tiny-dnn-autoencoder.cpp

#define CNN_USE_AVX //Enable the AVX backend for faster computing
#include "tiny_dnn/tiny_dnn.h"
using namespace tiny_dnn;
using namespace tiny_dnn::activation;
using namespace tiny_dnn::layers;

#include "AudioFile.h"

#include <vector>
#include <random>
using namespace std;

int main()
{
	//Some traning variable 
	const int SAMPLE_NUM = 1024;   //How much traning samples
	const int WINDOW_SIZE = 2048;  //How long is each sample
	const int BATCH_SIZE = 16;     //How many samples per batch
	const int NUM_EPOCHS = 2048;   //How much epoch we want to run

	//Create a sample buffer
	vector<vec_t> samples(SAMPLE_NUM, vec_t(WINDOW_SIZE));

	AudioFile<float> audioFile;
	audioFile.load("a.wav"); //Load audio
	auto& data = audioFile.samples[0]; //Just use the first channel

	//Create a RNG and a distribution to generate random numbers
	mt19937 rng;
	uniform_int_distribution<int> dist(0, data.size()-WINDOW_SIZE);
	//Generate samples
	for(auto& v : samples)
	{
		int offset = dist(rng);//Generate random offsets
		auto start = data.begin() + offset;
		//Copy data from source to sample
		copy(start, start+WINDOW_SIZE, v.begin());
	}
	
	//Create an autoencoder
	network<sequential> net;
	net << fully_connected_layer(WINDOW_SIZE, 512) << tanh_layer()
		<< fully_connected_layer(512, WINDOW_SIZE);
	
	//Helper class
	tiny_dnn::progress_display disp(SAMPLE_NUM);
	tiny_dnn::timer t;
	
	int currentEpoch = 0;
	
	//Callbacks when a mini bactch is done
	auto onMinibatch = [&]()
	{
		//This updates the progress display
		disp += BATCH_SIZE;	
	};
	
	//Callbacks when an epoch is done
	auto onEpoch = [&]()
	{
		std::cout << "Epoch " << ++currentEpoch << "/" << NUM_EPOCHS << "done. "
			<< t.elapsed() << "s elapsed." << std::endl;
			
		//Reset progress display and timer
		disp.restart(SAMPLE_NUM);
		t.restart();
	};
	
	//train the network with absolute(L1) error.
	adagrad optimizer;
	net.fit<absolute>(optimizer, samples, samples, BATCH_SIZE, NUM_EPOCHS
		, onMinibatch, onEpoch);
	net.save("net");

	//Let's try the network
	vector<float> result(data.size());
	for(int i=0;i<data.size();i+=WINDOW_SIZE)
	{
		//Input to the neural network
		vec_t input(WINDOW_SIZE);
		//copy data into the input vector
		copy(data.begin()+i, data.begin()+i+WINDOW_SIZE, input.begin());

		//Run the neural network then copy it to the result buffer
		vec_t predict = net.predict(input);
		copy(result.begin()+i, result.begin()+i+WINDOW_SIZE, predict.begin());
	}
	
	//Save the audio we ganarated
	AudioFile<float> saveFile;
	AudioFile<float>::AudioBuffer buffer(1);
	buffer[0] = result;
	audioFile.setAudioBuffer(buffer);
	audioFile.save("audioFile2.wav");

	
}