Deep Learning Framework

Introduction

A deep learning framework with c language. I try to build a deep learning framework with few features to reduce the usage of cpu. And let it has a user-friendly api.

Feature

• Shared library
• Lite
• Tensorflow-like programming pattern

Future work

• Add a csv reader
• Add convolution and pooling
• Design a more robust algorithm for backpropagation
• Improve the precision of the partial differential
• Improve the speed of matrix calculation

Demo

A simple neural network demo

1. make
2. make demo

Quick start

1. Type command make to build shared library
2. Create a c script test.c and write down the below:

#include <stdlib.h>
#include "dl.h"

int main()
{
    int dim[2] = {2, 3};
    float constant[6] = {1, 2, 3, 4, 5, 6};
    Node *n = dl.node.constant(constant, dim, 2, "Hello World");

    dl.debug.info(n, 0);

    free(n);
    return 0;
}

3. Type command gcc -o test test.c lib/libdl.so -Isrc -Wl,-rpath=lib -lm to compile
4. Type command ./test and see result:

---------------------
Name: Hello World
---------------------
Type: Constant
Dimension: 2-D
Dimension Lenght: 2 3
Value: 1.000000 2.000000 3.000000 4.000000 5.000000 6.000000
Expression Type: None
---------------------

API

Node *dl.node.variable(uint32_t *dim, uint32_t num_dims, char *name)

Description: Create a node which has trainable variables, variables will be initialized via sampling normal distribution

Args:

• dim: The length of every dimension
• num_dims: The count of dimensions, maximum is 4
• name: The name of the node

Node *dl.node.constant(float *val, uint32_t *dim, uint32_t num_dims, char *name)

Description: Create a node which has constant, constant will not change during optimization

Args:

• val: The constant value of the node
• dim: The length of every dimension
• num_dims: The count of dimensions, maximum is 4
• name: The name of the node

Node *dl.node.placeholder(uint32_t *dim, uint32_t num_dims, char *name)

Description: Create a node which doesn't have any value

Args:

• dim: The length of every dimension
• num_dims: The count of dimensions, maximum is 4
• name: The name of the node

void dl.debug.info(Node *n, int ignore_val)

Description: Show the info of the node

Args:

• n: The node you want to check
• ignore_val: set 1 not to display all values in the node, otherwise set 0