Related
Would this be a valid implementation of a cross entropy loss that takes the ordinal structure of the GT y into consideration? y_hat is the prediction from a neural network.
ce_loss = F.cross_entropy(y_hat, y, reduction="none")
distance_weight = torch.abs(y_hat.argmax(1) - y) + 1
ordinal_ce_loss = torch.mean(distance_weight * ce_loss)
I'll attempt to answer this question by first fully defining the task, since the question is a bit sparse on details.
I have a set of ordinal classes (e.g. first, second, third, fourth,
etc.) and I would like to predict the class of each data example from
among this set. I would like to define an entropy-based loss-function
for this problem. I would like this loss function to weight the loss
between a predicted class torch.argmax(y_hat) and the true class y
according to the ordinal distance between the two classes. Does the
given loss expression accomplish this?
Short answer: sure, it is "valid". You've roughly implemented L1-norm ordinal class weighting. I'd question whether this is truly the correct weighting strategy for this problem.
For instance, consider that for a true label n, the bin n response is weighted by 1, but the bin n+1 and n-1 responses are weighted by 2. This means that a lot more emphasis will be placed on NOT predicting false positives than on correctly predicting true positives, which may imbue your model with some strange bias.
It also means that examples on the edge will result in a larger total sum of weights, meaning that you'll be weighting examples where the true label is say "first" or "last" more highly than the intermediate classes. (Say you have 5 classes: 1,2,3,4,5. A true label of 1 will require distance_weight of [1,2,3,4,5], the sum of which is 15. A true label of 3 will require distance_weight of [3,2,1,2,3], the sum of which is 11.
In general, classification problems and entropy-based losses are underpinned by the assumption that no set of classes or categories is any more or less related than any other set of classes. In essence, the input data is embedded into an orthogonal feature space where each class represents one vector in the basis. This is quite plainly a bad assumption in your case, meaning that this embedding space is probably not particularly elegant: thus, you have to correct for it with sort of a hack-y weight fix. And in general, this assumption of class non-correlation is probably not true in a great many classification problems (consider e.g. the classic ImageNet classification problem, wherein the class pairs [bus,car], and [bus,zebra] are treated as equally dissimilar. But this is probably a digression into the inherent lack of usefulness of strict ontological structuring of information which is outside the scope of this answer...)
Long Answer: I'd highly suggest moving into a space where the ordinal value you care about is instead expressed in a continuous space. (In the first, second, third example, you might for instance output a continuous value over the range [1,max_place]. This allows you to benefit from loss functions that already capture well the notion that predictions closer in an ordered space are better than predictions farther away in an ordered space (e.g. MSE, Smooth-L1, etc.)
Let's consider one more time the case of the [first,second,third,etc.] ordinal class example, and say that we are trying to predict the places of a set of runners in a race. Consider two races, one in which the first place runner wins by 30% relative to the second place runner, and the second in which the first place runner wins by only 1%. This nuance is entirely discarded by the ordinal discrete classification. In essence, the selection of an ordinal set of classes truncates the amount of information conveyed in the prediction, which means not only that the final prediction is less useful, but also that the loss function encodes this strange truncation and binarization, which is then reflected (perhaps harmfully) in the learned model. This problem could likely be much more elegantly solved by regressing the finishing position, or perhaps instead by regressing the finishing time, of each athlete, and then performing the final ordinal classification into places OUTSIDE of the network training.
In conclusion, you might expect a well-trained ordinal classifier to produce essentially a normal distribution of responses across the class bins, with the distribution peak on the true value: a binned discretization of a space that almost certainly could, and likely should, be treated as a continuous space.
I have an environment in open AI gym, where the observation space is like [12,12.5,16.7], one value is discrete and the other two are continues, how can I define this in Gym?
I have tried to use the multi-discrete and discrete but it doesn't cover the continuous space, and I also tried box but the first integer one was problematic.
In reinforcement learning you usually want to normalize the observations in the range 0-1 (especially if you are using neural networks as function approximators). Therefore, it makes sense to use Boxes in the range 0-1.
How should stereo (2 channel) audio data be represented for FFT? Do you
A. Take the average of the two channels and assign it to the real component of a number and leave the imaginary component 0.
B. Assign one channel to the real component and the other channel to the imag component.
Is there a reason to do one or the other? I searched the web but could not find any definite answers on this.
I'm doing some simple spectrum analysis and, not knowing any better, used option A). This gave me an unexpected result, whereas option B) went as expected. Here are some more details:
I have a WAV file of a piano "middle-C". By definition, middle-C is 260Hz, so I would expect the peak frequency to be at 260Hz and smaller peaks at harmonics. I confirmed this by viewing the spectrum via an audio editing software (Sound Forge). But when I took the FFT myself, with option A), the peak was at 520Hz. With option B), the peak was at 260Hz.
Am I missing something? The explanation that I came up with so far is that representing stereo data using a real and imag component implies that the two channels are independent, which, I suppose they're not, and hence the mess-up.
I don't think you're taking the average correctly. :-)
C. Process each channel separately, assigning the amplitude to the real component and leaving the imaginary component as 0.
Option B does not make sense. Option A, which amounts to convert the signal to mono, is OK (if you are interested in a global spectrum).
Your problem (double freq) is surely related to some misunderstanding in the use of your FFT routines.
Once you take the FFT you need to get the Magnitude of the complex frequency spectrum. To get the magnitude you take the absolute of the complex spectrum |X(w)|. If you want to look at the power spectrum you square the magnitude spectrum, |X(w)|^2.
In terms of your frequency shift I think it has to do with you setting the imaginary parts to zero.
If you imagine the complex Frequency spectrum as a series of complex vectors or position vectors in a cartesian space. If you took one discrete frequency bin X(w), there would be one real component representing its direction in the real axis (x -direction), and one imaginary component in the in the imaginary axis (y - direction). There are four important values about this discrete frequency, 1. real value, 2. imaginary value, 3. Magnitude and, 4. phase. If you just take the real value and set imaginary to 0, you are setting Magnitude = real and phase = 0deg or 90deg. You have hence forth modified the resulting spectrum, and applied a bias to every frequency bin. Take a look at the wiki on Magnitude of a vector, also called the Euclidean norm of a vector to brush up on your understanding. Leonbloy was correct, but I hope this was more informative.
Think of the FFT as a way to get information from a single signal. What you are asking is what is the best way to display data from two signals. My answer would be to treat each independently, and display an FFT for each.
If you want a really fast streaming FFT you can read about an algorithm I wrote here: www.depthcharged.us/?p=176
I'd like to write a program that lets users draw points, lines, and circles as though with a straightedge and compass. Then I want to be able to answer the question, "are these three points collinear?" To answer correctly, I need to avoid rounding error when calculating the points.
Is this possible? How can I represent the points in memory?
(I looked into some unusual numeric libraries, but I didn't find anything that claimed to offer both exact arithmetic and exact comparisons that are guaranteed to terminate.)
Yes.
I highly recommend Introduction to constructions, which is a good basic guide.
Basically you need to be able to compute with constructible numbers - numbers that are either rational, or of the form a + b sqrt(c) where a,b,c were previously created (see page 6 on that PDF). This could be done with algebraic data type (e.g. data C = Rational Integer Integer | Root C C C in Haskell, where Root a b c = a + b sqrt(c)). However, I don't know how to perform tests with that representation.
Two possible approaches are:
Constructible numbers are a subset of algebraic numbers, so you can use algebraic numbers.
All algebraic numbers can be represented using polynomials of whose they are roots. The operations are computable, so if you represent a number a with polynomial p and b with polynomial q (p(a) = q(b) = 0), then it is possible to find a polynomial r such that r(a+b) = 0. This is done in some CASes like Mathematica, example. See also: Computional algebraic number theory - chapter 4
Use Tarski's test and represent numbers. It is slow (doubly exponential or so), but works :) Example: to represent sqrt(2), use the formula x^2 - 2 && x > 0. You can write equations for lines there, check if points are colinear etc. See A suite of logic programs, including Tarski's test
If you turn to computable numbers, then equality, colinearity etc. get undecidable.
I think the only way this would be possible is if you used a symbolic representation,
as opposed to trying to represent coordinate values directly -- so you would have
to avoid trying to coerce values like sqrt(2) into some numerical format. You will
be dealing with irrational numbers that are not finitely representable in binary,
decimal, or any other positional notation.
To expand on Jim Lewis's answer slightly, if you want to operate on points that are constructible from the integers with exact arithmetic, you will need to be able to operate on representations of the form:
a + b sqrt(c)
where a, b, and c are either rational numbers, or representations in the form given above. Wikipedia has a pretty decent article on the subject of what points are constructible.
Answering the question of exact equality (as necessary to establish colinearity) with such representations is a rather tricky problem.
If you try to compare co-ordinates for your points, then you have a problem. Leaving aside co-linearity for a moment, how about just working out whether two points are the same or not?
Supposing that one has given co-ordinates, and the other is a compass-straightedge construction starting from certain other co-ordinates, you want to determine with certainty whether they're the same point or not. Either way is a theorem of Euclidean geometry, it's not something you can just measure. You can prove they aren't the same by spotting some difference in their co-ordinates (for example by computing decimal places of each until you encounter a difference). But in general to prove they are the same cannot be done by approximate methods. Compute as many decimal places as you like of some expansions of 1/sqrt(2) and sqrt(2)/2, and you can prove they're very close together but you won't ever prove they're equal. That takes algebra (or geometry).
Similarly, to show that three points are co-linear you will need theorem-proving software. Represent the points A, B, C by their constructions, and attempt to prove the theorem "A, B and C are colinear". This is very hard - your program will prove some theorems but not others. Much easier is to ask the user for a proof that they are co-linear, and then verify (or refute) that proof, but that's probably not what you want.
In general, constructable points may have an arbitrarily complex symbolic form, so you must use a symbolic representation to work them exactly. As Stephen Canon noted above, you often need numbers of the form a+b*sqrt(c), where a and b are rational and c is an integer. All numbers of this form form a closed set under arithmetic operations. I have written some C++ classes (see rational_radical1.h) to work with these numbers if that is all you need.
It is also possible to construct numbers which are sums of any number of terms of rational multiples of radicals. When dealing with more than a single radicand, the numbers are no longer closed under multiplication and division, so you will need to store them as variable length rational coefficient arrays. The time complexity of operations will then be quadratic in the number of terms.
To go even further, you can construct the square root of any given number, so you could potentially have nested square roots. Here, the representations must be tree-like structures to deal with root hierarchy. While difficult to implement, there is nothing in principle preventing you from working with these representations. I'm not sure just what additional numbers can be constructed, but beyond a certain point, your symbolic representation will be expressive enough to handle very large classes of numbers.
Addendum
Found this Google Books link.
If the grid axes are integer valued then the answer is fairly straight forward, the points are either exactly colinear or they are not.
Typically however, one works with real numbers (well, floating points) and then draws the rounded values on the screen which does exist in integer space. In this case you have no choice but to pick a tolerance and use it to determine colinearity. Keep it small and the users will never know the difference.
You seem to be asking, in effect, "Can the normal mathematics (integer or floating point) used by computers be made to represent real numbers perfectly, with no rounding errors?" And, of course, the answer to that is "No." If you want theoretical correctness, then you will be stuck with the much harder problem of symbolic manipulation and coding up the equivalent of the inferences that are done in geometry. (In short, I'm agreeing with Steve Jessop, above.)
Some thoughts in the hope that they might help.
The sort of constructions you're talking about will require multiplication and division, which means that to preserve exactness you'll have to use rational numbers, which are generally easy to implement on top of a suitable sort of big integer (i.e., of unbounded magnitude). (Common Lisp has these built-in, and there have to be other languages.)
Now, you need to represent square roots of arbitrary numbers, and these have to be mixed in.
Therefore, a number is one of: a rational number, a rational number multiplied by a square root of a rational number (or, alternately, just the square root of a rational), or a sum of numbers. In order to prove anything, you're going to have to get these numbers into some sort of canonical form, which for all I can figure offhand may be annoying and computationally expensive.
This of course means that the users will be restricted to rational points and cannot use arbitrary rotations, but that's probably not important.
I would recommend no to try to make it perfectly exact.
The first reason for this is what you are asking here, the rounding error and all that stuff that comes with floating point calculations.
The second one is that you have to round your input as the mouse and screen work with integers. So, initially all user input would be integers, and your output would be integers.
Beside, from a usability point of view, its easier to click in the neighborhood of another point (in a line for example) and that the interface consider you are clicking in the point itself.
I want to invert a 4x4 matrix. My numbers are stored in fixed-point format (1.15.16 to be exact).
With floating-point arithmetic I usually just build the adjoint matrix and divide by the determinant (e.g. brute force the solution). That worked for me so far, but when dealing with fixed point numbers I get an unacceptable precision loss due to all of the multiplications used.
Note: In fixed point arithmetic I always throw away some of the least significant bits of immediate results.
So - What's the most numerical stable way to invert a matrix? I don't mind much about the performance, but simply going to floating-point would be to slow on my target architecture.
Meta-answer: Is it really a general 4x4 matrix? If your matrix has a special form, then there are direct formulas for inverting that would be fast and keep your operation count down.
For example, if it's a standard homogenous coordinate transform from graphics, like:
[ux vx wx tx]
[uy vy wy ty]
[uz vz wz tz]
[ 0 0 0 1]
(assuming a composition of rotation, scale, translation matrices)
then there's an easily-derivable direct formula, which is
[ux uy uz -dot(u,t)]
[vx vy vz -dot(v,t)]
[wx wy wz -dot(w,t)]
[ 0 0 0 1 ]
(ASCII matrices stolen from the linked page.)
You probably can't beat that for loss of precision in fixed point.
If your matrix comes from some domain where you know it has more structure, then there's likely to be an easy answer.
I think the answer to this depends on the exact form of the matrix. A standard decomposition method (LU, QR, Cholesky etc.) with pivoting (an essential) is fairly good on fixed point, especially for a small 4x4 matrix. See the book 'Numerical Recipes' by Press et al. for a description of these methods.
This paper gives some useful algorithms, but is behind a paywall unfortunately. They recommend a (pivoted) Cholesky decomposition with some additional features too complicated to list here.
I'd like to second the question Jason S raised: are you certain that you need to invert your matrix? This is almost never necessary. Not only that, it is often a bad idea. If you need to solve Ax = b, it is more numerically stable to solve the system directly than to multiply b by A inverse.
Even if you have to solve Ax = b over and over for many values of b, it's still not a good idea to invert A. You can factor A (say LU factorization or Cholesky factorization) and save the factors so you're not redoing that work every time, but you'd still solve the system each time using the factorization.
You might consider doubling to 1.31 before doing your normal algorithm. It'll double the number of multiplications, but you're doing a matrix invert and anything you do is going to be pretty tied to the multiplier in your processor.
For anyone interested in finding the equations for a 4x4 invert, you can use a symbolic math package to resolve them for you. The TI-89 will do it even, although it'll take several minutes.
If you give us an idea of what the matrix invert does for you, and how it fits in with the rest of your processing we might be able to suggest alternatives.
-Adam
Let me ask a different question: do you definitely need to invert the matrix (call it M), or do you need to use the matrix inverse to solve other equations? (e.g. Mx = b for known M, b) Often there are other ways to do this w/o explicitly needing to calculate the inverse. Or if the matrix M is a function of time & it changes slowly then you could calculate the full inverse once, & there are iterative ways to update it.
If the matrix represents an affine transformation (many times this is the case with 4x4 matrices so long as you don't introduce a scaling component) the inverse is simply the transpose of the upper 3x3 rotation part with the last column negated. Obviously if you require a generalized solution then looking into Gaussian elimination is probably the easiest.