Skip to content

Latest commit

 

History

History
executable file
·
3186 lines (2235 loc) · 93.4 KB

maths.md

File metadata and controls

executable file
·
3186 lines (2235 loc) · 93.4 KB

Math Functions

Torch provides MATLAB-like functions for manipulating Tensor objects. Functions fall into several types of categories:

By default, all operations allocate a new Tensor to return the result. However, all functions also support passing the target Tensor(s) as the first argument(s), in which case the target Tensor(s) will be resized accordingly and filled with result. This property is especially useful when one wants have tight control over when memory is allocated.

The Torch package adopts the same concept, so that calling a function directly on the Tensor itself using an object-oriented syntax is equivalent to passing the Tensor as the optional resulting Tensor. The following two calls are equivalent.

torch.log(x, x)
x:log()

Similarly, torch.conv2 function can be used in the following manner.

> x = torch.rand(100, 100)
> k = torch.rand(10, 10)
> res1 = torch.conv2(x, k)   -- case 1

> res2 = torch.Tensor()
> torch.conv2(res2, x, k)     -- case 2

> res2:dist(res1)
0

The advantage of second case is, same res2 Tensor can be used successively in a loop without any new allocation.

-- no new memory allocations...
> for i = 1, 100 do
     torch.conv2(res2, x, k)
  end

> res2:dist(res1)
0

Construction or extraction functions

[res] torch.cat( [res,] x_1, x_2, [dimension] )

[res] torch.cat( [res,] {x_1, x_2, ...}, [dimension] )

x = torch.cat(x_1, x_2, [dimension]) returns a Tensor x which is the concatenation of Tensors x_1 and x_2 along dimension dimension.

If dimension is not specified or if it is -1, it is the maximum last dimension over all input tensors, except if all tensors are empty, then it is 1.

The other dimensions of x_1 and x_2 have to be equal.

Also supports arrays with arbitrary numbers of Tensors as inputs.

Empty tensors are ignored during catting, and thus do not throw an error. Performing cat on empty tensors only will always result in an empty tensor.

Examples:

> torch.cat(torch.ones(3), torch.zeros(2))
 1
 1
 1
 0
 0
[torch.DoubleTensor of size 5]

> torch.cat(torch.ones(3, 2), torch.zeros(2, 2), 1)
 1  1
 1  1
 1  1
 0  0
 0  0
[torch.DoubleTensor of size 5x2]

> torch.cat(torch.ones(2, 2), torch.zeros(2, 2), 1)
 1  1
 1  1
 0  0
 0  0
[torch.DoubleTensor of size 4x2]

> torch.cat(torch.ones(2, 2), torch.zeros(2, 2), 2)
 1  1  0  0
 1  1  0  0
[torch.DoubleTensor of size 2x4]

> torch.cat(torch.cat(torch.ones(2, 2), torch.zeros(2, 2), 1), torch.rand(3, 2), 1)
 1.0000  1.0000
 1.0000  1.0000
 0.0000  0.0000
 0.0000  0.0000
 0.3227  0.0493
 0.9161  0.1086
 0.2206  0.7449
[torch.DoubleTensor of size 7x2]

> torch.cat({torch.ones(2, 2), torch.zeros(2, 2), torch.rand(3, 2)}, 1)
 1.0000  1.0000
 1.0000  1.0000
 0.0000  0.0000
 0.0000  0.0000
 0.3227  0.0493
 0.9161  0.1086
 0.2206  0.7449
[torch.DoubleTensor of size 7x2]

> torch.cat({torch.Tensor(), torch.rand(3, 2)}, 1)
 0.3227  0.0493
 0.9161  0.1086
 0.2206  0.7449
[torch.DoubleTensor of size 3x2]

[res] torch.diag([res,] x [,k])

y = torch.diag(x) when x is of dimension 1 returns a diagonal matrix with diagonal elements constructed from x.

y = torch.diag(x) when x is of dimension 2 returns a Tensor of dimension 1 with elements constructed from the diagonal of x.

y = torch.diag(x, k) returns the k-th diagonal of x, where k = 0 is the main diagonal, k > 0 is above the main diagonal and k < 0 is below the main diagonal.

[res] torch.eye([res,] n [,m])

y = torch.eye(n) returns the n × n identity matrix.

y = torch.eye(n, m) returns an n × m identity matrix with ones on the diagonal and zeros elsewhere.

[res] torch.histc([res,] x [,nbins, min_value, max_value])

y = torch.histc(x) returns the histogram of the elements in x. By default the elements are sorted into 100 equally spaced bins between the minimum and maximum values of x.

y = torch.histc(x, n) same as above with n bins.

y = torch.histc(x, n, min, max) same as above with n bins and [min, max] as elements range.

[res] torch.bhistc([res,] x [,nbins, min_value, max_value])

y = torch.bhistc(x) returns the histogram of the elements in 2d tensor x along the last dimension. By default the elements are sorted into 100 equally spaced bins between the minimum and maximum values of x.

y = torch.bhistc(x, n) same as above with n bins.

y = torch.bhistc(x, n, min, max) same as above with n bins and [min, max] as elements range.

x = torch.Tensor(3, 6)

> x[1] = torch.Tensor{ 2, 4, 2, 2, 5, 4 }
> x[2] = torch.Tensor{ 3, 5, 1, 5, 3, 5 }
> x[3] = torch.Tensor{ 3, 4, 2, 5, 5, 1 }

> x
 2  4  2  2  5  4
 3  5  1  5  3  5
 3  4  2  5  5  1
[torch.DoubleTensor of size 3x6]

> torch.bhistc(x, 5, 1, 5)
 0  3  0  2  1
 1  0  2  0  3
 1  1  1  1  2
[torch.DoubleTensor of size 3x5]

> y = torch.Tensor(1, 6):copy(x[1])

> torch.bhistc(y, 5)
 3  0  2  0  1
[torch.DoubleTensor of size 1x5]

[res] torch.linspace([res,] x1, x2, [,n])

y = torch.linspace(x1, x2) returns a one-dimensional Tensor of size 100 equally spaced points between x1 and x2.

y = torch.linspace(x1, x2, n) returns a one-dimensional Tensor of n equally spaced points between x1 and x2.

[res] torch.logspace([res,] x1, x2, [,n])

y = torch.logspace(x1, x2) returns a one-dimensional Tensor of 100 logarithmically eqally spaced points between 10^x1 and 10^x2.

y = torch.logspace(x1, x2, n) returns a one-dimensional Tensor of n logarithmically equally spaced points between 10^x1 and 10^x2.

[res] torch.multinomial([res,], p, n, [,replacement])

y = torch.multinomial(p, n) returns a Tensor y where each row contains n indices sampled from the multinomial probability distribution located in the corresponding row of Tensor p.

The rows of p do not need to sum to one (in which case we use the values as weights), but must be non-negative and have a non-zero sum. Indices are ordered from left to right according to when each was sampled (first samples are placed in first column).

If p is a vector, y is a vector size n.

If p is a m-rows matrix, y is an m × n matrix.

If replacement is true, samples are drawn with replacement. If not, they are drawn without replacement, which means that when a sample index is drawn for a row, it cannot be drawn again for that row. This implies the constraint that n must be lower than p length (or number of columns of p if it is a matrix).

The default value for replacement is false.

p = torch.Tensor{1, 1, 0.5, 0}
a = torch.multinomial(p, 10000, true)

> a
...
[torch.LongTensor of dimension 10000]

> for i = 1, 4 do print(a:eq(i):sum()) end
3967
4016
2017
0

Note: If you use the function with a given result Tensor, i.e. of the function prototype: torch.multinomial(res, p, n [, replacement]) then you will have to call it slightly differently as:

p.multinomial(res, p, n, replacement) -- p.multinomial instead of torch.multinomial

This is due to the fact that the result here is of a LongTensor type, and we do not define a torch.multinomial over long Tensors.

[state] torch.multinomialAliasSetup(probs)

[res] torch.multinomialAlias(output, state)

state = torch.multinomialAliasSetup(probs) returns a table state consisting of two tensors : probability table and an alias table. This is required once for each probs vectors. We can then sample from the multinomial distribution multiple times by consulting these tensors state table.

torch.multinomialAlias(output, state) returns output filled with indices drawn from the multinomial distribution probs. output itself is filled with the indices and it is not necessary to get the return value of the statement.

The sampling is done through a technique defined in a very simple way in this blog about The Alias Method. The paper that describes this technique is present here. This can only sample with replacement.

The output Tensor that is fed into the multinomialAlias method need not be contiguous. The output tensor can only be a 1d tensor. If you are required to fill a nd tensor enter a 1d view of the same tensor. This method is exceptionally faster than torch.multinomial when you want to sample a lot of samples from the same distrbution or sample from the same distribution a large number of times. torch.multinomial is faster for sampling few samples from a distribution once because the multinomialAliasSetup method takes some time in this case. To see and compare how these two methods differ in speed run th test/test_aliasMultinomial.lua.

> state = torch.multinomialAliasSetup(probs)
> state
{
  1 : LongTensor - size: 4
  2 : DoubleTensor - size: 4
}
> output = torch.LongTensor(2,3)
> torch.multinomialAlias(output:view(-1), state)
 4
 1
 2
 3
 2
 2
[torch.LongTensor of size 6]
> output
 4  1  2
 3  2  2
[torch.LongTensor of size 2x3]

You can also allocate memory and reuse it for the state table.

> state = {torch.LongTensor(), torch.DoubleTensor()}
> probs = torch.DoubleTensor({0.2, 0.3, 0.5})
> state = torch.multinomialAliasSetup(probs, state)
> state
{
  1 : LongTensor - size: 3
  2 : DoubleTensor - size: 3
}
> output = torch.LongTensor(7)
> torch.multinomialAlias(output, state)
 2
 2
 3
 1
 2
 2
 2
[torch.LongTensor of size 7]

[res] torch.ones([res,] m [,n...])

y = torch.ones(n) returns a one-dimensional Tensor of size n filled with ones.

y = torch.ones(m, n) returns a m × n Tensor filled with ones.

For more than 4 dimensions, you can use a storage as argument: y = torch.ones(torch.LongStorage{m, n, k, l, o}).

[res] torch.rand([res,] [gen,] m [,n...])

y = torch.rand(n) returns a one-dimensional Tensor of size n filled with random numbers from a uniform distribution on the interval [0, 1).

y = torch.rand(m, n) returns a m × n Tensor of random numbers from a uniform distribution on the interval [0, 1).

For more than 4 dimensions, you can use a storage as argument: y = torch.rand(torch.LongStorage{m, n, k, l, o}).

y = torch.rand(gen, m, n) returns a m × n Tensor of random numbers from a uniform distribution on the interval [0, 1), using a non-global random number generator gen created by torch.Generator().

[res] torch.randn([res,] [gen,] m [,n...])

y = torch.randn(n) returns a one-dimensional Tensor of size n filled with random numbers from a normal distribution with mean zero and variance one.

y = torch.randn(m, n) returns a m × n Tensor of random numbers from a normal distribution with mean zero and variance one.

For more than 4 dimensions, you can use a storage as argument: y = torch.randn(torch.LongStorage{m, n, k, l, o}).

y = torch.randn(gen, m, n) returns a m × n Tensor of random numbers from a normal distribution with mean zero and variance one, using a non-global random number generator gen created by torch.Generator().

[res] torch.range([res,] x, y [,step])

y = torch.range(x, y) returns a Tensor of size floor((y - x) / step) + 1 with values from x to y with step step (default to 1).

> torch.range(2, 5)
 2
 3
 4
 5
[torch.DoubleTensor of size 4]

> torch.range(2, 5, 1.2)
 2.0000
 3.2000
 4.4000
[torch.DoubleTensor of size 3]

[res] torch.randperm([res,] [gen,] n)

y = torch.randperm(n) returns a random permutation of integers from 1 to n.

y = torch.randperm(gen, n) returns a random permutation of integers from 1 to n, using a non-global random number generator gen created by torch.Generator().

[res] torch.reshape([res,] x, m [,n...])

y = torch.reshape(x, m, n) returns a new m × n Tensor y whose elements are taken rowwise from x, which must have m * n elements. The elements are copied into the new Tensor.

For more than 4 dimensions, you can use a storage: y = torch.reshape(x, torch.LongStorage{m, n, k, l, o}).

[res] torch.tril([res,] x [,k])

y = torch.tril(x) returns the lower triangular part of x, the other elements of y are set to 0.

torch.tril(x, k) returns the elements on and below the k-th diagonal of x as non-zero. k = 0 is the main diagonal, k > 0 is above the main diagonal and k < 0 is below the main diagonal.

[res] torch.triu([res,] x, [,k])

y = torch.triu(x) returns the upper triangular part of x, the other elements of y are set to 0.

torch.triu(x, k) returns the elements on and above the k-th diagonal of x as non-zero. k = 0 is the main diagonal, k > 0 is above the main diagonal and k < 0 is below the main diagonal.

[res] torch.zeros([res,] x)

y = torch.zeros(n) returns a one-dimensional Tensor of size n filled with zeros.

y = torch.zeros(m, n) returns a m × n Tensor filled with zeros.

For more than 4 dimensions, you can use a storage: y = torch.zeros(torch.LongStorage{m, n, k, l, o}).

Element-wise Mathematical Operations

[res] torch.abs([res,] x)

y = torch.abs(x) returns a new Tensor with the absolute values of the elements of x.

x:abs() replaces all elements in-place with the absolute values of the elements of x.

[res] torch.sign([res,] x)

y = torch.sign(x) returns a new Tensor with the sign (+/- 1) of the elements of x.

x:sign() replaces all elements in-place with the sign of the elements of x.

[res] torch.acos([res,] x)

y = torch.acos(x) returns a new Tensor with the arcosine of the elements of x.

x:acos() replaces all elements in-place with the arcosine of the elements of x.

[res] torch.asin([res,] x)

y = torch.asin(x) returns a new Tensor with the arcsine of the elements of x.

x:asin() replaces all elements in-place with the arcsine of the elements of x.

[res] torch.atan([res,] x)

y = torch.atan(x) returns a new Tensor with the arctangent of the elements of x.

x:atan() replaces all elements in-place with the arctangent of the elements of x.

[res] torch.atan2([res,] x, y)

y = torch.atan2(x, y) returns a new Tensor with the arctangent of the elements of x and y. Note that the arctangent of the elements x and y refers to the signed angle in radians between the rays ending at origin where the first one starts at (1, 0) and the second at (y, x).

x:atan2() replaces all elements in-place with the arctangent of the elements of x and y.

[res] torch.ceil([res,] x)

y = torch.ceil(x) returns a new Tensor with the values of the elements of x rounded up to the nearest integers.

x:ceil() replaces all elements in-place with the values of the elements of x rounded up to the nearest integers.

[res] torch.cos([res,] x)

y = torch.cos(x) returns a new Tensor with the cosine of the elements of x.

x:cos() replaces all elements in-place with the cosine of the elements of x.

[res] torch.cosh([res,] x)

y = torch.cosh(x) returns a new Tensor with the hyberbolic cosine of the elements of x.

x:cosh() replaces all elements in-place with the hyberbolic cosine of the elements of x.

[res] torch.exp([res,] x)

y = torch.exp(x) returns, for each element in x, e (Neper number, the base of natural logarithms) raised to the power of the element in x.

x:exp() returns, for each element in x, e raised to the power of the element in x.

[res] torch.floor([res,] x)

y = torch.floor(x) returns a new Tensor with the values of the elements of x rounded down to the nearest integers.

x:floor() replaces all elements in-place with the values of the elements of x rounded down to the nearest integers.

[res] torch.log([res,] x)

y = torch.log(x) returns a new Tensor with the natural logarithm of the elements of x.

x:log() replaces all elements in-place with the natural logarithm of the elements of x.

[res] torch.log1p([res,] x)

y = torch.log1p(x) returns a new Tensor with the natural logarithm of the elements of x + 1.

x:log1p() replaces all elements in-place with the natural logarithm of the elements of x + 1. This function is more accurate than log for small values of x.

x:neg()

x:neg() replaces all elements in-place with the sign-reversed values of the elements of x.

x:cinv()

x:cinv() replaces all elements in-place with 1.0 / x.

[res] torch.pow([res,] x, n)

Let x be a Tensor and n a number.

y = torch.pow(x, n) returns a new Tensor with the elements of x to the power of n.

y = torch.pow(n, x) returns, a new Tensor with n to the power of the elements of x.

x:pow(n) replaces all elements in-place with the elements of x to the power of n.

torch.pow(x, n, x) replaces all elements in-place with n to the power of the elements of x.

[res] torch.round([res,] x)

y = torch.round(x) returns a new Tensor with the values of the elements of x rounded to the nearest integers.

x:round() replaces all elements in-place with the values of the elements of x rounded to the nearest integers.

[res] torch.sin([res,] x)

y = torch.sin(x) returns a new Tensor with the sine of the elements of x.

x:sin() replaces all elements in-place with the sine of the elements of x.

[res] torch.sinh([res,] x)

y = torch.sinh(x) returns a new Tensor with the hyperbolic sine of the elements of x.

x:sinh() replaces all elements in-place with the hyperbolic sine of the elements of x.

[res] torch.sqrt([res,] x)

y = torch.sqrt(x) returns a new Tensor with the square root of the elements of x.

x:sqrt() replaces all elements in-place with the square root of the elements of x.

[res] torch.rsqrt([res,] x)

y = torch.rsqrt(x) returns a new Tensor with the reciprocal of the square root of the elements of x.

x:rsqrt() replaces all elements in-place with the reciprocal of the square root of the elements of x.

[res] torch.tan([res,] x)

y = torch.tan(x) returns a new Tensor with the tangent of the elements of x.

x:tan() replaces all elements in-place with the tangent of the elements of x.

[res] torch.tanh([res,] x)

y = torch.tanh(x) returns a new Tensor with the hyperbolic tangent of the elements of x.

x:tanh() replaces all elements in-place with the hyperbolic tangent of the elements of x.

[res] torch.sigmoid([res,] x)

y = torch.sigmoid(x) returns a new Tensor with the sigmoid of the elements of x.

x:sigmoid() replaces all elements in-place with the sigmoid of the elements of x.

[res] torch.trunc([res,] x)

y = torch.trunc(x) returns a new Tensor with the truncated integer values of the elements of x.

x:trunc() replaces all elements in-place with the truncated integer values of the elements of x.

[res] torch.frac([res,] x)

y = torch.frac(x) returns a new Tensor with the fractional portion of the elements of x.

x:frac() replaces all elements in-place with the fractional portion of the elements of x.

Basic operations

In this section, we explain basic mathematical operations for Tensors.

[boolean] equal([tensor1,] tensor2)

Returns true iff the dimensions and values of tensor1 and tensor2 are exactly the same.

x = torch.Tensor{1,2,3}
y = torch.Tensor{1,2,3}
> x:equal(y)
true

y = torch.Tensor{1,2,4}
> x:equal(y)
false

Note that a:equal(b) is more efficient that a:eq(b):all() as it avoids allocation of a temporary tensor and can short-circuit.

[res] torch.add([res,] tensor, value)

Add the given value to all elements in the Tensor.

y = torch.add(x, value) returns a new Tensor.

x:add(value) add value to all elements in place.

[res] torch.add([res,] tensor1, tensor2)

Add tensor1 to tensor2 and put result into res. The number of elements must match, but sizes do not matter.

> x = torch.Tensor(2, 2):fill(2)
> y = torch.Tensor(4):fill(3)
> x:add(y)
> x
 5  5
 5  5
[torch.DoubleTensor of size 2x2]

y = torch.add(a, b) returns a new Tensor.

torch.add(y, a, b) puts a + b in y.

a:add(b) accumulates all elements of b into a.

y:add(a, b) puts a + b in y.

[res] torch.add([res,] tensor1, value, tensor2)

Multiply elements of tensor2 by the scalar value and add it to tensor1. The number of elements must match, but sizes do not matter.

> x = torch.Tensor(2, 2):fill(2)
> y = torch.Tensor(4):fill(3)
> x:add(2, y)
> x
 8  8
 8  8
[torch.DoubleTensor of size 2x2]

x:add(value, y) multiply-accumulates values of y into x.

z:add(x, value, y) puts the result of x + value * y in z.

torch.add(x, value, y) returns a new Tensor x + value * y.

torch.add(z, x, value, y) puts the result of x + value * y in z.

tensor:csub(value)

Subtracts the given value from all elements in the Tensor, in place.

tensor:csub(tensor2)

Subtracts tensor2 from tensor, in place. The number of elements must match, but sizes do not matter.

> x = torch.Tensor(2, 2):fill(8)
> y = torch.Tensor(4):fill(3)
> x:csub(y)
> x
 5  5
 5  5
[torch.DoubleTensor of size 2x2]

a:csub(b) put a - b into a.

[res] torch.mul([res,] tensor1, value)

Multiply all elements in the Tensor by the given value.

z = torch.mul(x, 2) will return a new Tensor with the result of x * 2.

torch.mul(z, x, 2) will put the result of x * 2 in z.

x:mul(2) will multiply all elements of x with 2 in-place.

z:mul(x, 2) will put the result of x * 2 in z.

[res] torch.clamp([res,] tensor, min_value, max_value)

Clamp all elements in the Tensor into the range [min_value, max_value]. ie:

      ⎧ min_value, if x_i < min_value
y_i = ⎨ x_i,       if min_value ≤ x_i ≤ max_value
      ⎩ max_value, if x_i > max_value

z = torch.clamp(x, 0, 1) will return a new Tensor with the result of x bounded between 0 and 1.

torch.clamp(z, x, 0, 1) will put the result in z.

x:clamp(0, 1) will perform the clamp operation in place (putting the result in x).

z:clamp(x, 0, 1) will put the result in z.

[res] torch.cmul([res,] tensor1, tensor2)

Element-wise multiplication of tensor1 by tensor2. The number of elements must match, but sizes do not matter.

> x = torch.Tensor(2, 2):fill(2)
> y = torch.Tensor(4):fill(3)
> x:cmul(y)
> = x
 6  6
 6  6
[torch.DoubleTensor of size 2x2]

z = torch.cmul(x, y) returns a new Tensor.

torch.cmul(z, x, y) puts the result in z.

y:cmul(x) multiplies all elements of y with corresponding elements of x.

z:cmul(x, y) puts the result in z.

[res] torch.cpow([res,] tensor1, tensor2)

Element-wise power operation, taking the elements of tensor1 to the powers given by elements of tensor2. The number of elements must match, but sizes do not matter.

> x = torch.Tensor(2, 2):fill(2)
> y = torch.Tensor(4):fill(3)
> x:cpow(y)
> x
 8  8
 8  8
[torch.DoubleTensor of size 2x2]

z = torch.cpow(x, y) returns a new Tensor.

torch.cpow(z, x, y) puts the result in z.

y:cpow(x) takes all elements of y to the powers given by the corresponding elements of x.

z:cpow(x, y) puts the result in z.

[res] torch.addcmul([res,] x [,value], tensor1, tensor2)

Performs the element-wise multiplication of tensor1 by tensor2, multiply the result by the scalar value (1 if not present) and add it to x. The number of elements must match, but sizes do not matter.

> x = torch.Tensor(2, 2):fill(2)
> y = torch.Tensor(4):fill(3)
> z = torch.Tensor(2, 2):fill(5)
> x:addcmul(2, y, z)
> x
 32  32
 32  32
[torch.DoubleTensor of size 2x2]

z:addcmul(value, x, y) accumulates the result in z.

torch.addcmul(z, value, x, y) returns a new Tensor with the result.

torch.addcmul(z, z, value, x, y) puts the result in z.

[res] torch.div([res,] tensor, value)

Divide all elements in the Tensor by the given value.

z = torch.div(x, 2) will return a new Tensor with the result of x / 2.

torch.div(z, x, 2) will put the result of x / 2 in z.

x:div(2) will divide all elements of x with 2 in-place.

z:div(x, 2) puts the result of x / 2 in z.

[res] torch.cdiv([res,] tensor1, tensor2)

Performs the element-wise division of tensor1 by tensor2. The number of elements must match, but sizes do not matter.

> x = torch.Tensor(2, 2):fill(1)
> y = torch.range(1, 4)
> x:cdiv(y)
> x
 1.0000  0.5000
 0.3333  0.2500
[torch.DoubleTensor of size 2x2]

z = torch.cdiv(x, y) returns a new Tensor.

torch.cdiv(z, x, y) puts the result in z.

y:cdiv(x) divides all elements of y with corresponding elements of x.

z:cdiv(x, y) puts the result in z.

[res] torch.lshift([res,] tensor, value)

Left shift all elements in the Tensor by the given value.

z = torch.lshift(x, 2) will return a new Tensor with the result of x << 2.

torch.lshift(z, x, 2) will put the result of x << 2 in z.

x:lshift(2) will perform left shift operation all elements of x by 2 bits.

z:lshift(x, 2) puts the result of x << 2 in z.

Note: For float type tensors, x:lshift(value) evaluates x:mul(math.pow(2, value)) internally.

[res] torch.clshift([res,] tensor1, tensor2)

Performs the left shift operation of each element in tensor1 by each element in tensor2. The number of elements must match, but sizes do not matter.

> x = torch.LongTensor(2, 2):fill(1)
> y = torch.LongTensor(2, 2):range(1, 4)
> x:clshift(y)
> x
 2  4
 8 16
[torch.LongTensor of size 2x2]

z = torch.clshift(x, y) returns a new Tensor.

torch.clshift(z, x, y) puts the result in z.

y:clshift(x) left shifts all elements of y with corresponding elements of x.

z:clshift(x, y) puts the result in z.

[res] torch.rshift([res,] tensor, value)

Right shift all elements in the Tensor by the given value.

z = torch.rshift(x, 2) will return a new Tensor with the result of x >> 2.

torch.rshift(z, x, 2) will put the result of x >> 2 in z.

x:rshift(2) will perform right shift operation all elements of x by 2 bits.

z:rshift(x, 2) puts the result of x >> 2 in z.

Note: For float type tensors, x:lshift(value) evaluates x:div(math.pow(2, value)) internally.

[res] torch.crshift([res,] tensor1, tensor2)

Performs the right shift operation of each element in tensor1 by each element in tensor2. The number of elements must match, but sizes do not matter.

> x = torch.LongTensor(2, 2):fill(32)
> y = torch.LongTensor(2, 2):range(1, 4)
> x:crshift(y)
> x
 16 8
  4 2
[torch.LongTensor of size 2x2]

z = torch.crshift(x, y) returns a new Tensor.

torch.crshift(z, x, y) puts the result in z.

y:crshift(x) right shifts all elements of y with corresponding elements of x.

z:crshift(x, y) puts the result in z.

[res] torch.addcdiv([res,] x [,value], tensor1, tensor2)

Performs the element-wise division of tensor1 by tensor2, multiply the result by the scalar value and add it to x. The number of elements must match, but sizes do not matter.

> x = torch.Tensor(2, 2):fill(1)
> y = torch.range(1, 4)
> z = torch.Tensor(2, 2):fill(5)
> x:addcdiv(2, y, z)
> x
 1.4000  1.8000
 2.2000  2.6000
[torch.DoubleTensor of size 2x2]

z:addcdiv(value, x, y) accumulates the result in z.

torch.addcdiv(z, value, x, y) returns a new Tensor with the result.

torch.addcdiv(z, z, value, x, y) puts the result in z.

[res] torch.fmod([res,] tensor, value)

Computes remainder of division (rounded towards zero) of all elements in the Tensor by value. This works both for integer and floating point numbers. It behaves the same as Lua bulit-in function math.fmod() and a little bit different from torch.remainder() and % operator. For example:

> x = torch.Tensor({-3, 3})
> torch.fmod(x, 2)
-1
 1
[torch.DoubleTensor of size 2]

> torch.fmod(x, -2)
-1
 1
[torch.DoubleTensor of size 2]

> torch.remainder(x, 2)
 1
 1
[torch.DoubleTensor of size 2]

> torch.remainder(x, -2)
-1
-1
[torch.DoubleTensor of size 2]

z = torch.fmod(x, 2) will return a new Tensor with the result of math.fmod(x, 2).

torch.fmod(z, x, 2) will put the result of math.fmod(x, 2) in z.

x:fmod(2) will replace all elements of x the result of math.fmod(x, 2) in-place.

z:fmod(x, 2) puts the result of math.fmod(x, 2) in z.

[res] torch.remainder([res,] tensor, value)

Computes remainder of division (rounded to nearest) of all elements in the Tensor by value. This works both for integer and floating point numbers. It behaves the same as % operator and can be expressed as a % b = a - b * floor(a/b). See torch.fmod() for comparison.

z = torch.remainder(x, 2) will return a new Tensor with the result of x % 2.

torch.remainder(z, x, 2) will put the result of x % 2 in z.

x:remainder(2) will replace all elements of x the result of x % 2 in-place.

z:remainder(x, 2) puts the result of x % 2 in z.

[res] torch.mod([res,] tensor, value)

This function is deprecated and exists only for compatibility with previous versions. Please use torch.fmod() or torch.remainder() instead.

[res] torch.cfmod([res,] tensor1, tensor2)

Computes the element-wise remainder of the division (rounded towards zero) of tensor1 by tensor2. The number of elements must match, but sizes do not matter.

> x = torch.Tensor({{3, 3}, {-3, -3}})
> y = torch.Tensor({{2, -2}, {2, -2}})
> x:cfmod(y)
 1  1
-1 -1
[torch.DoubleTensor of size 2x2]

z = torch.cfmod(x, y) returns a new Tensor.

torch.cfmod(z, x, y) puts the result in z.

y:cfmod(x) replaces all elements of y by their remainders of division (rounded towards zero) by corresponding elements of x.

z:cfmod(x, y) puts the result in z.

[res] torch.cremainder([res,] tensor1, tensor2)

Computes element-wise remainder of the division (rounded to nearest) of tensor1 by tensor2. The number of elements must match, but sizes do not matter.

> x = torch.Tensor({{3, 3}, {-3, -3}})
> y = torch.Tensor({{2, -2}, {2, -2}})
> x:cfmod(y)
 1  1
-1 -1
[torch.DoubleTensor of size 2x2]

z = torch.cremainder(x, y) returns a new Tensor.

torch.cremainder(z, x, y) puts the result in z.

y:cremainder(x) replaces all elements of y by their remainders of division (rounded to nearest) by corresponding elements of x.

z:cremainder(x, y) puts the result in z.

[res] torch.cmod([res,] tensor1, tensor2)

This function is deprecated and exists only for compatibility with previous versions. Please use torch.cfmod() or torch.cremainder() instead.

[res] torch.bitand([res,] tensor, value)

Performs bitwise and operation on all elements in the Tensor by the given value.

z = torch.bitand(x, value) will return a new Tensor with the result of x & value.

torch.bitand(z, x, value) will put the result of x & value in z.

x:bitand(value) will perform right shift operation all elements of x by value bits.

z:bitand(x, value) puts the result of x & value in z.

Note: This function is only supported for [Int|Long|Byte]Tensors

[res] torch.cbitand([res,] tensor1, tensor2)

Performs bitwise and operation of each element in tensor1 by each element in tensor2. The number of elements must match, but sizes do not matter.

> x = torch.LongTensor(4):fill(6)
> y = torch.LongTensor{1, 2, 4, 8}
> x:cbitand(y)
> x
  0
  2
  4
  0
[torch.LongTensor of size 4]

z = torch.cbitand(x, y) returns a new Tensor.

torch.cbitand(z, x, y) puts the result in z.

y:cbitand(x) performs bitwise and all elements of y with corresponding elements of x.

z:cbitand(x, y) puts the result in z.

Note: This function is only supported for [Int|Long|Byte]Tensors

[res] torch.bitor([res,] tensor, value)

Performs bitwise or operation on all elements in the Tensor by the given value.

z = torch.bitor(x, value) will return a new Tensor with the result of x & value.

torch.bitor(z, x, value) will put the result of x | value in z.

x:bitor(value) will perform right shift operation all elements of x by value bits.

z:bitor(x, value) puts the result of x | value in z.

Note: This function is only supported for [Int|Long|Byte]Tensors

[res] torch.cbitor([res,] tensor1, tensor2)

Performs bitwise or operation of each element in tensor1 by each element in tensor2. The number of elements must match, but sizes do not matter.

> x = torch.LongTensor(4):fill(3)
> y = torch.LongTensor{1, 2, 4, 8}
> x:cbitor(y)
> x
  3
  3
  7
 11
[torch.LongTensor of size 4]

z = torch.cbitor(x, y) returns a new Tensor.

torch.cbitor(z, x, y) puts the result in z.

y:cbitor(x) performs bitwise or all elements of y with corresponding elements of x.

z:cbitor(x, y) puts the result in z.

Note: This function is only supported for [Int|Long|Byte]Tensors

[res] torch.bitxor([res,] tensor, value)

Performs bitwise xor operation on all elements in the Tensor by the given value.

z = torch.bitxor(x, value) will return a new Tensor with the result of x & value.

torch.bitxor(z, x, value) will put the result of x ^ value in z.

x:bitxor(value) will perform right shift operation all elements of x by value bits.

z:bitxor(x, value) puts the result of x ^ value in z.

Note: This function is only supported for [Int|Long|Byte]Tensors

[res] torch.cbitxor([res,] tensor1, tensor2)

Performs bitwise xor operation of each element in tensor1 by each element in tensor2. The number of elements must match, but sizes do not matter.

> x = torch.LongTensor(4):fill(15)
> y = torch.LongTensor{1, 2, 4, 8}
> x:cbitxor(y)
> x
  14
  13
  11
   7
[torch.LongTensor of size 4]

z = torch.cbitxor(x, y) returns a new Tensor.

torch.cbitxor(z, x, y) puts the result in z.

y:cbitxor(x) performs bitwise xor all elements of y with corresponding elements of x.

z:cbitxor(x, y) puts the result in z.

Note: This function is only supported for [Int|Long|Byte]Tensors

[number] torch.dot(tensor1, tensor2)

Performs the dot product between tensor1 and tensor2. The number of elements must match: both Tensors are seen as a 1D vector.

> x = torch.Tensor(2, 2):fill(2)
> y = torch.Tensor(4):fill(3)
> x:dot(y)
24

torch.dot(x, y) returns dot product of x and y. x:dot(y) returns dot product of x and y.

[res] torch.addmv([res,] [v1,] vec1, [v2,] mat, vec2)

Performs a matrix-vector multiplication between mat (2D Tensor) and vec2 (1D Tensor) and add it to vec1.

Optional values v1 and v2 are scalars that multiply vec1 and vec2 respectively.

In other words,

res = (v1 * vec1) + (v2 * (mat * vec2))

Sizes must respect the matrix-multiplication operation: if mat is a n × m matrix, vec2 must be vector of size m and vec1 must be a vector of size n.

> x = torch.Tensor(3):fill(0)
> M = torch.Tensor(3, 2):fill(3)
> y = torch.Tensor(2):fill(2)
> x:addmv(M, y)
> x
 12
 12
 12
[torch.DoubleTensor of size 3]

torch.addmv(x, y, z) returns a new Tensor with the result.

torch.addmv(r, x, y, z) puts the result in r.

Differences when used as a method

x:addmv(y, z) does x = x + y * z

r:addmv(x, y, z) does r = x + y * z if x is a vector

r:addmv(s, y, z) does r = r + s * y * z if s is a scalar.

r:addmv(x, s, y, z) does r = x + s * y * z if s is a scalar and x is a vector.

r:addmv(s1, s2, y, z) does r = s1 * r + s2 * y * z if s1 and s2 are scalars.

The last example does not accurately fit into the function signature, and needs a special mention. It changes the function signature to:

[vec1] = vec1:addmv([v1,] [v2,] mat, vec2)

[res] torch.addr([res,] [v1,] mat, [v2,] vec1, vec2)

Performs the outer-product between vec1 (1D Tensor) and vec2 (1D Tensor).

Optional values v1 and v2 are scalars that multiply mat and vec1 [out] vec2 respectively.

In other words,

res_ij = (v1 * mat_ij) + (v2 * vec1_i * vec2_j)

If vec1 is a vector of size n and vec2 is a vector of size m, then mat must be a matrix of size n × m.

> x = torch.range(1, 3)
> y = torch.range(1, 2)
> M = torch.Tensor(3, 2):zero()
> M:addr(x, y)
 1  2         --     |0 0|     |1 2|
 2  4         -- = 1*|0 0| + 1*|2 4|
 3  6         --     |0 0|     |3 6|
[torch.DoubleTensor of size 3x2]
-- default values of v1 and v2 are 1.

> M:addr(2, 1, x, y)
  3   6        --     |1 2|     |1 2|
  6  12        -- = 2*|2 4| + 1*|2 4|
  9  18        --     |3 6|     |3 6|
[torch.DoubleTensor of size 3x2]

> A = torch.range(1, 6):resize(3, 2)
> A
 1  2
 3  4
 5  6
[torch.DoubleTensor of size 3x2]
> M:addr(2, A, 1, x, y)
  3   6        --   |1 2|     |1 2|
  8  12        -- 2*|3 4| + 1*|2 4|
 13  18        --   |5 6|     |3 6|
[torch.DoubleTensor of size 3x2]

torch.addr(M, x, y) returns the result in a new Tensor.

torch.addr(r, M, x, y) puts the result in r.

M:addr(x, y) puts the result in M.

r:addr(M, x, y) puts the result in r.

[res] torch.addmm([res,] [v1,] M, [v2,] mat1, mat2)

Performs a matrix-matrix multiplication between mat1 (2D Tensor) and mat2 (2D Tensor).

Optional values v1 and v2 are scalars that multiply M and mat1 * mat2 respectively.

In other words,

res = (v1 * M) + (v2 * mat1 * mat2)

If mat1 is a n × m matrix, mat2 a m × p matrix, M must be a n × p matrix.

torch.addmm(M, mat1, mat2) returns the result in a new Tensor.

torch.addmm(r, M, mat1, mat2) puts the result in r.

Differences when used as a method

M:addmm(mat1, mat2) does M = M + mat1 * mat2.

r:addmm(M, mat1, mat2) does r = M + mat1 * mat2.

r:addmm(v1, M, v2, mat1, mat2) does r = (v1 * M) + (v2 * mat1 * mat2).

M:addmm(v1, v2, mat1, mat2) does M = (v1 * M) + (v2 * mat1 * mat2).

The last example does not accurately fit into the function signature, and needs a special mention. It changes the function signature to:

[M] = M:addmm([v1,] [v2,] mat1, mat2)

[res] torch.addbmm([res,] [v1,] M, [v2,] batch1, batch2)

Batch matrix matrix product of matrices stored in batch1 and batch2, with a reduced add step (all matrix multiplications get accumulated in a single place).

batch1 and batch2 must be 3D Tensors each containing the same number of matrices. If batch1 is a b × n × m Tensor, batch2 a b × m × p Tensor, res will be a n × p Tensor.

In other words,

res = (v1 * M) + (v2 * sum(batch1_i * batch2_i, i = 1, b))

torch.addbmm(M, x, y) puts the result in a new Tensor.

M:addbmm(x, y) puts the result in M, resizing M if necessary.

M:addbmm(beta, M2, alpha, x, y) puts the result in M, resizing M if necessary.

[res] torch.baddbmm([res,] [v1,] M, [v2,] batch1, batch2)

Batch matrix matrix product of matrices stored in batch1 and batch2, with batch add.

batch1 and batch2 must be 3D Tensors each containing the same number of matrices. If batch1 is a b × n × m Tensor, batch2 a b × m × p Tensor, res will be a b × n × p Tensor.

In other words,

res_i = (v1 * M_i) + (v2 * batch1_i * batch2_i)

torch.baddbmm(M, x, y) puts the result in a new Tensor.

M:baddbmm(x, y) puts the result in M, resizing M if necessary.

M:baddbmm(beta, M2, alpha, x, y) puts the result in M, resizing M if necessary.

[res] torch.mv([res,] mat, vec)

Matrix vector product of mat and vec. Sizes must respect the matrix-multiplication operation: if mat is a n × m matrix, vec must be vector of size m and res must be a vector of size n.

torch.mv(x, y) puts the result in a new Tensor.

torch.mv(M, x, y) puts the result in M.

M:mv(x, y) puts the result in M.

[res] torch.mm([res,] mat1, mat2)

Matrix matrix product of mat1 and mat2. If mat1 is a n × m matrix, mat2 a m × p matrix, res must be a n × p matrix.

torch.mm(x, y) puts the result in a new Tensor.

torch.mm(M, x, y) puts the result in M.

M:mm(x, y) puts the result in M.

[res] torch.bmm([res,] batch1, batch2)

Batch matrix matrix product of matrices stored in batch1 and batch2. batch1 and batch2 must be 3D Tensors each containing the same number of matrices. If batch1 is a b × n × m Tensor, batch2 a b × m × p Tensor, res will be a b × n × p Tensor.

torch.bmm(x, y) puts the result in a new Tensor.

torch.bmm(M, x, y) puts the result in M, resizing M if necessary.

M:bmm(x, y) puts the result in M, resizing M if necessary.

[res] torch.ger([res,] vec1, vec2)

Outer product of vec1 and vec2. If vec1 is a vector of size n and vec2 is a vector of size m, then res must be a matrix of size n × m.

torch.ger(x, y) puts the result in a new Tensor.

torch.ger(M, x, y) puts the result in M.

M:ger(x, y) puts the result in M.

[res] torch.lerp([res,] a, b, weight)

Linear interpolation of two scalars or tensors based on a weight: res = a + weight * (b - a)

torch.lerp(a, b, weight) puts the result in a new Tensor if a and b are tensors. If a and b are scalars the functions returns a number.

torch.lerp(M, a, b, weight) puts the result in M.

M:lerp(a, b, weight) puts the result in M.

Overloaded operators

It is possible to use basic mathematical operators like +, -, /, * and % with Tensors. These operators are provided as a convenience. While they might be handy, they create and return a new Tensor containing the results. They are thus not as fast as the operations available in the previous section.

Another important point to note is that these operators are only overloaded when the first operand is a Tensor. For example, this will NOT work:

> x = 5 + torch.rand(3)

Addition and subtraction

You can add a Tensor to another one with the + operator. Subtraction is done with -. The number of elements in the Tensors must match, but the sizes do not matter. The size of the returned Tensor will be the size of the first Tensor.

> x = torch.Tensor(2, 2):fill(2)
> y = torch.Tensor(4):fill(3)
> = x + y
 5  5
 5  5
[torch.DoubleTensor of size 2x2]

> = y - x
 1
 1
 1
 1
[torch.DoubleTensor of size 4]

A scalar might also be added or subtracted to a Tensor. The scalar needs to be on the right of the operator.

> x = torch.Tensor(2, 2):fill(2)
> = x + 3
 5  5
 5  5
[torch.DoubleTensor of size 2x2]

Negation

A Tensor can be negated with the - operator placed in front:

> x = torch.Tensor(2, 2):fill(2)
> = -x
-2 -2
-2 -2
[torch.DoubleTensor of size 2x2]

Multiplication

Multiplication between two Tensors is supported with the * operators. The result of the multiplication depends on the sizes of the Tensors.

  • 1D and 1D: Returns the dot product between the two Tensors (scalar).
  • 2D and 1D: Returns the matrix-vector operation between the two Tensors (1D Tensor).
  • 2D and 2D: Returns the matrix-matrix operation between the two Tensors (2D Tensor).

Sizes must be conformant for the corresponding operation.

A Tensor might also be multiplied by a scalar. The scalar might be on the right or left of the operator.

Examples:

> M = torch.Tensor(2, 2):fill(2)
> N = torch.Tensor(2, 4):fill(3)
> x = torch.Tensor(2):fill(4)
> y = torch.Tensor(2):fill(5)
> = x * y -- dot product
40

> = M * x --- matrix-vector
 16
 16
[torch.DoubleTensor of size 2]

> = M * N -- matrix-matrix
 12  12  12  12
 12  12  12  12
[torch.DoubleTensor of size 2x4]

Division and Modulo (remainder)

Only the division of a Tensor by a scalar is supported with the operator /.

Example:

> x = torch.Tensor(2, 2):fill(2)
> = x/3
 0.6667  0.6667
 0.6667  0.6667
[torch.DoubleTensor of size 2x2]

Similarly, the remainder of the division of a Tensors elements by a scalar can be obtained with the operator %.

Example:

x = torch.Tensor{{1,2},{3,4}} = x % 3 1 2 0 1 [torch.Tensor of size 2x2]

Column or row-wise operations (dimension-wise operations)

[res] torch.cross([res,] a, b [,n])

y = torch.cross(a, b) returns the cross product of a and b along the first dimension of length 3.

y = torch.cross(a, b, n) returns the cross product of vectors in dimension n of a and b.

a and b must have the same size, and both a:size(n) and b:size(n) must be 3.

[res] torch.cumprod([res,] x [,dim])

y = torch.cumprod(x) returns the cumulative product of the elements of x, performing the operation over the last dimension.

y = torch.cumprod(x, n) returns the cumulative product of the elements of x, performing the operation over dimension n.

-- 1. cumulative product for a vector
> A = torch.range(1, 5)
> A
 1
 2
 3
 4
 5
[torch.DoubleTensor of size 5]

> B = torch.cumprod(A)
> B
   1     -- B(1) = A(1) = 1
   2     -- B(2) = A(1)*A(2) = 1*2 = 2
   6     -- B(3) = A(1)*A(2)*A(3) = 1*2*3 = 6
  24     -- B(4) = A(1)*A(2)*A(3)*A(4) = 1*2*3*4 = 24
 120     -- B(5) = A(1)*A(2)*A(3)*A(4)*A(5) =1*2*3*4*5 = 120
[torch.DoubleTensor of size 5]

-- 2. cumulative product for a matrix
> A = torch.LongTensor{{1, 4, 7}, {2, 5, 8}, {3, 6, 9}}
> A
 1  4  7
 2  5  8
 3  6  9
[torch.LongTensor of size 3x3]

> B = torch.cumprod(A)
> B
   1    4    7
   2   20   56
   6  120  504
[torch.LongTensor of size 3x3]

-- Why?
-- B(1, 1) = A(1, 1) = 1
-- B(2, 1) = A(1, 1)*A(2, 1) = 1*2 = 2
-- B(3, 1) = A(1, 1)*A(2, 1)*A(3, 1) = 1*2*3 = 6
-- B(1, 2) = A(1, 2) = 4
-- B(2, 2) = A(1, 2)*A(2, 2) = 4*5 = 20
-- B(3, 2) = A(1, 2)*A(2, 2)*A(3, 2) = 4*5*6 = 120
-- B(1, 3) = A(1, 3) = 7
-- B(2, 3) = A(1, 3)*A(2, 3) = 7*8 = 56
-- B(3, 3) = A(1, 3)*A(2, 3)*A(3, 3) = 7*8*9 = 504

-- 3. cumulative product along 2-dim
> B = torch.cumprod(A, 2)
> B
   1    4   28
   2   10   80
   3   18  162
[torch.LongTensor of size 3x3]

-- Why?
-- B(1, 1) = A(1, 1) = 1
-- B(1, 2) = A(1, 1)*A(1, 2) = 1*4 = 4
-- B(1, 3) = A(1, 1)*A(1, 2)*A(1, 3) = 1*4*7 = 28
-- B(2, 1) = A(2, 1) = 2
-- B(2, 2) = A(2, 1)*A(2, 2) = 2*5 = 10
-- B(2, 3) = A(2, 1)*A(2, 2)*A(2, 3) = 2*5*8 = 80
-- B(3, 1) = A(3, 1) = 3
-- B(3, 2) = A(3, 1)*A(2, 3) = 3*6 = 18
-- B(3, 3) = A(3, 1)*A(2, 3)*A(3, 3) = 3*6*9 = 162

[res] torch.cumsum([res,] x [,dim])

y = torch.cumsum(x) returns the cumulative sum of the elements of x, performing the operation over the first dimension.

y = torch.cumsum(x, n) returns the cumulative sum of the elements of x, performing the operation over dimension n.

torch.max([resval, resind,] x [,dim])

y = torch.max(x) returns the single largest element of x.

y, i = torch.max(x, 1) returns the largest element in each column (across rows) of x, and a Tensor i of their corresponding indices in x.

y, i = torch.max(x, 2) performs the max operation for each row.

y, i = torch.max(x, n) performs the max operation over the dimension n.

> x = torch.randn(3, 3)
> x
 1.1994 -0.6290  0.6888
-0.0038 -0.0908 -0.2075
 0.3437 -0.9948  0.1216
[torch.DoubleTensor of size 3x3]

> torch.max(x)
1.1993977428735

> torch.max(x, 1)
 1.1994 -0.0908  0.6888
[torch.DoubleTensor of size 1x3]

 1  2  1
[torch.LongTensor of size 1x3]

> torch.max(x, 2)
 1.1994
-0.0038
 0.3437
[torch.DoubleTensor of size 3x1]

 1
 1
 1
[torch.LongTensor of size 3x1]

[res] torch.mean([res,] x [,dim])

y = torch.mean(x) returns the mean of all elements of x.

y = torch.mean(x, 1) returns a Tensor y of the mean of the elements in each column of x.

y = torch.mean(x, 2) performs the mean operation for each row.

y = torch.mean(x, n) performs the mean operation over the dimension n.

torch.min([resval, resind,] x [,dim])

y = torch.min(x) returns the single smallest element of x.

y, i = torch.min(x, 1) returns the smallest element in each column (across rows) of x, and a Tensor i of their corresponding indices in x.

y, i = torch.min(x, 2) performs the min operation for each row.

y, i = torch.min(x, n) performs the min operation over the dimension n.

[res] torch.cmax([res,] tensor1, tensor2)

Compute the maximum of each pair of values in tensor1 and tensor2.

c = torch.cmax(a, b) returns a new Tensor containing the element-wise maximum of a and b.

a:cmax(b) stores the element-wise maximum of a and b in a.

c:cmax(a, b) stores the element-wise maximum of a and b in c.

> a = torch.Tensor{1, 2, 3}
> b = torch.Tensor{3, 2, 1}
> torch.cmax(a, b)
 3
 2
 3
[torch.DoubleTensor of size 3]

[res] torch.cmax([res,] tensor, value)

Compute the maximum between each value in tensor and value.

c = torch.cmax(a, v) returns a new Tensor containing the maxima of each element in a and v.

a:cmax(v) stores the maxima of each element in a and v in a.

c:cmax(a, v) stores the maxima of each element in a and v in c.

> a = torch.Tensor{1, 2, 3}
> torch.cmax(a, 2)
 2
 2
 3
[torch.DoubleTensor of size 3]

[res] torch.cmin([res,] tensor1, tensor2)

Compute the minimum of each pair of values in tensor1 and tensor2.

c = torch.cmin(a, b) returns a new Tensor containing the element-wise minimum of a and b.

a:cmin(b) stores the element-wise minimum of a and b in a.

c:cmin(a, b) stores the element-wise minimum of a and b in c.

> a = torch.Tensor{1, 2, 3}
> b = torch.Tensor{3, 2, 1}
> torch.cmin(a, b)
 1
 2
 1
[torch.DoubleTensor of size 3]

[res] torch.cmin([res,] tensor, value)

Compute the minimum between each value in tensor and value.

c = torch.cmin(a, v) returns a new Tensor containing the minima of each element in a and v.

a:cmin(v) stores the minima of each element in a and v in a.

c:cmin(a, v) stores the minima of each element in a and v in c.

> a = torch.Tensor{1, 2, 3}
> torch.cmin(a, 2)
 1
 2
 2
[torch.DoubleTensor of size 3]

torch.median([resval, resind,] x [,dim])

y = torch.median(x) performs the median operation over the last dimension of x (one-before-middle in the case of an even number of elements).

y, i = torch.median(x, 1) returns the median element in each column (across rows) of x, and a Tensor i of their corresponding indices in x.

y, i = torch.median(x, 2) performs the median operation for each row.

y, i = torch.median(x, n) performs the median operation over the dimension n.

> x = torch.randn(3, 3)
> x
 0.7860  0.7687 -0.9362
 0.0411  0.5407 -0.3616
-0.0129 -0.2499 -0.5786
[torch.DoubleTensor of size 3x3]

> y, i = torch.median(x)
> y
 0.7687
 0.0411
-0.2499
[torch.DoubleTensor of size 3x1]

> i
 2
 1
 2
[torch.LongTensor of size 3x1]

> y, i = torch.median(x, 1)
> y
 0.0411  0.5407 -0.5786
[torch.DoubleTensor of size 1x3]

> i
 2  2  3
[torch.LongTensor of size 1x3]

> y, i = torch.median(x, 2)
> y
 0.7687
 0.0411
-0.2499
[torch.DoubleTensor of size 3x1]

> i
 2
 1
 2
[torch.LongTensor of size 3x1]

torch.mode([resval, resind,] x [,dim])

y = torch.mode(x) returns the most frequent element of x over its last dimension.

y, i = torch.mode(x, 1) returns the mode element in each column (across rows) of x, and a Tensor i of their corresponding indices in x.

y, i = torch.mode(x, 2) performs the mode operation for each row.

y, i = torch.mode(x, n) performs the mode operation over the dimension n.

torch.kthvalue([resval, resind,] x, k [,dim])

y = torch.kthvalue(x, k) returns the k-th smallest element of x over its last dimension.

y, i = torch.kthvalue(x, k, 1) returns the k-th smallest element in each column (across rows) of x, and a Tensor i of their corresponding indices in x.

y, i = torch.kthvalue(x, k, 2) performs the k-th value operation for each row.

y, i = torch.kthvalue(x, k, n) performs the k-th value operation over the dimension n.

[res] torch.prod([res,] x [,n])

y = torch.prod(x) returns the product of all elements in x.

y = torch.prod(x, n) returns a Tensor y whom size in dimension n is 1 and where elements are the product of elements of x with respect to dimension n.

> a = torch.Tensor{{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}
> a
(1,.,.) =
  1  2
  3  4

(2,.,.) =
  5  6
  7  8
[torch.DoubleTensor of dimension 2x2x2]

> torch.prod(a, 1)
(1,.,.) =
   5  12
  21  32
[torch.DoubleTensor of dimension 1x2x2]

> torch.prod(a, 2)
(1,.,.) =
   3   8

(2,.,.) =
  35  48
[torch.DoubleTensor of size 2x1x2]

> torch.prod(a, 3)
(1,.,.) =
   2
  12

(2,.,.) =
  30
  56
[torch.DoubleTensor of size 2x2x1]

torch.sort([resval, resind,] x [,d] [,flag])

y, i = torch.sort(x) returns a Tensor y where all entries are sorted along the last dimension, in ascending order. It also returns a Tensor i that provides the corresponding indices from x.

y, i = torch.sort(x, d) performs the sort operation along a specific dimension d.

y, i = torch.sort(x) is therefore equivalent to y, i = torch.sort(x, x:dim())

y, i = torch.sort(x, d, true) performs the sort operation along a specific dimension d, in descending order.

> x = torch.randn(3, 3)
> x
-1.2470 -0.4288 -0.5337
 0.8836 -0.1622  0.9604
 0.6297  0.2397  0.0746
[torch.DoubleTensor of size 3x3]

> torch.sort(x)
-1.2470 -0.5337 -0.4288
-0.1622  0.8836  0.9604
 0.0746  0.2397  0.6297
[torch.DoubleTensor of size 3x3]

 1  3  2
 2  1  3
 3  2  1
[torch.LongTensor of size 3x3]

torch.topk([resval, resind,] x, k, [,dim] [,dir] [,sort])

y, i = torch.topk(x, k) returns all k smallest elements in x over its last dimension including their indices, in unsorted order.

y, i = torch.topk(x, k, dim) performs the same operation except over dimension dim.

y, i = torch.topk(x, k, dim, dir) adds a sorting direction that has the same sense as torch.sort; false returns the k smallest elements in the slice, true returns the k largest elements in the slice.

y, i = torch.topk(x, k, dim, dir, true) specifies that the results in y should be sorted with respect to dir; by default, the results are potentially unsorted since the computation may be faster, but if sorting is desired, the sort flag may be passed, in which case the results are returned from smallest to k-th smallest (dir == false) or highest to k-th highest (dir == true).

The implementation provides no guarantee of the order of selection (indices) among equivalent elements (e.g., topk k == 2 selection of a vector {1, 2, 1, 1}; the values returned could be any pair of 1 entries in the vector).

[res] torch.std([res,] x, [,dim] [,flag])

y = torch.std(x) returns the standard deviation of the elements of x.

y = torch.std(x, dim) performs the std operation over the dimension dim.

y = torch.std(x, dim, false) performs the std operation normalizing by n-1 (this is the default).

y = torch.std(x, dim, true) performs the std operation normalizing by n instead of n-1.

[res] torch.sum([res,] x)

y = torch.sum(x) returns the sum of the elements of x.

y = torch.sum(x, 2) performs the sum operation for each row.

y = torch.sum(x, n) performs the sum operation over the dimension n.

[res] torch.var([res,] x [,dim] [,flag])

y = torch.var(x) returns the variance of the elements of x.

y = torch.var(x, dim) performs the var operation over the dimension dim.

y = torch.var(x, dim, false) performs the var operation normalizing by n-1 (this is the default).

y = torch.var(x, dim, true) performs the var operation normalizing by n instead of n-1.

Matrix-wide operations (Tensor-wide operations)

Note that many of the operations in dimension-wise operations can also be used as matrix-wide operations, by just omitting the dim parameter.

torch.norm(x [,p] [,dim])

y = torch.norm(x) returns the 2-norm of the Tensor x.

y = torch.norm(x, p) returns the p-norm of the Tensor x.

y = torch.norm(x, p, dim) returns the p-norms of the Tensor x computed over the dimension dim.

torch.renorm([res], x, p, dim, maxnorm)

Renormalizes the sub-Tensors along dimension dim such that they do not exceed norm maxnorm.

y = torch.renorm(x, p, dim, maxnorm) returns a version of x with p-norms lower than maxnorm over non-dim dimensions. The dim argument is not to be confused with the argument of the same name in function norm. In this case, the p-norm is measured for each i-th sub-Tensor x:select(dim, i). This function is equivalent to (but faster than) the following:

function renorm(matrix, value, dim, maxnorm)
   local m1 = matrix:transpose(dim, 1):contiguous()
   -- collapse non-dim dimensions:
   m2 = m1:reshape(m1:size(1), m1:nElement()/m1:size(1))
   local norms = m2:norm(value, 2)
   -- clip
   local new_norms = norms:clone()
   new_norms[torch.gt(norms, maxnorm)] = maxnorm
   new_norms:cdiv(norms:add(1e-7))
   -- renormalize
   m1:cmul(new_norms:expandAs(m1))
   return m1:transpose(dim, 1)
end

x:renorm(p, dim, maxnorm) returns the equivalent of x:copy(torch.renorm(x, p, dim, maxnorm)).

Note: this function is particularly useful as a regularizer for constraining the norm of parameter Tensors. See Hinton et al. 2012, p. 2.

torch.dist(x, y)

y = torch.dist(x, y) returns the 2-norm of x - y.

y = torch.dist(x, y, p) returns the p-norm of x - y.

torch.numel(x)

y = torch.numel(x) returns the count of the number of elements in the matrix x.

torch.trace(x)

y = torch.trace(x) returns the trace (sum of the diagonal elements) of a matrix x. This is equal to the sum of the eigenvalues of x. The returned value y is a number, not a Tensor.

Convolution Operations

These functions implement convolution or cross-correlation of an input image (or set of input images) with a kernel (or set of kernels). The convolution function in Torch can handle different types of input/kernel dimensions and produces corresponding outputs. The general form of operations always remain the same.

[res] torch.conv2([res,] x, k, [, 'F' or 'V'])

This function computes 2 dimensional convolutions between x and k. These operations are similar to BLAS operations when number of dimensions of input and kernel are reduced by 2.

  • x and k are 2D: convolution of a single image with a single kernel (2D output). This operation is similar to multiplication of two scalars.
  • x (p × m × n) and k (p × ki × kj) are 3D: convolution of each input slice with corresponding kernel (3D output).
  • x (p × m × n) 3D, k (q × p × ki × kj) 4D: convolution of all input slices with the corresponding slice of kernel. Output is 3D (q × m × n). This operation is similar to matrix vector product of matrix k and vector x.

The last argument controls if the convolution is a full ('F') or valid ('V') convolution. The default is valid convolution.

x = torch.rand(100, 100)
k = torch.rand(10, 10)
c = torch.conv2(x, k)
> c:size()
 91
 91
[torch.LongStorage of size 2]

c = torch.conv2(x, k, 'F')
> c:size()
 109
 109
[torch.LongStorage of size 2]

[res] torch.xcorr2([res,] x, k, [, 'F' or 'V'])

This function operates with same options and input/output configurations as torch.conv2, but performs cross-correlation of the input with the kernel k.

[res] torch.conv3([res,] x, k, [, 'F' or 'V'])

This function computes 3 dimensional convolutions between x and k. These operations are similar to BLAS operations when number of dimensions of input and kernel are reduced by 3.

  • x and k are 3D: convolution of a single image with a single kernel (3D output). This operation is similar to multiplication of two scalars.
  • x (p × m × n × o) and k (p × ki × kj × kk) are 4D: convolution of each input slice with corresponding kernel (4D output).
  • x (p × m × n × o) 4D, k (q × p × ki × kj × kk) 5D: convolution of all input slices with the corresponding slice of kernel. Output is 4D q × m × n × o. This operation is similar to matrix vector product of matrix k and vector x.

The last argument controls if the convolution is a full ('F') or valid ('V') convolution. The default is valid convolution.

x = torch.rand(100, 100, 100)
k = torch.rand(10, 10, 10)
c = torch.conv3(x, k)
> c:size()
 91
 91
 91
[torch.LongStorage of size 3]

c = torch.conv3(x, k, 'F')
> c:size()
 109
 109
 109
[torch.LongStorage of size 3]

[res] torch.xcorr3([res,] x, k, [, 'F' or 'V'])

This function operates with same options and input/output configurations as torch.conv3, but performs cross-correlation of the input with the kernel k.

Eigenvalues, SVD, Linear System Solution

Functions in this section are implemented with an interface to LAPACK libraries. If LAPACK libraries are not found during compilation step, then these functions will not be available.

[x, lu] torch.gesv([resb, resa,] B, A)

X, LU = torch.gesv(B, A) returns the solution of AX = B and LU contains L and U factors for LU factorization of A.

A has to be a square and non-singular matrix (2D Tensor). A and LU are m × m, X is m × k and B is m × k.

If resb and resa are given, then they will be used for temporary storage and returning the result.

  • resa will contain L and U factors for LU factorization of A.
  • resb will contain the solution X.

Note: Irrespective of the original strides, the returned matrices resb and resa will be transposed, i.e. with strides 1, m instead of m, 1.

> a = torch.Tensor({{6.80, -2.11,  5.66,  5.97,  8.23},
                  {-6.05, -3.30,  5.36, -4.44,  1.08},
                  {-0.45,  2.58, -2.70,  0.27,  9.04},
                  {8.32,  2.71,  4.35,  -7.17,  2.14},
                  {-9.67, -5.14, -7.26,  6.08, -6.87}}):t()

> b = torch.Tensor({{4.02,  6.19, -8.22, -7.57, -3.03},
                  {-1.56,  4.00, -8.67,  1.75,  2.86},
                  {9.81, -4.09, -4.57, -8.61,  8.99}}):t()

> b
 4.0200 -1.5600  9.8100
 6.1900  4.0000 -4.0900
-8.2200 -8.6700 -4.5700
-7.5700  1.7500 -8.6100
-3.0300  2.8600  8.9900
[torch.DoubleTensor of dimension 5x3]

> a
 6.8000 -6.0500 -0.4500  8.3200 -9.6700
-2.1100 -3.3000  2.5800  2.7100 -5.1400
 5.6600  5.3600 -2.7000  4.3500 -7.2600
 5.9700 -4.4400  0.2700 -7.1700  6.0800
 8.2300  1.0800  9.0400  2.1400 -6.8700
[torch.DoubleTensor of dimension 5x5]


> x = torch.gesv(b, a)
> x
-0.8007 -0.3896  0.9555
-0.6952 -0.5544  0.2207
 0.5939  0.8422  1.9006
 1.3217 -0.1038  5.3577
 0.5658  0.1057  4.0406
[torch.DoubleTensor of dimension 5x3]

> b:dist(a * x)
1.1682163181673e-14

[x] torch.trtrs([resb, resa,] b, a [, 'U' or 'L'] [, 'N' or 'T'] [, 'N' or 'U'])

X = torch.trtrs(B, A) returns the solution of AX = B where A is upper-triangular.

A has to be a square, triangular, non-singular matrix (2D Tensor). A and resa are m × m, X and B are m × k. (To be very precise: A does not have to be triangular and non-singular, rather only its upper or lower triangle will be taken into account and that part has to be non-singular.)

The function has several options:

  • uplo ('U' or 'L') specifies whether A is upper or lower triangular; the default value is 'U'.
  • trans ('N' or 'T') specifies the system of equations: 'N' for A * X = B (no transpose), or 'T' for A^T * X = B (transpose); the default value is 'N'.
  • diag ('N' or 'U') 'U' specifies that A is unit triangular, i.e., it has ones on its diagonal; 'N' specifies that A is not (necessarily) unit triangular; the default value is 'N'.

If resb and resa are given, then they will be used for temporary storage and returning the result. resb will contain the solution X.

Note: Irrespective of the original strides, the returned matrices resb and resa will be transposed, i.e. with strides 1, m instead of m, 1.

> a = torch.Tensor({{6.80, -2.11,  5.66,  5.97,  8.23},
                  {0, -3.30,  5.36, -4.44,  1.08},
                  {0,  0, -2.70,  0.27,  9.04},
                  {0,  0,  0,  -7.17,  2.14},
                  {0,  0,  0,  0, -6.87}})

> b = torch.Tensor({{4.02,  6.19, -8.22, -7.57, -3.03},
                  {-1.56,  4.00, -8.67,  1.75,  2.86},
                  {9.81, -4.09, -4.57, -8.61,  8.99}}):t()

> b
 4.0200 -1.5600  9.8100
 6.1900  4.0000 -4.0900
-8.2200 -8.6700 -4.5700
-7.5700  1.7500 -8.6100
-3.0300  2.8600  8.9900
[torch.DoubleTensor of dimension 5x3]

> a
 6.8000 -2.1100  5.6600  5.9700  8.2300
 0.0000 -3.3000  5.3600 -4.4400  1.0800
 0.0000  0.0000 -2.7000  0.2700  9.0400
 0.0000  0.0000  0.0000 -7.1700  2.1400
 0.0000  0.0000  0.0000  0.0000 -6.8700
[torch.DoubleTensor of dimension 5x5]

> x = torch.trtrs(b, a)
> x
-3.5416 -0.2514  3.0847
 4.2072  2.0391 -4.5146
 4.6399  1.7804 -2.6077
 1.1874 -0.3683  0.8103
 0.4410 -0.4163 -1.3086
[torch.DoubleTensor of size 5x3]

> b:dist(a*x)
4.1895292266754e-15

torch.potrf([res,] A [, 'U' or 'L'] )

Cholesky Decomposition of 2D Tensor A. The matrix A has to be a positive-definite and either symmetric or complex Hermitian.

The factorization has the form

 A = U**T * U,   if UPLO = 'U', or
 A = L  * L**T,  if UPLO = 'L',

where U is an upper triangular matrix and L is lower triangular.

The optional character uplo = {'U', 'L'} specifies whether the upper or lower triangulardecomposition should be returned. By default, uplo = 'U'.

U = torch.potrf(A, 'U') returns the upper triangular Cholesky decomposition of A.

L = torch.potrf(A, 'L') returns the lower triangular Cholesky decomposition of A.

If Tensor res is provided, the resulting decomposition will be stored therein.

> A = torch.Tensor({
    {1.2705,  0.9971,  0.4948,  0.1389,  0.2381},
    {0.9971,  0.9966,  0.6752,  0.0686,  0.1196},
    {0.4948,  0.6752,  1.1434,  0.0314,  0.0582},
    {0.1389,  0.0686,  0.0314,  0.0270,  0.0526},
    {0.2381,  0.1196,  0.0582,  0.0526,  0.3957}})

> chol = torch.potrf(A)
> chol
 1.1272  0.8846  0.4390  0.1232  0.2112
 0.0000  0.4626  0.6200 -0.0874 -0.1453
 0.0000  0.0000  0.7525  0.0419  0.0738
 0.0000  0.0000  0.0000  0.0491  0.2199
 0.0000  0.0000  0.0000  0.0000  0.5255
[torch.DoubleTensor of size 5x5]

> torch.potrf(chol, A, 'L')
> chol
 1.1272  0.0000  0.0000  0.0000  0.0000
 0.8846  0.4626  0.0000  0.0000  0.0000
 0.4390  0.6200  0.7525  0.0000  0.0000
 0.1232 -0.0874  0.0419  0.0491  0.0000
 0.2112 -0.1453  0.0738  0.2199  0.5255
[torch.DoubleTensor of size 5x5]

torch.pstrf([res, piv, ] A [, 'U' or 'L'] )

Cholesky factorization with complete pivoting of a real symmetric positive semidefinite 2D Tensor A. The matrix A has to be a positive semi-definite and symmetric. The factorization has the form

P**T * A * P = U**T * U ,  if UPLO = 'U',
P**T * A * P = L  * L**T,  if UPLO = 'L',

where U is an upper triangular matrix and L is lower triangular, and P is stored as the vector piv. More specifically, piv is such that the nonzero entries are P[piv[k], k] = 1.

The optional character argument uplo = {'U', 'L'} specifies whether the upper or lower triangular decomposition should be returned. By default, uplo = 'U'.

U, piv = torch.sdtrf(A, 'U') returns the upper triangular Cholesky decomposition of A

L, piv = torch.potrf(A, 'L') returns the lower triangular Cholesky decomposition of A.

If tensors res and piv (an IntTensor) are provided, the resulting decomposition will be stored therein.

> A = torch.Tensor({
    {1.2705,  0.9971,  0.4948,  0.1389,  0.2381},
    {0.9971,  0.9966,  0.6752,  0.0686,  0.1196},
    {0.4948,  0.6752,  1.1434,  0.0314,  0.0582},
    {0.1389,  0.0686,  0.0314,  0.0270,  0.0526},
    {0.2381,  0.1196,  0.0582,  0.0526,  0.3957}})

> U, piv = torch.pstrf(A)
> U
 1.1272  0.4390  0.2112  0.8846  0.1232
 0.0000  0.9750 -0.0354  0.2942 -0.0233
 0.0000  0.0000  0.5915 -0.0961  0.0435
 0.0000  0.0000  0.0000  0.3439 -0.0854
 0.0000  0.0000  0.0000  0.0000  0.0456
[torch.DoubleTensor of size 5x5]

> piv
 1
 3
 5
 2
 4
[torch.IntTensor of size 5]

> Ap = U:t() * U
> Ap
 1.2705  0.4948  0.2381  0.9971  0.1389
 0.4948  1.1434  0.0582  0.6752  0.0314
 0.2381  0.0582  0.3957  0.1196  0.0526
 0.9971  0.6752  0.1196  0.9966  0.0686
 0.1389  0.0314  0.0526  0.0686  0.0270
[torch.DoubleTensor of size 5x5]

> -- Permute rows and columns
> Ap:indexCopy(1, piv:long(), Ap:clone())
> Ap:indexCopy(2, piv:long(), Ap:clone())
> (Ap - A):norm()
1.5731560566382e-16

torch.potrs([res,] B, chol [, 'U' or 'L'] )

Returns the solution to linear system AX = B using the Cholesky decomposition chol of 2D Tensor A.

Square matrix chol should be triangular; and, righthand side matrix B should be of full rank.

Optional character uplo = {'U', 'L'} specifies matrix chol as either upper or lower triangular; and, by default, equals 'U'.

If Tensor res is provided, the resulting decomposition will be stored therein.

> A = torch.Tensor({
    {1.2705,  0.9971,  0.4948,  0.1389,  0.2381},
    {0.9971,  0.9966,  0.6752,  0.0686,  0.1196},
    {0.4948,  0.6752,  1.1434,  0.0314,  0.0582},
    {0.1389,  0.0686,  0.0314,  0.0270,  0.0526},
    {0.2381,  0.1196,  0.0582,  0.0526,  0.3957}})

> B = torch.Tensor({
    {0.6219,  0.3439,  0.0431},
    {0.5642,  0.1756,  0.0153},
    {0.2334,  0.8594,  0.4103},
    {0.7556,  0.1966,  0.9637},
    {0.1420,  0.7185,  0.7476}})

> chol = torch.potrf(A)
> chol
 1.1272  0.8846  0.4390  0.1232  0.2112
 0.0000  0.4626  0.6200 -0.0874 -0.1453
 0.0000  0.0000  0.7525  0.0419  0.0738
 0.0000  0.0000  0.0000  0.0491  0.2199
 0.0000  0.0000  0.0000  0.0000  0.5255
[torch.DoubleTensor of size 5x5]

> solve = torch.potrs(B, chol)
> solve
  12.1945   61.8622   92.6882
 -11.1782  -97.0303 -138.4874
 -15.3442  -76.6562 -116.8218
   6.1930   13.5238   25.2056
  29.9678  251.7346  360.2301
[torch.DoubleTensor of size 5x3]

> A*solve
 0.6219  0.3439  0.0431
 0.5642  0.1756  0.0153
 0.2334  0.8594  0.4103
 0.7556  0.1966  0.9637
 0.1420  0.7185  0.7476
[torch.DoubleTensor of size 5x3]

> B:dist(A*solve)
4.6783066076306e-14

torch.potri([res,] chol [, 'U' or 'L'] )

Returns the inverse of 2D Tensor A given its Cholesky decomposition chol.

Square matrix chol should be triangular.

Optional character uplo = {'U', 'L'} specifies matrix chol as either upper or lower triangular; and, by default, equals 'U'.

If Tensor res is provided, the resulting inverse will be stored therein.

> A = torch.Tensor({
    {1.2705,  0.9971,  0.4948,  0.1389,  0.2381},
    {0.9971,  0.9966,  0.6752,  0.0686,  0.1196},
    {0.4948,  0.6752,  1.1434,  0.0314,  0.0582},
    {0.1389,  0.0686,  0.0314,  0.0270,  0.0526},
    {0.2381,  0.1196,  0.0582,  0.0526,  0.3957}})

> chol = torch.potrf(A)
> chol
 1.1272  0.8846  0.4390  0.1232  0.2112
 0.0000  0.4626  0.6200 -0.0874 -0.1453
 0.0000  0.0000  0.7525  0.0419  0.0738
 0.0000  0.0000  0.0000  0.0491  0.2199
 0.0000  0.0000  0.0000  0.0000  0.5255
[torch.DoubleTensor of size 5x5]

> inv = torch.potri(chol)
> inv
  42.2781  -39.0824    8.3019 -133.4998    2.8980
 -39.0824   38.1222   -8.7468  119.4247   -2.5944
   8.3019   -8.7468    3.1104  -25.1405    0.5327
-133.4998  119.4247  -25.1405  480.7511  -15.9747
   2.8980   -2.5944    0.5327  -15.9747    3.6127
[torch.DoubleTensor of size 5x5]

> inv:dist(torch.inverse(A))
2.8525852877633e-12

torch.gels([resb, resa,] b, a)

Solution of least squares and least norm problems for a full rank m × n matrix A.

  • If n ≤ m, then solve ||AX-B||_F.
  • If n > m , then solve min ||X||_F s.t. AX = B.

On return, first n rows of x matrix contains the solution and the rest contains residual information. Square root of sum squares of elements of each column of x starting at row n + 1 is the residual for corresponding column.

Note: Irrespective of the original strides, the returned matrices resb and resa will be transposed, i.e. with strides 1, m instead of m, 1.

> a = torch.Tensor({{ 1.44, -9.96, -7.55,  8.34,  7.08, -5.45},
                  {-7.84, -0.28,  3.24,  8.09,  2.52, -5.70},
                  {-4.39, -3.24,  6.27,  5.28,  0.74, -1.19},
                  {4.53,  3.83, -6.64,  2.06, -2.47,  4.70}}):t()

> b = torch.Tensor({{8.58,  8.26,  8.48, -5.28,  5.72,  8.93},
                  {9.35, -4.43, -0.70, -0.26, -7.36, -2.52}}):t()

> a
 1.4400 -7.8400 -4.3900  4.5300
-9.9600 -0.2800 -3.2400  3.8300
-7.5500  3.2400  6.2700 -6.6400
 8.3400  8.0900  5.2800  2.0600
 7.0800  2.5200  0.7400 -2.4700
-5.4500 -5.7000 -1.1900  4.7000
[torch.DoubleTensor of dimension 6x4]

> b
 8.5800  9.3500
 8.2600 -4.4300
 8.4800 -0.7000
-5.2800 -0.2600
 5.7200 -7.3600
 8.9300 -2.5200
[torch.DoubleTensor of dimension 6x2]

> x = torch.gels(b, a)
> x
 -0.4506   0.2497
 -0.8492  -0.9020
  0.7066   0.6323
  0.1289   0.1351
 13.1193  -7.4922
 -4.8214  -7.1361
[torch.DoubleTensor of dimension 6x2]

> b:dist(a*x:narrow(1, 1, 4))
17.390200628863

> math.sqrt(x:narrow(1, 5, 2):pow(2):sumall())
17.390200628863

torch.symeig([rese, resv,] a [, 'N' or 'V'] [, 'U' or 'L'])

e, V = torch.symeig(A) returns eigenvalues and eigenvectors of a symmetric real matrix A.

A and V are m × m matrices and e is a m dimensional vector.

This function calculates all eigenvalues (and vectors) of A such that A = V diag(e) V'.

Third argument defines computation of eigenvectors or eigenvalues only. If it is 'N', only eigenvalues are computed. If it is 'V', both eigenvalues and eigenvectors are computed.

Since the input matrix A is supposed to be symmetric, only upper triangular portion is used by default. If the 4th argument is 'L', then lower triangular portion is used.

Note: Irrespective of the original strides, the returned matrix V will be transposed, i.e. with strides 1, m instead of m, 1.

> a = torch.Tensor({{ 1.96,  0.00,  0.00,  0.00,  0.00},
                  {-6.49,  3.80,  0.00,  0.00,  0.00},
                  {-0.47, -6.39,  4.17,  0.00,  0.00},
                  {-7.20,  1.50, -1.51,  5.70,  0.00},
                  {-0.65, -6.34,  2.67,  1.80, -7.10}}):t()

> a
 1.9600 -6.4900 -0.4700 -7.2000 -0.6500
 0.0000  3.8000 -6.3900  1.5000 -6.3400
 0.0000  0.0000  4.1700 -1.5100  2.6700
 0.0000  0.0000  0.0000  5.7000  1.8000
 0.0000  0.0000  0.0000  0.0000 -7.1000
[torch.DoubleTensor of dimension 5x5]

> e = torch.symeig(a)
> e
-11.0656
 -6.2287
  0.8640
  8.8655
 16.0948
[torch.DoubleTensor of dimension 5]

> e, v = torch.symeig(a, 'V')
> e
-11.0656
 -6.2287
  0.8640
  8.8655
 16.0948
[torch.DoubleTensor of dimension 5]

> v
-0.2981 -0.6075  0.4026 -0.3745  0.4896
-0.5078 -0.2880 -0.4066 -0.3572 -0.6053
-0.0816 -0.3843 -0.6600  0.5008  0.3991
-0.0036 -0.4467  0.4553  0.6204 -0.4564
-0.8041  0.4480  0.1725  0.3108  0.1622
[torch.DoubleTensor of dimension 5x5]

> v*torch.diag(e)*v:t()
 1.9600 -6.4900 -0.4700 -7.2000 -0.6500
-6.4900  3.8000 -6.3900  1.5000 -6.3400
-0.4700 -6.3900  4.1700 -1.5100  2.6700
-7.2000  1.5000 -1.5100  5.7000  1.8000
-0.6500 -6.3400  2.6700  1.8000 -7.1000
[torch.DoubleTensor of dimension 5x5]

> a:dist(torch.triu(v*torch.diag(e)*v:t()))
1.0219480822443e-14

torch.eig([rese, resv,] a [, 'N' or 'V'])

e, V = torch.eig(A) returns eigenvalues and eigenvectors of a general real square matrix A.

A and V are m × m matrices and e is a m dimensional vector.

This function calculates all right eigenvalues (and vectors) of A such that A = V diag(e) V'.

Third argument defines computation of eigenvectors or eigenvalues only. If it is 'N', only eigenvalues are computed. If it is 'V', both eigenvalues and eigenvectors are computed.

The eigen values returned follow LAPACK convention and are returned as complex (real/imaginary) pairs of numbers (2 * m dimensional Tensor).

Note: Irrespective of the original strides, the returned matrix V will be transposed, i.e. with strides 1, m instead of m, 1.

> a = torch.Tensor({{ 1.96,  0.00,  0.00,  0.00,  0.00},
                  {-6.49,  3.80,  0.00,  0.00,  0.00},
                  {-0.47, -6.39,  4.17,  0.00,  0.00},
                  {-7.20,  1.50, -1.51,  5.70,  0.00},
                  {-0.65, -6.34,  2.67,  1.80, -7.10}}):t()

> a
 1.9600 -6.4900 -0.4700 -7.2000 -0.6500
 0.0000  3.8000 -6.3900  1.5000 -6.3400
 0.0000  0.0000  4.1700 -1.5100  2.6700
 0.0000  0.0000  0.0000  5.7000  1.8000
 0.0000  0.0000  0.0000  0.0000 -7.1000
[torch.DoubleTensor of dimension 5x5]

> b = a + torch.triu(a, 1):t()
> b

  1.9600 -6.4900 -0.4700 -7.2000 -0.6500
 -6.4900  3.8000 -6.3900  1.5000 -6.3400
 -0.4700 -6.3900  4.1700 -1.5100  2.6700
 -7.2000  1.5000 -1.5100  5.7000  1.8000
 -0.6500 -6.3400  2.6700  1.8000 -7.1000
[torch.DoubleTensor of dimension 5x5]

> e = torch.eig(b)
> e
 16.0948   0.0000
-11.0656   0.0000
 -6.2287   0.0000
  0.8640   0.0000
  8.8655   0.0000
[torch.DoubleTensor of dimension 5x2]

> e, v = torch.eig(b, 'V')
> e
 16.0948   0.0000
-11.0656   0.0000
 -6.2287   0.0000
  0.8640   0.0000
  8.8655   0.0000
[torch.DoubleTensor of dimension 5x2]

> v
-0.4896  0.2981 -0.6075 -0.4026 -0.3745
 0.6053  0.5078 -0.2880  0.4066 -0.3572
-0.3991  0.0816 -0.3843  0.6600  0.5008
 0.4564  0.0036 -0.4467 -0.4553  0.6204
-0.1622  0.8041  0.4480 -0.1725  0.3108
[torch.DoubleTensor of dimension 5x5]

> v * torch.diag(e:select(2, 1))*v:t()
 1.9600 -6.4900 -0.4700 -7.2000 -0.6500
-6.4900  3.8000 -6.3900  1.5000 -6.3400
-0.4700 -6.3900  4.1700 -1.5100  2.6700
-7.2000  1.5000 -1.5100  5.7000  1.8000
-0.6500 -6.3400  2.6700  1.8000 -7.1000
[torch.DoubleTensor of dimension 5x5]

> b:dist(v * torch.diag(e:select(2, 1)) * v:t())
3.5423944346685e-14

torch.svd([resu, ress, resv,] a [, 'S' or 'A'])

U, S, V = torch.svd(A) returns the singular value decomposition of a real matrix A of size n × m such that A = USV'*.

U is n × n, S is n × m and V is m × m.

The last argument, if it is string, represents the number of singular values to be computed. 'S' stands for some and 'A' stands for all.

Note: Irrespective of the original strides, the returned matrix U will be transposed, i.e. with strides 1, n instead of n, 1.

> a = torch.Tensor({{8.79,  6.11, -9.15,  9.57, -3.49,  9.84},
                  {9.93,  6.91, -7.93,  1.64,  4.02,  0.15},
                  {9.83,  5.04,  4.86,  8.83,  9.80, -8.99},
                  {5.45, -0.27,  4.85,  0.74, 10.00, -6.02},
                  {3.16,  7.98,  3.01,  5.80,  4.27, -5.31}}):t()

> a
  8.7900   9.9300   9.8300   5.4500   3.1600
  6.1100   6.9100   5.0400  -0.2700   7.9800
 -9.1500  -7.9300   4.8600   4.8500   3.0100
  9.5700   1.6400   8.8300   0.7400   5.8000
 -3.4900   4.0200   9.8000  10.0000   4.2700
  9.8400   0.1500  -8.9900  -6.0200  -5.3100

> u, s, v = torch.svd(a)
> u
-0.5911  0.2632  0.3554  0.3143  0.2299
-0.3976  0.2438 -0.2224 -0.7535 -0.3636
-0.0335 -0.6003 -0.4508  0.2334 -0.3055
-0.4297  0.2362 -0.6859  0.3319  0.1649
-0.4697 -0.3509  0.3874  0.1587 -0.5183
 0.2934  0.5763 -0.0209  0.3791 -0.6526
[torch.DoubleTensor of dimension 6x5]

> s
 27.4687
 22.6432
  8.5584
  5.9857
  2.0149
[torch.DoubleTensor of dimension 5]

> v
-0.2514  0.8148 -0.2606  0.3967 -0.2180
-0.3968  0.3587  0.7008 -0.4507  0.1402
-0.6922 -0.2489 -0.2208  0.2513  0.5891
-0.3662 -0.3686  0.3859  0.4342 -0.6265
-0.4076 -0.0980 -0.4933 -0.6227 -0.4396
[torch.DoubleTensor of dimension 5x5]

> u * torch.diag(s) * v:t()
  8.7900   9.9300   9.8300   5.4500   3.1600
  6.1100   6.9100   5.0400  -0.2700   7.9800
 -9.1500  -7.9300   4.8600   4.8500   3.0100
  9.5700   1.6400   8.8300   0.7400   5.8000
 -3.4900   4.0200   9.8000  10.0000   4.2700
  9.8400   0.1500  -8.9900  -6.0200  -5.3100
[torch.DoubleTensor of dimension 6x5]

> a:dist(u * torch.diag(s) * v:t())
2.8923773593204e-14

torch.inverse([res,] x)

Computes the inverse of square matrix x.

torch.inverse(x) returns the result as a new matrix.

torch.inverse(y, x) puts the result in y.

Note: Irrespective of the original strides, the returned matrix y will be transposed, i.e. with strides 1, m instead of m, 1.

> x = torch.rand(10, 10)
> y = torch.inverse(x)
> z = x * y
> z
 1.0000 -0.0000  0.0000 -0.0000  0.0000  0.0000  0.0000 -0.0000  0.0000  0.0000
 0.0000  1.0000 -0.0000 -0.0000  0.0000  0.0000 -0.0000 -0.0000 -0.0000  0.0000
 0.0000 -0.0000  1.0000 -0.0000  0.0000  0.0000 -0.0000 -0.0000  0.0000  0.0000
 0.0000 -0.0000 -0.0000  1.0000 -0.0000  0.0000  0.0000 -0.0000 -0.0000  0.0000
 0.0000 -0.0000  0.0000 -0.0000  1.0000  0.0000  0.0000 -0.0000 -0.0000  0.0000
 0.0000 -0.0000  0.0000 -0.0000  0.0000  1.0000  0.0000 -0.0000 -0.0000  0.0000
 0.0000 -0.0000  0.0000 -0.0000  0.0000  0.0000  1.0000 -0.0000  0.0000  0.0000
 0.0000 -0.0000 -0.0000 -0.0000  0.0000  0.0000  0.0000  1.0000  0.0000  0.0000
 0.0000 -0.0000 -0.0000 -0.0000  0.0000  0.0000 -0.0000 -0.0000  1.0000  0.0000
 0.0000 -0.0000  0.0000 -0.0000  0.0000  0.0000  0.0000 -0.0000  0.0000  1.0000
[torch.DoubleTensor of dimension 10x10]

> torch.max(torch.abs(z - torch.eye(10))) -- Max nonzero
2.3092638912203e-14

torch.qr([q, r], x)

Compute a QR decomposition of the matrix x: matrices q and r such that x = q * r, with q orthogonal and r upper triangular. This returns the thin (reduced) QR factorization.

torch.qr(x) returns the Q and R components as new matrices.

torch.qr(q, r, x) stores them in existing Tensors q and r.

Note that precision may be lost if the magnitudes of the elements of x are large.

Note also that, while it should always give you a valid decomposition, it may not give you the same one across platforms - it will depend on your LAPACK implementation.

Note: Irrespective of the original strides, the returned matrix q will be transposed, i.e. with strides 1, m instead of m, 1.

> a = torch.Tensor{{12, -51, 4}, {6, 167, -68}, {-4, 24, -41}}
> a
  12  -51    4
   6  167  -68
  -4   24  -41
[torch.DoubleTensor of dimension 3x3]

> q, r = torch.qr(a)
> q
-0.8571  0.3943  0.3314
-0.4286 -0.9029 -0.0343
 0.2857 -0.1714  0.9429
[torch.DoubleTensor of dimension 3x3]

> r
 -14.0000  -21.0000   14.0000
   0.0000 -175.0000   70.0000
   0.0000    0.0000  -35.0000
[torch.DoubleTensor of dimension 3x3]

> (q * r):round()
  12  -51    4
   6  167  -68
  -4   24  -41
[torch.DoubleTensor of dimension 3x3]

> (q:t() * q):round()
 1  0  0
 0  1  0
 0  0  1
[torch.DoubleTensor of dimension 3x3]

torch.geqrf([m, tau], a)

This is a low-level function for calling LAPACK directly. You'll generally want to use torch.qr() instead.

Computes a QR decomposition of a, but without constructing Q and R as explicit separate matrices. Rather, this directly calls the underlying LAPACK function ?geqrf which produces a sequence of 'elementary reflectors'. See LAPACK documentation for further details.

torch.orgqr([q], m, tau)

This is a low-level function for calling LAPACK directly. You'll generally want to use torch.qr() instead.

Constructs a Q matrix from a sequence of elementary reflectors, such as that given by torch.geqrf. See LAPACK documentation for further details.

torch.ormqr([res], m, tau, mat [, 'L' or 'R'] [, 'N' or 'T'])

Multiply a matrix with Q as defined by the elementary reflectors and scalar factors returned by geqrf. This is a low-level function for calling LAPACK directly. You'll generally want to use torch.qr() instead.

  • side ('L' or 'R') specifies whether mat should be left-multiplied, mat * Q, or right-multiplied, Q * mat.
  • trans ('N' or 'T') specifies whether Q should be transposed before being multiplied.

See LAPACK documentation for further details.

Logical Operations on Tensors

These functions implement logical comparison operators that take a Tensor as input and another Tensor or a number as the comparison target. They return a ByteTensor in which each element is 0 or 1 indicating if the comparison for the corresponding element was false or true respectively.

torch.lt(a, b)

Implements < operator comparing each element in a with b (if b is a number) or each element in a with corresponding element in b.

torch.le(a, b)

Implements <= operator comparing each element in a with b (if b is a number) or each element in a with corresponding element in b.

torch.gt(a, b)

Implements > operator comparing each element in a with b (if b is a number) or each element in a with corresponding element in b.

torch.ge(a, b)

Implements >= operator comparing each element in a with b (if b is a number) or each element in a with corresponding element in b.

torch.eq(a, b)

Implements == operator comparing each element in a with b (if b is a number) or each element in a with corresponding element in b.

torch.ne(a, b)

Implements ~= operator comparing each element in a with b (if b is a number) or each element in a with corresponding element in b.

torch.all(a)

torch.any(a)

Additionally, any and all logically sum a ByteTensor returning true if any or all elements are logically true respectively. Note that logically true here is meant in the C sense (zero is false, non-zero is true) such as the output of the Tensor element-wise logical operations.

> a = torch.rand(10)
> b = torch.rand(10)
> a
 0.5694
 0.5264
 0.3041
 0.4159
 0.1677
 0.7964
 0.0257
 0.2093
 0.6564
 0.0740
[torch.DoubleTensor of dimension 10]

> b
 0.2950
 0.4867
 0.9133
 0.1291
 0.1811
 0.3921
 0.7750
 0.3259
 0.2263
 0.1737
[torch.DoubleTensor of dimension 10]

> torch.lt(a, b)
 0
 0
 1
 0
 1
 0
 1
 1
 0
 1
[torch.ByteTensor of dimension 10]

> torch.eq(a, b)
0
0
0
0
0
0
0
0
0
0
[torch.ByteTensor of dimension 10]

> torch.ne(a, b)
 1
 1
 1
 1
 1
 1
 1
 1
 1
 1
[torch.ByteTensor of dimension 10]

> torch.gt(a, b)
 1
 1
 0
 1
 0
 1
 0
 0
 1
 0
[torch.ByteTensor of dimension 10]

> a[torch.gt(a, b)] = 10
> a
 10.0000
 10.0000
  0.3041
 10.0000
  0.1677
 10.0000
  0.0257
  0.2093
 10.0000
  0.0740
[torch.DoubleTensor of dimension 10]

> a[torch.gt(a, 1)] = -1
> a
-1.0000
-1.0000
 0.3041
-1.0000
 0.1677
-1.0000
 0.0257
 0.2093
-1.0000
 0.0740
[torch.DoubleTensor of dimension 10]

> a = torch.ones(3):byte()
> torch.all(a)
true

> a[2] = 0
> torch.all(a)
false

> torch.any(a)
true

> a:zero()
> torch.any(a)
false