operations can make some of the output not-equal.)
All random numbers in the program are psuedo-random numbers used in a consistent way from the same seeds across all runs.
No uninitialized values. Python is generally safe in that way but numpy.empty() returns a new array without initializing entries. And it's not clear that it's much faster in practice. So beware!
#PaulPanzer's test shows that numpy.empty() does return an uninitialized array and it can easily and quickly recycle a recent array:
import numpy as np
np.arange(100); np.empty(100, int); np.empty(100, int)
np.arange(100, 200.0); np.empty(100, float); np.empty(100, float)
It's tricky to get useful timing measurements for these routines! In a timeit loop, numpy.empty() can just keep reallocating the same one or two memory nodes. The time is independent of the array size. To prevent recycling:
from timeit import timeit
timeit('l.append(numpy.empty(100000))', 'import numpy; l = []')
timeit('l.append(numpy.zeros(100000))', 'import numpy; l = []')
but reducing that array size to numpy.zeros(10000) takes 15x as long; reducing it to numpy.zeros(1000) takes 1.3x as long (on my MBP). Puzzling.
See also:
Hash values are salted in Python 3 and each dict preserves insertion order. That could vary the order of operations from run to run. [I'm wrangling with this problem in Python 2.7.15.]
I found that most (not all) of the non-determinism problems I'm experiencing seem to be fixed in the code for OpenBLAS 0.3.5.
A bunch of threading problems in earlier versions of OpenBLAS are fixed in release 0.3.4, but that release has a macOS compatibility bug that's fixed in the code for release 0.3.5. The bugs also occurs with Apple's Accelerate framework version 1.1 and Intel's MKL mkl==2019.0.
See how to install OpenBLAS and compile NumPy and SciPy on it.
Perhaps the remaining problems I'm experiencing are due to other libraries linked to Accelerate?
Note: I'm still open to more answers to this question.
I use the scipy.optimize.minimize ( https://docs.scipy.org/doc/scipy/reference/tutorial/optimize.html ) function with method='L-BFGS-B.
An example of what it returns is here above:
fun: 32.372210618549758
hess_inv: <6x6 LbfgsInvHessProduct with dtype=float64>
jac: array([ -2.14583906e-04, 4.09272616e-04, -2.55795385e-05,
3.76587650e-05, 1.49213975e-04, -8.38440428e-05])
message: 'CONVERGENCE: REL_REDUCTION_OF_F_<=_FACTR*EPSMCH'
nfev: 420
nit: 51
status: 0
success: True
x: array([ 0.75739412, -0.0927572 , 0.11986434, 1.19911266, 0.27866406,
-0.03825225])
The x value correctly contains the fitted parameters. How do I compute the errors associated to those parameters?
TL;DR: You can actually place an upper bound on how precisely the minimization routine has found the optimal values of your parameters. See the snippet at the end of this answer that shows how to do it directly, without resorting to calling additional minimization routines.
The documentation for this method says
The iteration stops when (f^k - f^{k+1})/max{|f^k|,|f^{k+1}|,1} <= ftol.
Roughly speaking, the minimization stops when the value of the function f that you're minimizing is minimized to within ftol of the optimum. (This is a relative error if f is greater than 1, and absolute otherwise; for simplicity I'll assume it's an absolute error.) In more standard language, you'll probably think of your function f as a chi-squared value. So this roughly suggests that you would expect
Of course, just the fact that you're applying a minimization routine like this assumes that your function is well behaved, in the sense that it's reasonably smooth and the optimum being found is well approximated near the optimum by a quadratic function of the parameters xi:
where Δxi is the difference between the found value of parameter xi and its optimal value, and Hij is the Hessian matrix. A little (surprisingly nontrivial) linear algebra gets you to a pretty standard result for an estimate of the uncertainty in any quantity X that's a function of your parameters xi:
which lets us write
That's the most useful formula in general, but for the specific question here, we just have X = xi, so this simplifies to
Finally, to be totally explicit, let's say you've stored the optimization result in a variable called res. The inverse Hessian is available as res.hess_inv, which is a function that takes a vector and returns the product of the inverse Hessian with that vector. So, for example, we can display the optimized parameters along with the uncertainty estimates with a snippet like this:
ftol = 2.220446049250313e-09
tmp_i = np.zeros(len(res.x))
for i in range(len(res.x)):
tmp_i[i] = 1.0
hess_inv_i = res.hess_inv(tmp_i)[i]
uncertainty_i = np.sqrt(max(1, abs(res.fun)) * ftol * hess_inv_i)
tmp_i[i] = 0.0
print('x^{0} = {1:12.4e} ± {2:.1e}'.format(i, res.x[i], uncertainty_i))
Note that I've incorporated the max behavior from the documentation, assuming that f^k and f^{k+1} are basically just the same as the final output value, res.fun, which really ought to be a good approximation. Also, for small problems, you can just use np.diag(res.hess_inv.todense()) to get the full inverse and extract the diagonal all at once. But for large numbers of variables, I've found that to be a much slower option. Finally, I've added the default value of ftol, but if you change it in an argument to minimize, you would obviously need to change it here.
One approach to this common problem is to use scipy.optimize.leastsq after using minimize with 'L-BFGS-B' starting from the solution found with 'L-BFGS-B'. That is, leastsq will (normally) include and estimate of the 1-sigma errors as well as the solution.
Of course, that approach makes several assumption, including that leastsq can be used and may be appropriate for solving the problem. From a practical view, this requires the objective function return an array of residual values with at least as many elements as variables, not a cost function.
You may find lmfit (https://lmfit.github.io/lmfit-py/) useful here: It supports both 'L-BFGS-B' and 'leastsq' and gives a uniform wrapper around these and other minimization methods, so that you can use the same objective function for both methods (and specify how to convert the residual array into the cost function). In addition, parameter bounds can be used for both methods. This makes it very easy to first do a fit with 'L-BFGS-B' and then with 'leastsq', using the values from 'L-BFGS-B' as starting values.
Lmfit also provides methods to more explicitly explore confidence limits on parameter values in more detail, in case you suspect the simple but fast approach used by leastsq might be insufficient.
It really depends what you mean by "errors". There is no general answer to your question, because it depends on what you're fitting and what assumptions you're making.
The easiest case is one of the most common: when the function you are minimizing is a negative log-likelihood. In that case the inverse of the hessian matrix returned by the fit (hess_inv) is the covariance matrix describing the Gaussian approximation to the maximum likelihood.The parameter errors are the square root of the diagonal elements of the covariance matrix.
Beware that if you are fitting a different kind of function or are making different assumptions, then that doesn't apply.
Quite simply, what I want to do is the following
A = np.ones((3,3)) #arbitrary matrix
B = np.ones((2,2)) #arbitrary matrix
A[1:,1:] = A[1:,1:] + B
except in Tensorflow (where the matrices can be arbitrarily complicated tensor expressions). Neither A nor B is a Tensorflow Variable, but just a run-of-the-mill tensor.
What I have gathered so far: tensors are immutable, so I cannot assign to a submatrix. tf.scatter_nd is the current option for sub-assignment, but does not appear to support sub-matrices, only slices.
Methods that should work, but are perhaps not ideal:
I could pad B with zeros, but I'm sure this leads to instantiation of
an unnecessarily large B - can it be made sparse, maybe?
I could use the padding idea, but write it as a low-rank decomposition, e.g. in Numpy: A+U.dot(B).U.T where U is a stacked zero and identity matrix. I'm not sure this is actually advantageous.
I could split A into submatrices, and stack them back together. Might be the most efficient, but sounds like the code would be convoluted.
Ideally, I want to do this operation N times for progressively smaller matrices, resulting in one large final result, but this is tangential.
I'll use one of the hacks for now, but I'm hoping someone can tell me what the idiomatic version is!
Minimally, I would like to know how to achieve what is stated in the title. Specifically, signal.lfilter seems like the only implementation of a difference equation filter in scipy, but it is 1D, as shown in the docs. I would like to know how to implement a 2D version as described by this difference equation. If that's as simple as "bro, use this function," please let me know, pardon my naiveté, and feel free to disregard the rest of the post.
I am new to DSP and acknowledging there might be a different approach to answering my question so I will explain the broader goal and give context for the question in the hopes someone knows how do want I want with Scipy, or perhaps a better way than what I explicitly asked for.
To get straight into it, broadly speaking I am using vectorized computation methods (Numpy/Scipy) to implement a Monte Carlo simulation to improve upon a naive for loop. I have successfully abstracted most of my operations to array computation / linear algebra, but a few specific ones (recursive computations) have eluded my intuition and I continually end up in the digital signal processing world when I go looking for how this type of thing has been done by others (that or machine learning but those "frameworks" are much opinionated). The reason most of my google searches end up on scipy.signal or scipy.ndimage library references is clear to me at this point, and subsequent to accepting the "signal" representation of my data, I have spent a considerable amount of time (about as much as reasonable for a field that is not my own) ramping up the learning curve to try and figure out what I need from these libraries.
My simulation entails updating a vector of data representing the state of a system each period for n periods, and then repeating that whole process a "Monte Carlo" amount of times. The updates in each of n periods are inherently recursive as the next depends on the state of the prior. It can be characterized as a difference equation as linked above. Additionally this vector is theoretically indexed on an grid of points with uneven stepsize. Here is an example vector y and its theoretical grid t:
y = np.r_[0.0024, 0.004, 0.0058, 0.0083, 0.0099, 0.0133, 0.0164]
t = np.r_[0.25, 0.5, 1, 2, 5, 10, 20]
I need to iteratively perform numerous operations to y for each of n "updates." Specifically, I am computing the curvature along the curve y(t) using finite difference approximations and using the result at each point to adjust the corresponding y(t) prior to the next update. In a loop this amounts to inplace variable reassignment with the desired update in each iteration.
y += some_function(y)
Not only does this seem inefficient, but vectorizing things seems intuitive given y is a vector to begin with. Furthermore I am interested in preserving each "updated" y(t) along the n updates, which would require a data structure of dimensions len(y) x n. At this point, why not perform the updates inplace in the array? This is wherein lies the question. Many of the update operations I have succesfully vectorized the "Numpy way" (such as adding random variates to each point), but some appear overly complex in the array world.
Specifically, as mentioned above the one involving computing curvature at each element using its neighbouring two elements, and then imediately using that result to update the next row of the array before performing its own curvature "update." I was able to implement a non-recursive version (each row fails to consider its "updated self" from the prior row) of the curvature operation using ndimage generic_filter. Given the uneven grid, I have unique coefficients (kernel weights) for each triplet in the kernel footprint (instead of always using [1,-2,1] for y'' if I had a uniform grid). This last part has already forced me to use a spatial filter from ndimage rather than a 1d convolution. I'll point out, something conceptually similar was discussed in this math.exchange post, and it seems to me only the third response saliently addressed the difference between mathematical notion of "convolution" which should be associative from general spatial filtering kernels that would require two sequential filtering operations or a cleverly merged kernel.
In any case this does not seem to actually address my concern as it is not about 2D recursion filtering but rather having a backwards looking kernel footprint. Additionally, I think I've concluded it is not applicable in that this only allows for "recursion" (backward looking kernel footprints in the spatial filtering world) in a manner directly proportional to the size of the recursion. Meaning if I wanted to filter each of n rows incorporating calculations on all prior rows, it would require a convolution kernel far too big (for my n anyways). If I'm understanding all this correctly, a recursive linear filter is algorithmically more efficient in that it returns (for use in computation) the result of itself applied over the previous n samples (up to a level where the stability of the algorithm is affected) using another companion vector (z). In my case, I would only need to look back one step at output signal y[n-1] to compute y[n] from curvature at x[n] as the rest works itself out like a cumsum. signal.lfilter works for this, but I can't used that to compute curvature, as that requires a kernel footprint that can "see" at least its left and right neighbors (pixels), which is how I ended up using generic_filter.
It seems to me I should be able to do both simultaneously with one filter namely spatial and recursive filtering; or somehow I've missed the maths of how this could be mathematically simplified/combined (convolution of multiples kernels?).
It seems like this should be a common problem, but perhaps it is rarely relevant to do both at once in signal processing and image filtering. Perhaps this is why you don't use signals libraries solely to implement a fast monte carlo simulation; though it seems less esoteric than using a tensor math library to implement a recursive neural network scan ... which I'm attempting to do right now.
EDIT: For those familiar with the theoretical side of DSP, I know that what I am describing, the process of designing a recursive filters with arbitrary impulse responses, is achieved by employing a mathematical technique called the z-transform which I understand is generally used for two things:
converting between the recursion coefficients and the frequency response
combining cascaded and parallel stages into a single filter
Both are exactly what I am trying to accomplish.
Also, reworded title away from FIR / IIR because those imply specific definitions of "recursion" and may be confusing / misnomer.
Do we have a GPU accelerated of version of numpy.max(X, axis=None) in Theano.
I looked into the documentation and found theano.tensor.max(X, axis=None), but it is 4-5 times slower than the numpy implementation.
I can assure you, it is not slow because of some bad choice of matrix size. Same matrix under theano.tensor.exp is 40 times faster than its numpy counterpart.
Any suggestions?
The previous answer is partial. The suggestion should not work, as the work around is the one used in the final compiled code. There is optimization that will do this transformation automatically.
The title of the question isn't the same as the content. They differ by the axis argument. I'll answer both questions.
If the axis is 0 or None we support this on the GPU for that operation for matrix. If the axis is None, we have a basic implementation that isn't well optimized as it is harder to parallelize. If the axis is 0, we have a basic implementation, but it is faster as it is easier to parallelize.
Also, how did you do your timing? If you just make one function with only that operation and test it via the device=gpu flags to do your comparison, this will include the transfer time between CPU and GPU. This is a memory bound operation, so if you include the transfer in your timming, personnaly I don't expect any speed op for that case. To see only the GPU operation, use Theano profiler: run with the Theano flag profile=True.
The max and exp operations are fundamentally different; exp (and other operations like addition, sin, etc.) is an elementwise operation that is embarrassingly parallelizable, while max requires a parallel-processing scan algorithm that basically builds up a tree of pairwise comparisons over an array. It's not impossible to speed up max, but it's not as easy as exp.
Anyway, the theano implementation of max basically consists of the following lines (in theano/tensor/basic.py):
try:
out = max_and_argmax(x, axis)[0]
except Exception:
out = CAReduce(scal.maximum, axis)(x)
where max_and_argmax is a bunch of custom code that, to my eye, implements a max+argmax operation using numpy, and CAReduce is a generic GPU-accelerated scan operation used as a fallback (which, according to the comments, doesn't support grad etc.). You could try using the fallback directly and see whether that is faster, maybe something like this:
from theano.tensor.elemwise import CAReduce
from theano.scalar import maximum
def mymax(X, axis=None):
CAReduce(maximum, axis)(X)