Simple Matrix

Simple Matrix is a lightweight matrix operations shard.

• It provides matrix calculations:
  • Sum
  • Dot Product
  • Element-wise Multiplication
  • Apply functions
  • Convolution
  • Transpose
• It uses methods sum, dot, etc. with three (instead of two) arguments: source, target matrices and a result matrix like a buffer
• With previous, SimpleMatrix is mostly looks like part of a functional language in Crystal syntax
• The engine does not have any matrix dimensions checking - this is on the developer's conscience
• It does not have internal mallocs because of the specific logic:
  1. Take one matrix
  2. Make an operation with this and other matrix
  3. Put result into third matrix

The module is specifically designed to have increased performance reserves. It can be used when designing the interaction of layers within neural network engines.

Please read the notes on operations below.

Installation

1. Add the dependency to your shard.yml:

   dependencies:
     simple_matrix:
       github: fruktorum/simple_matrix
   

2. Run shards install

Usage

require "simple_matrix"

For all calculations described below please pay attention for the:

• Matrices types must match

• Almost all operations can be produced with 2D arrays instead of matrices:

  matrix1.dot matrix2, result_matrix
  
  # same as
  matrix1.dot matrix2.buffer, result_matrix.buffer
  
  # same as
  dot_array = Array.new( matrix1.width ){ Array.new( 5 ){ rand } }
  result_array = Array.new( matrix1.height ){ Array.new( dot_array.first.size ){ 0 } }
  matrix1.dot dot_array, result_array
  

• With buffers operations (arrays instead of matrices) they types must match

Initialization

# height = 4, width = 3 (filled by 0)
matrix = SimpleMatrix( UInt32 ).new 4, 3

# type = UInt8, height = 4, width = 3, values = 99
matrix = SimpleMatrix( UInt8 ).new 4, 3, 99

# Custom values:
# 0  1  2
# 3  4  5
# 6  7  8
# 9 10 11
width, height = 3, 4
matrix = SimpleMatrix( UInt16 ).new( height, width ){ |y, x| ( y * width + x ).to_u16 }

# Matrix has public method to access internal buffer:
matrix.buffer # => Array( Array( UInt16 ) )

To initialize identity (in example: 5x5) matrix (zero-filled matrix which main diagonal filled by 1):

matrix = SimpleMatrix( UInt8 ).identity 5

┌ 1 0 0 0 0 ┐
│ 0 1 0 0 0 │
│ 0 0 1 0 0 │
│ 0 0 0 1 0 │
└ 0 0 0 0 1 ┘

Output

puts matrix # prints to stdout in pretty form

Operations

For the below assume that matrices are declared as m1, m2, etc. Result matrix is declared as r.

Multiplication.

m1.dot m2, r
m1.dot m2.buffer, r.buffer

Notes:

• Width of the first matrix must match the height of the second matrix. Without internal checks which are mentioned before, program will fail with Index out of bounds (IndexError).
• Width of the result (r) matrix must match the width of the second matrix.
• Height of the result (r) matrix must match the height of the first matrix.

Elementwise multiplication.

m1.mul m2, r
m1.mul m2.buffer, r.buffer

Notes:

• Dimenstions of the matrices must match
• If dimensions of m1 matrix are less than m2 matrix, only these dimensions will be used
• If dimensions of r matrix are greater than m1 matrix, the "tail" will not be used
• Otherwise program will fail with Index out of bounds (IndexError)

Summation.

m1.sum m2, r
m1.sum m2.buffer, r.buffer

Notes:

• Dimenstions of the matrices must match
• If dimensions of m1 matrix are less than m2 matrix, only these dimensions will be used
• If dimensions of r matrix are greater than m1 matrix, the "tail" will not be used
• Otherwise program will fail with Index out of bounds (IndexError)

Transpose.

m.transpose r
m.transpose r.buffer

Notes:

• Width of the m (source) matrix must match the height of the r (result) matrix
• Height of the m (source) matrix must match the height of the r (result) matrix

Convolution. (window, or kernel)

m.convolve k, r, padding
m.convolve k.buffer, r.buffer, padding

Notes:

• Kernel dimensions must be less than or equal to matrix dimensions
• Padding must be greater than or equal to zero
• Dimensions of a result matrix r should be calculated taking into account matrix m dimensions and kernel k dimensions. Convolution example can be found in convolution specs.

Apply function elementwise.

This is similar to applying the activation function to each neurons.

m.apply r do |m_matrix_value, m_y, m_x|
  # do something...
end

m.apply r.buffer do |m_matrix_value, m_y, m_x|
  # do something...
end

Notes:

• Result must have the same type as matrices
• Matrices width and height must match

Development

Please feel free to make an PR!

Please be sure that engine is still have the desired performance stats: no mallocs + best benchmark ips.

Contributing

1. Fork it (https://github.com/fruktorum/simple_matrix/fork)
2. Create your feature branch (git checkout -b my-new-feature)
3. Commit your changes (git commit -am 'Add some feature')
4. Push to the branch (git push origin my-new-feature)
5. Create a new Pull Request

Contributors

• SlayerShadow - creator and maintainer