golem 0.6.0

A simple tensor library for the computational graph on CPU


To use this package, run the following command in your project's root directory:

Manual usage
Put the following dependency into your project's dependences section:

golem

D

A library that provides tensors and computational graphs for machine learning.

Concept

  1. Lower the hurdle and learning curve to getting started.
  2. Use language features to make model description simple and safe.
  3. Unify the code and language for training and production environments.

Features

  • A statically size checked tensor
  • with shape-safe operators
  • Statically omit grads from tensor
  • UseGradient.no or No.gradient
  • Some friendly error messages
  • Simple SGD , Adam or AdaBelief optimizer

Usage

command

dub init
dub add golem

dub.json

{
    "dependencies": {
        "golem": "~>0.6.0"
    }
}

dub.sdl

dependency "golem" version="~>0.6.0"

Examples

import golem;

// statically sized tensor
auto x = tensor!([2, 2])([
        [0.1, 0.2],
        [0.3, 0.4],
    ]);
auto y = tensor!([2, 2])([
        [-0.1, 0.2],
        [0.3, -0.4],
    ]);

auto z = x + y;

assert(z.value[0, 0] == 0.0);
import golem.random : randn;

// no grads tensor
Tensor!(float, [3, 3], UseGradient.no) x = randn!(float, [3, 3], No.gradient);

Tensor Shape

// 3 x 2
auto x = tensor!([3, 2])(
    [1.0, 2.0],
    [3.0, 4.0],
    [5.0, 6.0],
);
// N x 2
auto y = tensor!([0, 2])([
    [1.0, 2.0],
    [3.0, 4.0],
]);

assert(x.shape == [3, 2]);
assert(y.shape == [2, 2]);

static assert(x.staticShape == [3, 2]);
static assert(y.staticShape == [0, 2]);

assert(x.runtimeShape == [3, 2]);
assert(y.runtimeShape == [2, 2]);

const batchSize = x.shape[0];
auto x = tensor!([3, 2])(
        [1.0, 2.0],
        [3.0, 4.0],
        [5.0, 6.0],
    );
auto y = tensor!([2, 2])([
        [1.0, 2.0],
        [3.0, 4.0],
    ]);
auto z = tensor!([0, 2])([
        [1.0, 2.0],
        [3.0, 4.0],
    ]);

// can not compile
static assert(!__traits(compiles, {
        auto a = x + y;
    }));

// runtime error
assertThrown!AssertError(x + z);
// more dimentions
// like PyTorch (batch, channel, width, height)
Tensor!(float, [0, 3, 224, 224]) images;

// (batch, width, height)
Tensor!(float, [0, 28, 28], UseGradient.no) mnistImages;

Broadcast Operation

Tensor!(float, [0, 28, 28]) images;
Tensor!(float, [28, 28]) offset;

// shape rule for broadcast
assert(images.shape.length > offset.shape.length);
assert(images.shape[$ - offset.shape.length .. $] == offset.shape);

// broadcast
auto result1 = broadcastOp!"+"(images, offset);
auto result2 = broadcastOp!"-"(images, offset);
auto result3 = broadcastOp!"*"(images, offset);

// implement broadcast!"/"
auto result4 = broadcastOp!"/"(images, onesLike(offset) / offset);

Linear

import golem;

// prepare datasets with dynamic batch sizes
auto data = tensor!([0, 2])([
        [0.1, 0.2],
        [0.1, 0.3],
        [0.15, 0.4],
        [0.2, 0.5],
    ]);
auto label = tensor!([0, 1])([
        [0.4],
        [0.5],
        [0.6],
        [0.7],
    ]);

// init
auto linear = new Linear!(double, 2, 1);
auto optimizer = createOptimizer!SGD(linear);

// train
foreach (epoch; 0 .. 10_000)
{
    auto y = linear(data);
    auto diff = label - y;
    auto loss = mean(diff * diff);

    optimizer.resetGrads();
    loss.backward();
    optimizer.trainStep();
}

// result
auto y = linear(data);
writeln(y.value);

Optimizer

import golem.nn : Linear;
import golem.optimizer;

auto fc1 = new Linear!(float, 28 * 28, 100);
auto fc2 = new Linear!(float, 100, 10);

// create instance with parameters
auto sgd = createOptimizer!SGD(fc1, fc2);
auto adam = createOptimizer!Adam(fc1, fc2);

// reset grads
sgd.resetGrads();
adam.resetGrads();

// train step
sgd.trainStep();
adam.trainStep();
// configure Parameters

// SGD
auto sgd = createOptimizer!SGD(fc1, fc2);
sgd.config.learningRate = 0.1;  // default 0.01
sgd.config.momentumRate = 0.95; // default 0.9
adam.config.weightDecay = 1e-3; // default 0

// Adam
auto adam = createOptimizer!Adam(fc1, fc2);
adam.config.learningRate = 0.1; // default 0.001
adam.config.beta1 = 0.95;       // default 0.9
adam.config.beta2 = 0.99;       // default 0.999
adam.config.eps = 1e-6;         // default 1e-8
adam.config.weightDecay = 1e-3; // default 0

// AdaBelief
// https://arxiv.org/abs/2010.07468
auto adabelief = createOptimizer!AdaBelief(fc1, fc2);
adabelief.config.learningRate = 0.1; // default 0.001
adabelief.config.beta1 = 0.95;       // default 0.9
adabelief.config.beta2 = 0.99;       // default 0.999
adabelief.config.eps = 1e-6;         // default 1e-8
adabelief.config.weightDecay = 1e-3; // default 0

Custom Model

Perceptron

  • Dim : Input -> Hidden -> Output
  • Activation : sigmoid
class Perceptron(size_t Input, size_t Hidden, size_t Output)
{
    // layers
    Linear!(float, Input, Hidden) fc1;
    Linear!(float, Hidden, Output) fc2;

    // targets of the optimization
    alias parameters = AliasSeq!(fc1, fc2);

    this()
    {
        // init layers
        foreach (ref p; parameters)
            p = new typeof(p);
    }

    auto forward(T)(T x)
    {
        auto h = sigmoid(fc1(x));
        auto o = sigmoid(fc2(h));
        return o;
    }
}

AutoEncoder

  • Dim : 10 -> 8 -> |3| -> 8 -> 10
class AutoEncoder
{
    // Nested custom model
    Perceptron!(float, 10, 8, 3) encoder;
    Perceptron!(float, 3, 8, 10) decoder;

    alias parameters = AliasSeq!(encoder, decoder);

    this()
    {
        foreach (ref p; parameters)
            p = new typeof(p);
    }

    auto forward(T)(T x)
    {
        auto encoded = encode(x);
        auto decoded = decode(encoded);
        return decoded;
    }

    auto encode(T)(T x)
    {
        return encoder.forward(x);
    }

    auto decode(T)(T x)
    {
        return decoder.forward(x);
    }
}
Use Sequence
alias Perceptron(size_t Input, size_t Hidden, size_t Output) = Sequence!(
        Linear!(float, Input, Hidden),
        Activation!sigmoid,
        Linear!(float, Hidden, Output),
        Activation!sigmoid,
    );

auto net = new Perceptron!(2, 2, 1);
auto x = tensor!([0, 2])([1.0f, 2.0f]);
auto y = net(x);

K-Fold

import golem.data.common;

auto source = [1.0, 2.0, 3.0, 4.0, 5.0];
auto datasets = source.kfold!5();

foreach (dataset; parallel(datasets[]))
{
    auto train = dataset[0];
    auto test = dataset[1];

    // train
}

Save & Load

auto model = new Model;
auto archiver = new ModelArchiver("model_data");
archiver.load(model); // recent saved parameters

foreach (epoch; 0 .. N)
{
    // train

    archiver.save(model); // save each epoch
}

filename format

./model_data
    model_yyyyMMdd-hhmmss.dat

Modules

Dependency Graph

dependency graph

Authors:
  • lempiji
Dependencies:
numir, msgpack-d, mir-blas, mir-algorithm
Versions:
0.6.0 2021-Apr-04
0.5.1 2021-Mar-07
0.5.0 2021-Mar-07
0.4.0 2021-Mar-03
0.3.0 2021-Feb-16
Show all 9 versions
Download Stats:
  • 0 downloads today

  • 0 downloads this week

  • 2 downloads this month

  • 7 downloads total

Score:
1.3
Short URL:
golem.dub.pm