num.crScientific computing in pure Crystal

联合创作 · 2023-09-26 02:33

num.cr

Num.cr is the core shard needed for scientific computing with Crystal

It provides:

  • An n-dimensional Tensor data structure
  • Efficient map, reduce and accumulate routines
  • GPU accelerated routines backed by OpenCL
  • Linear algebra routines backed by LAPACK and BLAS

Prerequisites

Num.cr aims to be a scientific computing library written in pure Crystal. All standard operations and data structures are written in Crystal. Certain routines, primarily linear algebra routines, are instead provided by a BLAS or LAPACK implementation.

Several implementations can be used, including Cblas, Openblas, and the Accelerate framework on Darwin systems. For GPU accelerated BLAS routines, the ClBlast library is required.

Num.cr also supports Tensors stored on a GPU. This is currently limited to OpenCL, and a valid OpenCL installation and device(s) are required.

Installation

Add this to your applications shard.yml

dependencies:
  num:
    github: crystal-data/num.cr

Several third-party libraries are required to use certain features of Num.cr. They are:

  • BLAS
  • LAPACK
  • OpenCL
  • ClBlast

While not at all required, they provide additional functionality than is provided by the basic library.

Just show me the code

The core data structure implemented by Num.cr is the Tensor, an N-dimensional data structure. A Tensor supports slicing, mutation, permutation, reduction, and accumulation. A Tensor can be a view of another Tensor, and can support either C-style or Fortran-style storage.

Creation

There are many ways to initialize a Tensor. Most creation methods can allocate a Tensor backed by either CPU or GPU based storage.

[1, 2, 3].to_tensor
Tensor.from_array [1, 2, 3]
Tensor(UInt8).zeros([3, 3, 2])
Tensor.random(0.0...1.0, [2, 2, 2])

ClTensor(Float32).zeros([3, 2, 2])
ClTensor(Float64).full([3, 4, 5], 3.8)

Operations

A Tensor supports a wide variety of numerical operations. Many of these operations are provided by Num.cr, but any operation can be mapped across one or more Tensors using sophisticated broadcasted mapping routines.

a = [1, 2, 3, 4].to_tensor
b = [[3, 4, 5, 6], [5, 6, 7, 8]].to_tensor

# Convert a Tensor to a GPU backed Tensor
acl = a.astype(Float64).gpu

puts Num.add(a, b)

# a is broadcast to b's shape
# [[ 4,  6,  8, 10],
#  [ 6,  8, 10, 12]]

When operating on more than two Tensors, it is recommended to use map rather than builtin functions to avoid the allocation of intermediate results. All map operations support broadcasting.

a = [1, 2, 3, 4].to_tensor
b = [[3, 4, 5, 6], [5, 6, 7, 8]].to_tensor
c = [3, 5, 7, 9].to_tensor

a.map(b, c) do |i, j, k|
  i + 2 / j + k * 3.5
end

# [[12.1667, 20     , 27.9   , 35.8333],
#  [11.9   , 19.8333, 27.7857, 35.75  ]]

Mutation

Tensors support flexible slicing and mutation operations. Many of these operations return views, not copies, so any changes made to the results might also be reflected in the parent.

a = Tensor.new([3, 2, 2]) { |i| i }

puts a.transpose

# [[[ 0,  4,  8],
#   [ 2,  6, 10]],
#
#  [[ 1,  5,  9],
#   [ 3,  7, 11]]]

puts a.reshape(6, 2)

# [[ 0,  1],
#  [ 2,  3],
#  [ 4,  5],
#  [ 6,  7],
#  [ 8,  9],
#  [10, 11]]

puts a[..., 1]

# [[ 2,  3],
#  [ 6,  7],
#  [10, 11]]

puts a[1..., {..., -1}]

# [[[ 6,  7],
#   [ 4,  5]],
#
#  [[10, 11],
#   [ 8,  9]]]

puts a[0, 1, 1].value

# 3

Linear Algebra

Tensors provide easy access to power Linear Algebra routines backed by LAPACK and BLAS implementations, and ClBlast for GPU backed Tensors.

a = [[1, 2], [3, 4]].to_tensor.map &.to_f32

puts a.inv

# [[-2  , 1   ],
#  [1.5 , -0.5]]

puts a.eigvals

# [-0.372281, 5.37228  ]

acl = a.opencl
bcl = a.opencl

puts acl.gemm(bcl).cpu

# [[7 , 10],
#  [15, 22]]

puts a.matmul(a)

# [[7 , 10],
#  [15, 22]]

Machine Learning

Num::Grad provides a pure-crystal approach to find derivatives of mathematical functions. Use a Num::Grad::Variable with a Num::Grad::Context to easily compute these derivatives.

ctx = Num::Grad::Context(Tensor(Float64)).new

x = ctx.variable([3.0])
y = ctx.variable([2.0])

# f(x) = x ** y
f = x ** y
puts f # => [9]

f.backprop

# df/dx = y * x = 6.0
puts x.grad # => [6.0]

Num::NN contains an extension to Num::Grad that provides an easy-to-use interface to assist in creating neural networks. Designing and creating a network is simple using Crystal's block syntax.

ctx = Num::Grad::Context(Tensor(Float64)).new

x_train = [[0.0, 0.0], [1.0, 0.0], [0.0, 1.0], [1.0, 1.0]].to_tensor
y_train = [[0.0], [1.0], [1.0], [0.0]].to_tensor

x = ctx.variable(x_train)

net = Num::NN::Network.new(ctx) do
  input [2]
  # A basic network with a single hidden layer using
  # a ReLU activation function
  linear 3
  relu
  linear 1

  # SGD Optimizer
  sgd 0.7

  # Sigmoid Cross Entropy to calculate loss
  sigmoid_cross_entropy_loss
end

500.times do |epoch|
  y_pred = net.forward(x)
  loss = net.loss(y_pred, y_train)
  puts "Epoch: #{epoch} - Loss #{loss}"
  loss.backprop
  net.optimizer.update
end

# Clip results to make a prediction
puts net.forward(x).value.map { |el| el > 0 ? 1 : 0}

# [[0],
#  [1],
#  [1],
#  [0]]

Review the documentation for full implementation details, and if something is missing, open an issue to add it!

浏览 6
点赞
评论
收藏
分享

手机扫一扫分享

编辑 分享
举报
评论
图片
表情
推荐
点赞
评论
收藏
分享

手机扫一扫分享

编辑 分享
举报