Ceras

What I cannot create, I do not understand. -- Richard Feynman

Ceras is yet another deep learning engine aiming to reinvent Keras, in C++20 and header-only.

Getting Started

A model is a way to organize layers. Here is an example to build a sequential model.

First, we include the header and use the namespace of ceras:

#include "./include/ceras.hpp"
using namespace ceras;

Then we compose layers using a functional interface. This builds up a computation graph:

auto input = Input({28, 28}); // shape( 28, 28 )
auto l0 = Reshape({28*28,})( input );
auto l1 = ReLU( Dense( 512 )( l0 ) );
auto l2 = ReLU( Dense( 256 )( l1 ) );
auto output = Dense( 10  )( l2 );

We generate a model by collecting the input layer and the output layer of this computation graph:

auto m = model{ input, output };

We can dump the structure to double check the architecture visually:

m.summary( "./mnist_minimal.dot" );

After generating a dot file 'mnist_minimal.dot', we convert it to a '.png' file by running dot -Tpng ./mnist_minimal.dot -o ./mnist_minimal.png. Here dot is an external command from package ImageMagick.

Before feeding a training set to this model, we neet to configure the training hyper-parameters

unsigned long batch_size = 10;
float learning_rate = 0.01;
auto cm = m.compile( CategoricalCrossentropy(), SGD(batch_size, learning_rate) );

Then we can train this model. Typical training API is:

unsigned long epoch = 50;
int verbose = 1;
double validation_split = 0.1;
cm.fit( input_data_of_784, output_data_of_10, batch_size, epoch, verbose, validation_split );

We can also train this model batch by batch:

cm.train_on_batch( input_batch_of_784, output_batch_of_10 );

We can evaluate the performance this way:

auto error = cm.evaluate( test_data_of_784, test_data_of_10, batch_size );

We can also make predictions from the new samples:

auto prediction = cm.predict( new_data_of_784 );

Check out a minimal example from here. On Linux/Unix, type make mnist_minimal && ./bin/test_mnist_minimal or make mnist_conv2d_minimal && ./bin/test_mnist_conv2d_minimal to try.

Some more examples:

vgg16: model and visualization.
unet: model and visualization.

Implementation using mid-level APIs

First we define the input layer:

// define computation graph, a 3-layered dense net with topology 784x256x128x10
using namespace ceras;
auto input = place_holder<tensor<float>>{}; // 1-D, 28x28 pixels

Then we define the first layer with a relu activation

// 1st layer
auto w1 = variable{ randn<float>( {28*28, 256}, 0.0, 10.0/(28.0*16.0) ) };
auto b1 = variable{ zeros<float>( { 1, 256 } ) };
auto l1 = relu( input * w1 + b1 );

The second layer with a relu or a sigmoid activation

// 2nd layer
auto w2 = variable{ randn<float>( {256, 128}, 0.0, 3.14/(16.0*11.2 )) };
auto b2 = variable{ zeros<float>( { 1, 128 } ) };
//auto l2 = relu( l1 * w2 + b2 );
auto l2 = sigmoid( l1 * w2 + b2 );

And the last layer

// 3rd layer
auto w3 = variable{ randn<float>( {128, 10}, 0.0, 1.0/35.8 ) };
auto b3 = variable{ zeros<float>( { 1, 10 } ) };
auto output = l2 * w3 + b3;

After defining the model, we can define the loss

auto ground_truth = place_holder<tensor<float>>{}; // 1-D, 10
auto loss = cross_entropy_loss( ground_truth, output );

And create a session

// creating session
session<tensor<float>> s;

tensor<float> input_images = ...; // (batch_size, 784)
tensor<float> output_labels = ...; // (batch_size, 10)
s.bind( input, input_images );
s.bind( ground_truth, output_labels );

Defining an optimizer

float learning_rate = 1.0e-1f;
auto optimizer = gradient_descent{ loss, batch_size, learning_rate };

And starting training the loss and the optimizer

for ( auto e : range( epoch ) )
{
    for ( auto i : range( iteration_per_epoch ) )
    {
        // here we update the contents in input_images and output_labels
        s.run( loss ); //forward pass
        s.run( optimizer ); //backward pass
    }
}

At last we can make prediction, this is done by rebinding the input layer

std::size_t new_batch_size = 1;
tensor<float> new_input_images{ {new_batch_size, 28 * 28} };
s.bind( input, new_input_images );

auto precition = s.run( output );

Checkout full example from this file.

Features

Fast and GPU memory friendly:
- 98% accuracy on MNIST in 10 epochs in 30s (loading dataset, training and validation on a laptop with Intel(R) Core(TM) i7-7700HQ and a mobile GTX 1060);
- only time-consuming operations are accelerated by GPU (such as GEMM), and your GPU memory size is not the restriction for large batch size.
Portable:
- header-only;
- CUDA acceleration is optional, not a must.
Simply implemented:
- mimicking Tensorflow/keras grammar for ease of use;
- minimizing the levels of indirections to expose as many implementation details as possible.

Usage

Using this library:

copy the include directory to the working directory, then include the header file

#include "ceras.hpp"

Compile/link:

g++ -c -std=c++20 -Wall -Wextra -ferror-limit=1 -ftemplate-backtrace-limit=0 -funsafe-math-optimizations  -Ofast -flto -pipe -march=native -DNDEBUG -o ./obj/test_mnist.o test/mnist.cc
g++ -o ./bin/test_mnist ./obj/test_mnist.o -funsafe-math-optimizations  -Ofast -flto -pipe -march=native

CBLAS can be optionally enabled by define macro CBLAS (tested with cblas 3.10.0, g++ 12.2.0):

g++ -c -std=c++20 -ftemplate-backtrace-limit=0 -funsafe-math-optimizations  -Ofast -flto -pipe -march=native -DCBLAS -o ./obj/test_mnist.o test/mnist.cc
g++ -funsafe-math-optimizations  -Ofast -flto -pipe -march=native -o ./bin/test_mnist ./obj/test_mnist.o -L/opt/cuda/lib64 -pthread  -lcblas

CUDA/CUBLAS could be optionally enabled by defining macro CUDA: (tested with cuda 11.2.r11.2, gcc 12.2.0, note the compile/link options)

g++ -c -std=c++20 -Wall -Wextra -fmax-errors=1 -ftemplate-backtrace-limit=0 -funsafe-math-optimizations  -Ofast -flto -pipe -march=native -DCUDA -DNDEBUG -o ./obj/test_mnist.o test/mnist.cc
g++ -funsafe-math-optimizations  -Ofast -flto -pipe -march=native -o ./bin/test_mnist ./obj/test_mnist.o -L/opt/cuda/lib64 -pthread  -lcudart -lcublas

However, this will override CBLAS.

Note: As Non-Type Template Parameters is not yet implemented in clang, only gcc works with this library.

Design

tensor

A tensor variable holds a multiple dimensional array. It can be created directly by its constructor:

ceras::tensor<float> a{{2, 1, 2}, {0.0f, 0.1f, 0.2f, 0.3f}};

in which the template parameter float is for the data type, the first argument {2, 1, 2} is for the tensor shape, and the second argument {0.0f, 0.1f, 0.2f, 0.3f} is for the data stored in the tensor.

Quite a few operations, such as +, -, *, abs, random, randn, reduce and max are implemented for tensor. But these operations are there to serve the purpose of deep learning, not intend to make a generic tensor library.

creating tensors

An empty tensor could be directly created using its constructor by passing a shape parameter:

auto data = ceras::tensor<float>{{16, 28, 28, 77}};

Or creating a zeros tensor:

auto empty = ceras::zeros<float>( {16, 28, 28, 21} );
auto empty_2 = ceras::zeros_like( data );

or creating a ones tensor:

auto one = ceras::ones<float>( { 28, 28, 21} );
auto one_2 = ceras::ones_like( data );

or a tensor filled with random values

auto r = ceras::random<float>( {12, 34} ); // U(0, 1)
auto r_1 = ceras::random<float>( {12, 34}, -10.0, 10.0 ); // U(-10, 10)
auto r_2 = ceras::random_like( data ); // U(0, 1)
auto r_3 = ceras::random_like( data, -10.0, 10.0 ); // U(-10, 10)

or a tensor sampling values from a Normal distribution

auto n = ceras::randn<float>( {12, 34} ); // N(0, 1)
auto n_1 = ceras::randn<float>( {12, 34}, 1.0, 10.0 ); // N(1, 10)
auto n_2 = ceras_randn_like( data ); // N(0, 1)
auto n_3 = ceras_randn_like( data, 1.0, 10.0 ); // N(1, 10)

access elements

It is possible to access elements by iterators to read

std::copy( data.begin(), data.end(), std::ostream_iterator<float>( std::cout, " " ) );

or write a tensor

std::fill( data.begin(), data.end(), 0.0 );

1D view is enabled by default:

data[0] = 1.0;
data[100] = -1.0;

2D view is possible by exposing the tensor data and by setting up two dimensional parameters:

auto v2 = ceras::view_2d{ data.data(), 16, 28*28*77 };
v2[11][28] = 0.0;

For 3D view, three parameters are required

auto v3 = ceras::view_3d{ data.data(), 16, 28, 28*77 };
v3[11][20][40] = 0.0;

For 4D view, four parameters are required

auto v4 = ceras::view_3d{ data.data(), 16, 28, 28, 77 };
v4[11][20][10][20] = 0.0;

constant

A constant variable holds a tensor instance, and this tensor is not supposed to be updated in its life-time.

ceras::tensor<float> eye{{2, 2}, {1.0f, 0.0f, 0.0f, 1.0f}};
ceras::constant<ceras::tensor<float>> c_eye{eye};

place_holder

A place_holder variable holds a position that a tensor will be fed later.

ceras::place_holder<ceras::tensor<float>> input{};
// ......
auto&s = ceras::get_default_session<ceras::tensor<float>>();
ceras::tensor<float> a{{2, 1, 2}, {0.0f, 0.1f, 0.2f, 0.3f}};
s.bind(input, a ); // binding a tensor to a place_holder

variable

A variable variable holds a stateful tensor, and this tensor will be updated anytime. This is designed for the weights in a neural network, which will be updated in every epoch of the training.

auto w = ceras::variable{ ceras::randn<float>( {28*28, 256}, 0.0, 10.0/(28.0*16.0) ) };

We can also apply regularizers to this variable:

float const l1_regularizer = 1.0e-5;
float const l2_regularizer = 1.0e-5;
auto w = ceras::variable{ ceras::randn<float>( {28*28, 256}, 0.0, 10.0/(28.0*16.0) ), l1_regularizer, l2_regularizer };

value

A value variable holds a constant value. The operations between a value and another operator (such as a constant, a place holder, a variable or an operation) is element-wised.

auto v = ceras::value{ 0.0f };
auto op = ...; // a constant, a place holder, a variable, or an operation
auto xx = ceras::maximum( v op );
//auto xx = v + op;
//auto xx = v - op;
//auto xx = v * op;

operation and computation graph

ceras uses expression template to represent a computation graph. A computation graph is a directed graph in which each node corresponds to a variable, a place_holder, a constant or an operation. In ceras, these node types are grouped in a Expression concept.

For example, a computation graph computes output Expression z of two input Expression x and y. Here x and y are two input nodes of z, and z is the consumer of x and y.

If x and y are two tensors are to be binded in a later stage, the corresponding code is

auto x = ceras::place_holder<ceras::tensor<float>>{};
auto y = ceras::place_holder<ceras::tensor<float>>{};
auto z = x + y;

This kind of expression is more useful when the computation is getting more complex, for example z = σ(A*x+b)

in which x, A and b are variables / place_holders / constants, and *, + and σ are operationss.

If A and b are two variables, and x is a place_holder, then the corresponding code is

auto x = ceras::place_holder<ceras::tensor<float>>{};
auto A = ceras::variable{ ceras::ones<float>({3, 3}) };//just for demostration, should not be initialized to ones
auto b = ceras::variable{ ceras::zeros<float>({3,}) };
auto z = sigmoid( A*x + b );

session

To evaluate the operations (computation graph), we need a session.

auto&s = ceras::get_default_session<ceras::tensor<float>>();

Then we can bind a tensor to x,

auto X = ceras::tensor<float>{{3,}, {1.0f, 2.0f, 3.0f}};
s.bind(x, X);

And evaluate the output at node z:

auto result = s.run(z);

This will generate a result tensor with shape (3,) and values (0.997527, 0.997527,0.997527). In addition, the x, A and b can also be evaluated by calling

auto _x = s.run(x);
auto _A = s.run(A);
auto _b = s.run(b);

By design, an instance of an expression has a built-in forward() method. When a session runs an expression, the forward() method will be invoked.

Please find the complete code from this file.

A session can be serialized to a file

s.save( "./test/mnist.session" );
// or
s.serialize( "./test/mnist.session" );

Also it can be deserialized from a file

s.restore( "./test/mnist.session" );
// or
s.deserialized( "./test/mnist.session" );

loss

A loss variable provides a metric between the expected output and the actual output of the computation graph. And a loss is implemented as an Expression. For example, the mae loss can be defined as

template < Expression Lhs_Expression, Expression Rhs_Expression >
auto constexpr mae( Lhs_Expression const& ground_truth, Rhs_Expression const& output ) noexcept
{
    return mean_reduce(abs(ground_truth - output));
};

in which mean_reduce, abs and - are predefined operations. Usually the ground_truth is just a place_holder variable, and will be rebinded at every training epoch.

We can define our loss operation with a place_holder for the ground_truth

auto ground_truth = ceras::place_holder<tensor<float>>{};
auto loss = mae(ground_truth, z);

optimizer

An optimizer variable holds an instance of an expression of loss. When an session runs an optimizer, the builtin method forward() will be invoked. And we define an optimizer this way:

unsigned long batch_size = ...;
float learning_rate = ...;
auto optimizer = ceras::sgd{loss, batch_size, learning_rate};

In a single epoch, we first execute a forward pass on the loss, with input x and ground_truth having been binded:

s.bind( x, ...);
s.bind(ground_truth, ...);
s.run(loss);

then we execute a backward pass with the optimizer:

s.run(optimizer);

By repeating forward pass and backward pass multiple times, the weights A and the bias b can be gradually updated.

model

model groups an expression template into an object with training and inference features.

We can first define an expression template

auto input = Input();
auto l1 = relu( Dense( 512, 28*28 )( input ) );
auto l2 = relu( Dense( 256, 512 )( l1 ) );
auto output = sigmoid( Dense( 10, 256 )( l2 ) );

Then we train this little model by defining a loss and an optimizer

// training

Afterwards, we can build a model for later use,

auto m = model{ input, output };

Then we can make some predictions using this model

auto prediction = m.predict( some_input_dataset );

more details

TODO

Examples

implement VGG16

There are a few pre-defined layers in file ./include/layer.hpp, such as Input, Conv2D and Dense. Starting from these layers, we are already able to build a VGG16 model.

The input layer for VGG16 is defined as

auto input = Input(); //  3D tensor input, (batch_size, 224, 224, 3)

followed by a convolutional layer and a relu activation

auto l0 = relu( Conv2D( 64, {3, 3}, {224, 3, 3}, "same" )(input) ); // 224, 224, 64

and a max pooling layer

auto l1 = max_pooling_2d( 2 ) ( l0 ); // 112, 112, 64

Then 2 convolutional layers and a max pooling layer

auto l2 = relu( Conv2D( 128, {3, 3}, {112, 112, 64}, "same" )( l1 ) ); // 112, 112, 128
auto l3 = relu( Conv2D( 128, {3, 3}, {112, 112, 128}, "same" )( l2 ) ); // 112, 112, 128
auto l4 = max_pooling_2d( 2 ) ( l3 ); // 56, 56, 128

followed by 3 convolutional layers and a max pooling layer

auto l5 = relu( Conv2D( 256, {3, 3}, {56, 56, 128}, "same" )( l4 ) ); // 56, 56, 256
auto l6 = relu( Conv2D( 256, {3, 3}, {56, 56, 256}, "same" )( l5 ) ); // 56, 56, 256
auto l7 = relu( Conv2D( 256, {3, 3}, {56, 56, 256}, "same" )( l6 ) ); // 56, 56, 256
auto l8 = max_pooling_2d( 2 ) ( l7 ); // 28, 28, 256

followed by another 3 convolutional layers and a max pooling layer

auto l9 = relu( Conv2D( 512, {3, 3}, {28, 28, 256}, "same" )( l8 ) ); // 28, 28, 512
auto l10 = relu( Conv2D( 512, {3, 3}, {28, 28, 512}, "same" )( l9 ) ); // 28, 28, 512
auto l11 = relu( Conv2D( 512, {3, 3}, {28, 28, 512}, "same" )( l10 ) ); // 28, 28, 512
auto l12 = max_pooling_2d( 2 ) ( l11 ); // 14, 14, 512

and again

auto l13 = relu( Conv2D( 512, {3, 3}, {14, 14, 512}, "same" )( l12 ) ); // 14, 14, 512
auto l14 = relu( Conv2D( 512, {3, 3}, {14, 14, 512}, "same" )( l13 ) ); // 14, 14, 512
auto l15 = relu( Conv2D( 512, {3, 3}, {14, 14, 512}, "same" )( l14 ) ); // 14, 14, 512
auto l16 = max_pooling_2d( 2 ) ( l15 ); // 7, 7, 512

then this 3d layer is flattened to 1d

auto l17 = flatten( l16 ); // 7x7x512

followed by a dense layer

auto l18 = relu( Dense( 4096, 7*7*512 )( l17 ) ); // 4096

and then 2 dense layers to the output layer

auto l19 = relu( Dense( 4096, 4096 )( l18 ) ); // 4096
auto l20 = relu( Dense( 1000, 4096 )( l19 ) ); // 1000
auto output = l20;

With above codes, VGG16 model has been build. However, we not train this model here as we do not have the training set yet. But we can demonstrate the training process with mnist, which is a dataset much smaller than imagenet. The computation graph can be found from this file and this image.

defining a 3-layered NN, 256+128 hidden units for mnist

define a 3 layer model

// define computation graph, a 3-layered dense net with topology 784x256x128x10
using namespace ceras;
auto input = Input();

// 1st layer
auto l1 = relu( Dense( 256, 28*28 )( input ) );
// or enabling BN
//auto l1 = relu( BatchNormalization( {256,} )( Dense( 256, 28*28 )( input ) ) );

// 2nd layer
auto l2 = sigmoid( Dense( 128, 256 )( l1 ) );

// 3rd layer
auto output = Dense( 10, 128 )( l2 );

auto ground_truth = place_holder<tensor<float>>{}; // 1-D, 10
auto loss = cross_entropy_loss( ground_truth, output );

preparing dataset

std::size_t const batch_size = 10;
tensor<float> input_images{ {batch_size, 28*28} };
tensor<float> output_labels{ {batch_size, 10} };

std::size_t const epoch = 1;
std::size_t const iteration_per_epoch = 60000/batch_size;

prepare session

// creating session
auto s = ceras::session<ceras::tensor<float>>{};
s.bind( input, input_images );
s.bind( ground_truth, output_labels );

define optimizer

float learning_rate = 1.0e-1f;
auto optimizer = gradient_descent{ loss, batch_size, learning_rate };

start training

for ( auto e : range( epoch ) )
{
    for ( auto i : range( iteration_per_epoch ) )
    {
        // update input_images, output_labels for the current batch
        s.run( loss ); //forward pass
        s.run( optimizer ); //backward pass
    }
}

make prediction

std::size_t new_batch_size = 1;
tensor<float> new_input_images{ {new_batch_size, 28 * 28} };
s.bind( input, new_input_images );

for ( auto i : range( tests ) )
{
    //prepare new_input_images as inputs
    auto precition = s.run( output );
    //post precess prediction
}

[alternative] define a convolutional model

using namespace ceras;
auto input = Input(); // 28*28
auto l0 = reshape( {28, 28, 1} )( input ); // 28, 28, 1
auto l1 = relu( Conv2D( 32, {3, 3}, {28, 28, 1}, "valid" )( l0 ) );
auto l2 = max_pooling_2d( 2 ) ( l1 ); // 13, 13, 32
auto l3 = relu( Conv2D( 64, {3, 3}, {13, 13, 32}, "valid" )( l2 ) );
auto l4 = max_pooling_2d( 2 )( l3 ); //5, 5, 64
auto l5 = drop_out(0.5)( flatten( l4 ) );
auto output = Dense( 10, 5*5*64 )( l5 );

auto ground_truth = place_holder<tensor<float>>{}; // 1-D, 10
auto loss = cross_entropy_loss( ground_truth, output );

Note: this convolutional model uses drop_out, when training this model, we should set ceras::learning_phase = 1;, which is the default value; and when doing prediction using this model, we should set ceras::learning_phase = 0;. This is also the case for BatchNormalization. The reason is that, the forward propagation behaviours for drop_out and BatchNormalization layers are different between the training and the prediction phase.

Supported layers

Operations:
- plus, or operator +;
- minus, or operator -;
- divide, or operator /;
- multiply, or operator *, note this operation implies matrix-matrix multiplication, i.e., dot in numpy;
- log;
- negative;
- elementwise_product, or hadamard_product;
- expand_dims;
- sum_reduct;
- mean_reduce;
- minus;
- square;
- abs;
- exp;
- clip;
- reshape;
- flatten;
- flip;
- identity;
- transpose;
- conv2d;
- conv2d_transpose;
- drop_out;
- max_pooling_2d;
- zero_padding_2d;
- repeat;
- average_pooling_2d;
- up_sampling_2d;
- batch_normalization;
- instance_normalization;
- concatenate, or concat;
- maximum;
- minimum;
- random_normal_like.
- sqrt.
- hypot.
- ones_like.
- zeros_like.
- atan2.
- equal.
- sign.
- reduce_min.
- reduce_max.
- reduce_sum.
- abs.
- acos.
- acosh.
- asin.
- asinh.
- atan.
- atanh.
- cbrt.
- ceil.
- cos.
- cosh.
- erf.
- erfc.
- exp.
- exp2.
- expm1.
- fabs.
- floor.
- llrint.
- llround.
- log.
- log10.
- log1p.
- log2.
- lrint.
- lround.
- nearbyint.
- rint.
- round.
- sin.
- sinh.
- sqrt.
- tan.
- tanh.
- trunc.
- argmin.
- argmax.
Complex
- +
- -
- *
- real
- imag
- abs
- norm
- conj
- polar
- arg
Activations:
- softmax;
- selu;
- softplus;
- softsign;
- sigmoid;
- tanh;
- relu;
- leaky_relu;
- elu;
- exponential;
- hard_sigmoid;
- gelu.
- swish.
- silu.
- crelu.
- tank_shrink.
- mish.
- lisht.
Losses:
- mae;
- mse;
- cross_entropy;
- binary_entropy;
- hinge_loss.
Optimizers:
- sgd;
- adagrad;
- rmsprop;
- adadelta;
- adam;
- gradient_descent.

ExpandDims

expand_dims expands the dimensions of the input layer at the given axis by 1.

    auto x = variable{ ones<float>({2, 3, 4}) };
    auto x0 = expand_dims(0)( x ); // new shape is ( 1, 2, 3, 4 )
    auto x1 = expand_dims(1)( x ); // new shape is ( 2, 1, 3, 4 )
    auto x2 = expand_dims(2)( x ); // new shape is ( 2, 3, 1, 4 )
    auto x3 = expand_dims(-1)( x ); // new shape is ( 2, 3, 4, 1 )

plus

plus or + does element-wise addition. (note broadcasting is permitted.)

    auto a = variable{ ones<float>( {2, 2} ) };
    auto b = variable{ zeros<float>( {2, 2} ) };
    auto ab = a+b; // or 'auto ab = plus( a, b );'
    ceras::session<ceras::tensor<double>> s;
    std::cout <<  s.run( ab );

this will produce a 2x2 matrix of [ [1, 1], [1, 1] ]. Full code is here.

multiply

multiply or * does matrix multiplication.

    auto a = variable{ ones<float>( {2, 2} ) };
    auto b = variable{ ones<float>( {2, 2} ) };
    auto ab = a*b; // or 'auto ab = multiply( a, b );'
    ceras::session<ceras::tensor<double>> s;
    std::cout <<  s.run( ab );

this will produce a 2x2 matrix of [[2, 2], [2, 2]]. Full code is here.

log

log does element-wise logarithm on each element.

    auto a = variable{ ones<float>( {2, 2} ) };
    auto la = log(a);
    ceras::session<ceras::tensor<double>> s;
    std::cout <<  s.run( la );

this will produce a 2x2 matrix of [[0, 0], [0, ]]. Full code is here.

softmax

softmax applies softmax on last channel elements.

    auto a = variable{ ones<float>( {2, 2} ) };
    auto ls = softmax(a);
    ceras::session<ceras::tensor<double>> s;
    std::cout <<  s.run( ls );

this will produce a 2x2 matrix of [[0.5, 0.5], [0.5, 0.5]]. Full code is here.

mae

mae gives out mean absolute error.

    auto a = variable{ ones<float>( {2, 2} ) };
    auto b = variable{ zeros<float>( {2, 2} ) };
    auto ab = mae(a, b);
    ceras::session<ceras::tensor<double>> s;
    std::cout <<  s.run( ab );

this will produce a 1x1 matrix of [1]. Full code is here.

hingeloss

hinge_loss gives hinge loss between y_true and y_pred. For example:

auto a = ceras::random<float>( {3, 3} );
ceras::for_each( a.begin(), a.end(), []( auto& v ){ v = v > 0.5f ? 1.0 : -1.0; } );
std::cout << "a created with:\n" << a << std::endl;

auto b = ceras::random<float>( {3, 3} );
ceras::for_each( b.begin(), b.end(), []( auto& v ){ v = v > 0.5f ? 1.0 : -1.0; } );
std::cout << "b created with:\n" << b << std::endl;

auto va = ceras::variable{ a };
auto vb = ceras::variable{ b };
auto diff = ceras::hinge_loss( va, vb );

ceras::session<ceras::tensor<float>> s;
auto d = s.run( diff );
std::cout << "hinge loss is\n" << d << std::endl;

for an example tensor a =

1       1       1
-1      -1      1
1       1       -1

and b =

1       1       -1
-1      1       -1
-1      -1      -1

the computed hinge loss is 1.111.

gradient_descent

gradient_decent is an optimizer taking 3 arguments:

a loss expression
a batch_size
a learning rate

A typical optimizer instance is auto optimizer = gradient_decent{ loss, batch_size, learning_rate };

    // define model, a single layer NN, using softmax activation
    auto x = place_holder<tensor<double>>{};
    auto W = variable{ tensor<double>{ {2, 2}, {1.0, -1.0, 1.0, -1.0} } };
    auto b = variable{ tensor<double>{{1,2}, {0.0, 0.0} } };
    auto p = softmax( x * W + b ); // p is our model

    // preparing input for the model
    unsigned long const N = 512;
    auto blues = randn<double>( {N, 2} ) - 2.0 * ones<double>( {N, 2} );
    auto reds = randn<double>( {N, 2} ) + 2.0 * ones<double>( {N, 2} );
    auto _x = concatenate( blues, reds, 0 );

    // binding input to layer x
    session<tensor<double>> s;
    s.bind( x, _x );

    // define loss here
    auto c = place_holder<tensor<double>>{};
    auto J = cross_entropy( c, p );

    // generating output/ground_truth for the model
    auto c_blue = tensor<double>{{1, 2}, {1.0, 0.0} };
    auto c_blues = repmat( c_blue, N, 1 );
    auto c_red = tensor<double>{{1, 2}, {0.0, 1.0} };
    auto c_reds = repmat( c_red, N, 1 );
    auto _c = concatenate( c_blues, c_reds, 0 );

    // binding output to the model
    s.bind( c, _c );
    // define optimizer here
    double const learning_rate = 1.0e-3;
    auto optimizer = gradient_descent{ J, 1, learning_rate }; // J is the loss, 1 is the batch size, learning_rate is the hyper-parameter

    auto const iterations = 32UL;
    for ( auto idx = 0UL; idx != iterations; ++idx )
    {
        // first do forward propagation
        auto J_result = s.run( J );
        std::cout << "J at iteration " << idx+1 << ": " << J_result[0] << std::endl;
        // then do backward propagation
        s.run( optimizer );
    }

Fixing the random seed to 42 by random_generator.seed( 42 );, we can get output below:

J at iteration 1: 8165.29
J at iteration 2: 643.3
J at iteration 3: 48.2642
J at iteration 4: 43.2
J at iteration 5: 39.3805
J at iteration 6: 36.3763
J at iteration 7: 33.9391
J at iteration 8: 31.9142
J at iteration 9: 30.1999
J at iteration 10: 28.726
J at iteration 11: 27.4427
J at iteration 12: 26.3131
J at iteration 13: 25.3096
J at iteration 14: 24.4111
J at iteration 15: 23.601
J at iteration 16: 22.866
J at iteration 17: 22.1955
J at iteration 18: 21.5809
J at iteration 19: 21.0151
J at iteration 20: 20.492
J at iteration 21: 20.0068
J at iteration 22: 19.5551
J at iteration 23: 19.1335
J at iteration 24: 18.7387
J at iteration 25: 18.3682
J at iteration 26: 18.0197
J at iteration 27: 17.691
J at iteration 28: 17.3805
J at iteration 29: 17.0865
J at iteration 30: 16.8077
J at iteration 31: 16.5429
J at iteration 32: 16.2909

The full code is here.

TODO

mimicking Tensorflow::Keras grammar, as close as possible
recurrent operations
provide a single-header file

Documentation

License

BSD

Name		Name	Last commit message	Last commit date
Latest commit History 480 Commits
.github/workflows		.github/workflows
assets		assets
bin		bin
dataset		dataset
doc		doc
examples		examples
include		include
obj		obj
test		test
tmp		tmp
.gitignore		.gitignore
BSD		BSD
Makefile		Makefile
ReadMe.md		ReadMe.md

fengwang/ceras

Folders and files

Latest commit

History

Repository files navigation