DeepDendro

Prerequisites

• Eigen library

  • Download link

  • Move it to the path /usr/local/include/eigen3 (or create a symbolic link):

    sudo ln -s $INSTALLATION_PATH$ /usr/local/include/eigen3

• C++ 20

Props

• indicators library

Usage example

---

Features

Convolutional layers

2D

3D

Convolutional3D conv3d{
        N_Filters,
        {3, 3, 1},
        activation::relu,
        {28, 28, 1}};

To better understand the dimensions, one can use the .print_structure() method. For example, here the output of the layer would be a Tensor with shape $(26, 26, N_{Filters})$.

conv3d.print_structure();

---

Pooling layers

2D

Shape input_shape{4, 16};
Shape grid{2, 2};
Shape stride{2, 2};

MaxPool2D maxPool2D{input_shape, grid, stride};

3D

Shape input_shape{20, 20, 16};
Shape grid{2, 2, 1};
Shape stride{2, 2, 1};

MaxPool3D maxPool3D{input_shape, grid, stride};

This is the example of a 3D Max Pooling layer. The output of the layer would be a Tensor with shape $(10, 10, 16)$. The pooling layers are used to decrease the dimensions of the input, and to reduce the number of parameters in the network.

Flattening Layers

Flattening layers are used when one needs to reshape a multidimensional Tensor into a 1D Tensor, or simply a Vector. These layers are essential when one wants to combine Convolutional and Dense layers in a network.

2D

Eigen::Tensor<double, 2> input_tensor{4, 16};

FlatteningLayer2D flattening2D;
Eigen::VectorXd flattened = flattening2D.flatten(input_tensor);

Eigen::Tensor<double, 2> original_tensor = flattening2D.back_to_tensor(flattened);

3D

Eigen::Tensor<double, 3> input_tensor(4, 16, 3);

FlatteningLayer3D flattening3D;
Eigen::VectorXd flattened = flattening3D.flatten(input_tensor);

Eigen::Tensor<double, 3> original_tensor = flattening3D.back_to_tensor(flattened);

In both examples, the input_tensor and original_tensor are equal.

---

Testing

To test the library yourself, you can download CIFAR10 dataset

sh CIFAR10/download_cifar10.sh