singlang.cpp

//tiny-dnn headers
#define CNN_USE_AVX
#include <tiny_dnn/tiny_dnn.h>
using namespace tiny_dnn;
using namespace tiny_dnn::activation;
using namespace tiny_dnn::layers;

//tiny-dnn comes with xtensor, use it
#include <tiny_dnn/xtensor/xnpy.hpp>
#include <tiny_dnn/xtensor/xview.hpp>
#include <tiny_dnn/xtensor/xsort.hpp>

#include <iostream>
#include <random>
#include <tuple>

constexpr int NUM_BATCHES = 50;
constexpr int BATCH_SIZE = 32;

//Convert data to a form tiny-dnn can read
decltype(auto) convertDataset(const xt::xarray<float>& X, const xt::xarray<float>& Y)
{
	size_t numData = X.shape()[0];
	std::vector<vec_t> x(numData);
	std::vector<label_t> y(numData);
	for(size_t i=0;i<numData;i++)
	{
		auto xb = xt::view(X, i, xt::all());
		x[i] = vec_t(xb.begin(), xb.end());
		
		auto yb = xt::view(Y, i, xt::all());
		y[i] = xt::argmax(yb)[0];
	}
	
	//shuffle dataset
	shuffle(x.begin(), x.end(), std::mt19937(42));
	shuffle(y.begin(), y.end(), std::mt19937(42));
	
	return std::tuple(x, y);
}

int main()
{
	//Load dataset
	//X.npy is saved in fp32
	xt::xarray<float> X = xt::load_npy<float>("X.npy");
	//But Y.npy is saved in fp64, needs to convert
	xt::xarray<float> Y = xt::cast<float>((xt::xarray<float>)xt::load_npy<double>("Y.npy"));
	assert(X.shape()[0] == Y.shape()[0]);
	
	network<sequential> net;
	net << conv(64, 64, 5, 1, 6) << leaky_relu() // in: 64x64x1, out 6 chanels, kernel size: 5
		<< max_pool(60, 60, 6, 2) // in: 60x60x6, 2x2 pooling
 		<< conv(30, 30, 5, 6, 9) << leaky_relu() // in: 30x30x6, out 9 channels, kernel size: 5
		<< max_pool(26, 26, 9, 2) // in:26x26x9, 2x2 pooling
		<< conv(13, 13, 6, 9, 12) << leaky_relu()// in: 13x13x9, out 12 channels, kernel size: 6
		<< fc(8*8*12, 10)
		<< softmax();
	auto [x, y] = convertDataset(X, Y);
	
	//Setup callbacks so we can see the traning progress
	size_t numData = x.size();
	progress_display disp(numData);
	timer t;
	int epoch = 1;
	auto onEnumerateEpoch = [&]()
	{
		std::cout << std::endl;
		std::cout << "Epoch " << epoch << "/" << NUM_BATCHES << " finished. "
			<< t.elapsed() << "s elapsed." << std::endl;
		++epoch;

		disp.restart(numData);
		t.restart();
	};

	auto onEnumerateMinibatch = [&](){disp += BATCH_SIZE;};
	
	//Train the model
	adam optimizer;
	net.train<mse>(optimizer, x, y, BATCH_SIZE, NUM_BATCHES
		,onEnumerateMinibatch, onEnumerateEpoch);
	
	//Test the netowrk
	net.test(x, y).print_detail(std::cout);
	
	net.save("model.net");
}