{{ message }}
/ keops Public
Switch branches/tags
Nothing to show

Cannot retrieve contributors at this time
262 lines (195 sloc) 18.2 KB

# Formulas and syntax

KeOps lets us define any reduction operation of the form

\alpha_i = \operatorname{Reduction}_j\limits \big[ F(x^0_{\iota_0}, ... , x^{n-1}_{\iota_{n-1}})  \big]


or

\beta_j = \operatorname{Reduction}_i\limits \big[ F(x^0_{\iota_0}, ... , x^{n-1}_{\iota_{n-1}})  \big]


where F is a symbolic formula, the x^k_{\iota_k}'s are vector variables and \text{Reduction} is a Sum, LogSumExp or any other standard operation (see :ref:part.reduction for the full list of supported reductions).

We now describe the symbolic syntax that can be used through all KeOps bindings.

## Variables: category, index and dimension

At a low level, every variable x^k_{\iota_k} is specified by its category \iota_k\in\{i,j,\emptyset\} (meaning that the variable is indexed by i, by j, or is a fixed parameter across indices), its positional index k and its dimension d_k.

In practice, the category \iota_k is given through a keyword

keyword meaning
Vi variable indexed by i
Vj variable indexed by j
Pm parameter

followed by a (k,d_k) or (index,dimension) pair of integers. For instance, Vi(2,4) specifies a variable indexed by i, given as the third (k=2) input in the function call, and representing a vector of dimension d_k=4.

N.B.: Using the same index k for two variables with different dimensions or categories is not allowed and will be rejected by the compiler.

## Reserved words

keyword meaning
Ind indexes sequences

followed by a sequence (i_0, i_1, \cdots) of integers. For instance, Ind(2,4,2,5,12) can be used as parameters for some operations.

## Math operators

To define formulas with KeOps, you can use simple arithmetics:

 f * g scalar-vector multiplication (if f is scalar) or vector-vector element-wise multiplication f + g addition of two vectors f - g difference between two vectors or minus sign f / g element-wise division (N.B. f can be scalar, in fact f / g is the same as f * Inv(g)) (f | g) scalar product between vectors

Elementary functions:

 Minus(f) element-wise opposite of f Inv(f) element-wise inverse 1 ./ f Exp(f) element-wise exponential function Log(f) element-wise natural logarithm XLogX(f) computes f * log(f) element-wise (with value 0 at 0) Sin(f) element-wise sine function SinXDivX(f) function sin(f)/f element-wise (with value 1 at 0) Asin(f) element-wise arc-sine function Cos(f) element-wise cosine function Acos(f) element-wise arc-cosine function Atan(f) element-wise arc-tangent function Atan2(f,g) element-wise 2-argument arc-tangent function Pow(f, P) P-th power of f (element-wise), where P is a fixed integer Powf(f, g) power operation, alias for Exp(g*Log(f)) Square(f) element-wise square, faster than Pow(f,2) Sqrt(f) element-wise square root, faster than Powf(f,.5) Rsqrt(f) element-wise inverse square root, faster than Powf(f,-.5) Abs(f) element-wise absolute value Sign(f) element-wise sign function (-1 if f<0, 0 if f=0, 1 if f>0) Step(f) element-wise step function (0 if f<0, 1 if f>=0) ReLU(f) element-wise ReLU function (0 if f<0, f if f>=0) Clamp(f,a,b) element-wise Clamp function (a if f=0, h if f<0) Mod(f,m,off) element-wise Modulo function, with offset (computes f - m * floor((f - off)/m)) Round(f,d) element-wise Round function, with d decimal rounding

Simple vector operations:

 SqNorm2(f) squared L2 norm, same as (f|f) Norm2(f) L2 norm, same as Sqrt((f|f)) Normalize(f) normalize vector, same as Rsqrt(SqNorm2(f)) * f SqDist(f, g) squared L2 distance, same as SqNorm2(f - g)

Generic squared Euclidean norms, with support for scalar, diagonal and full (symmetric) matrices. If f is a vector of size N, depending on the size of s, WeightedSqNorm(s,f) may refer to:

• a weighted L2 norm s\cdot\sum_{0\leqslant i < N} f[i]^2 if s is a vector of size 1.
• a separable norm \sum_{0\leqslant i < N} s[i]\cdot f[i]^2 if s is a vector of size N.
• a full anisotropic norm \sum_{0\leqslant i,j < N} s[iN+j] f[i] f[j] if s is a vector of size N * N such that s[i*N+j]=s[j*N+i] (i.e. stores a symmetric matrix).
 WeightedSqNorm(s, f) generic squared euclidean norm WeightedSqDist(s, f, g) generic squared distance, same as WeightedSqNorm(s, f-g)

Operations involving complex numbers:

 ComplexReal(f) Real part of complex (vectorized) ComplexImag(f) Imaginary part of complex (vectorized) Real2Complex(f) convert real vector to complex vector with zero imaginary part (F+0*i) Imag2Complex(f) convert real vector to complex vector with zero real part (0+i*F) Conj(f) Complex conjugate (vectorized) ComplexAbs(f) Absolute value or modulus of complex (vectorized) ComplexSquareAbs(f) Square of modulus of complex (vectorized) ComplexAngle(f) Angle of complex (vectorized) ComplexSum(f) Sum of complex vector ComplexSumT(f,dim) Adjoint operation of ComplexSum - replicates f (complex scalar) dim times ComplexMult(f,g) Complex multiplication of f and g (vectorized) ComplexScal(f,g) Multiplication of f (complex scalar) with g (complex vector) ComplexRealScal(f,g) Multiplication of f (real scalar) with g (complex vector) ComplexDivide(f,g) Complex division of f and g (vectorized)

 IntCst(N) integer constant N IntInv(N) alias for Inv(IntCst(N)) : 1/N Zero(N) vector of zeros of size N Sum(f) sum of elements of vector f Max(f) max of elements of vector f Min(f) min of elements of vector f ArgMax(f) argmax of elements of vector f ArgMin(f) argmin of elements of vector f Elem(f, M) extract M-th element of vector f ElemT(f, N, M) insert scalar value f at position M in a vector of zeros of length N Extract(f, M, D) extract sub-vector from vector f (M is starting index, D is dimension of sub-vector) ExtractT(f, M, D) insert vector f in a larger vector of zeros (M is starting index, D is dimension of output) Concat(f, g) concatenation of vectors f and g OneHot(f, D) encodes a (rounded) scalar value as a one-hot vector of dimension D

Elementary dot products:

 MatVecMult(f, g) matrix-vector product f x g: f is vector interpreted as matrix (column-major), g is vector VecMatMult(f, g) vector-matrix product f x g: f is vector, g is vector interpreted as matrix (column-major) TensorProd(f, g) tensor cross product f x g^T: f and g are vectors of sizes M and N, output is of size MN. TensorDot(f, g, dimf, dimg, contf, contg) tensordot product f : g(similar to numpy\'s tensordot _ in the spirit): f and g are tensors of sizes listed in dimf and dimg :ref:index sequences  and contracted along the dimensions listed in contf and contg :ref:index sequences . The MatVecMult, VecMatMult and TensorProd operations are special cases of TensorDot.

 Grad(f,x,e) gradient of f with respect to the variable x, with e as the "grad_input" to backpropagate GradMatrix(f, v) matrix of gradient (i.e. transpose of the jacobian matrix)

## Reductions

The operations that can be used to reduce an array are described in the following table.

code name arguments mathematical expression (reduction over j) remarks
Sum f \sum_j f_{ij}
Max_SumShiftExp f (scalar) (m_i,s_i) with \left\{\begin{array}{l}m_i=\max_j f_{ij}\\s_i=\sum_j\exp(f_{ij}-m_i)\end{array}\right.
• core KeOps reduction for LogSumExp.
LogSumExp f (scalar) \log\left(\sum_j\exp(f_{ij})\right) only in Python bindings
Max_SumShiftExpWeight f (scalar), g (m_i,s_i) with \left\{\begin{array}{l}m_i=\max_j f_{ij}\\s_i=\sum_j\exp(f_{ij}-m_i)g_{ij}\end{array}\right.
• core KeOps reduction for LogSumExpWeight and SumSoftMaxWeight.
LogSumExpWeight f (scalar), g \log\left(\sum_j\exp(f_{ij})g_{ij}\right) only in Python bindings
SumSoftMaxWeight f (scalar), g \left(\sum_j\exp(f_{ij})g_{ij}\right)/\left(\sum_j\exp(f_{ij})\right) only in Python bindings
Min f \min_j f_{ij} no gradient
ArgMin f \text{argmin}_jf_{ij} gradient xreturns zeros
Min_ArgMin f \left(\min_j f_{ij} ,\text{argmin}_j f_{ij}\right) no gradient
Max f \max_j f_{ij} no gradient
ArgMax f \text{argmax}_j f_{ij} gradient returns zeros
Max_ArgMax f \left(\max_j f_{ij},\text{argmax}_j f_{ij}\right) no gradient
KMin f, K (int) \begin{array}{l}\left[\min_j f_{ij},\ldots,\min^{(K)}_jf_{ij}\right] \\(\min^{(k)}\text{means k-th smallest value})\end{array} no gradient
ArgKMin f, K (int) \left[\text{argmin}_jf_{ij},\ldots,\text{argmin}^{(K)}_j f_{ij}\right] gradient returns zeros
KMin_ArgKMin f, K (int) \left([\min^{(1...K)}_j f_{ij} ],[\text{argmin}^{(1...K)}_j f_{ij}]\right) no gradient

N.B.: All these reductions, except Max_SumShiftExp and LogSumExp, are vectorized : whenever the input f or g is vector-valued, the output will be vector-valued, with the corresponding reduction applied element-wise to each component.

N.B.: All reductions accept an additional optional argument that specifies wether the reduction is performed over the j or the i index. (see :ref:part.cppapi and :ref:part.genred)

## An example

Assume we want to compute the sum

F(p,x,y,a)_i = \left(\sum_{j=1}^N (p -a_j )^2 \exp(x_i^u + y_j^u) \right)_{i=1,\ldots,M, u=1,2,3} \in \mathbb R^{M\times 3}


where:

• p \in \mathbb R is a parameter,
• x \in \mathbb R^{M\times 3} is an x-variable indexed by i,
• y \in \mathbb R^{N\times 3} is an y-variable indexed by j,
• a \in \mathbb R^N is an y-variable indexed by j.

Using the variable placeholders presented above and the mathematical operations listed in :ref:part.mathOperation, we can define F as a symbolic string

F = "Sum_Reduction( Square( Pm(0,1) - Vj(3,1) )  *  Exp( Vi(1,3) + Vj(2,3) ), 1 )"

in which + and - denote the usual addition of vectors, Exp is the (element-wise) exponential function and * denotes scalar-vector multiplication. The second argument 1 of the Sum_Reduction operator indicates that the summation is performed with respect to the j index: a 0 would have been associated with an i-reduction.

Note that in all bindings, variables can be defined through aliases. In this example, we may write p=Pm(0,1), x=Vi(1,3), y=Vj(2,3), a=Vj(3,1) and thus give F through a much friendlier expression:

F = "Sum_Reduction( Square(p - a) * Exp(x + y), 1 )"