347 lines
12 KiB
C++
347 lines
12 KiB
C++
/*
|
|
* <Short Description>
|
|
* Copyright (C) 2024 Brett Terpstra
|
|
*
|
|
* This program is free software: you can redistribute it and/or modify
|
|
* it under the terms of the GNU General Public License as published by
|
|
* the Free Software Foundation, either version 3 of the License, or
|
|
* (at your option) any later version.
|
|
*
|
|
* This program is distributed in the hope that it will be useful,
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
* GNU General Public License for more details.
|
|
*
|
|
* You should have received a copy of the GNU General Public License
|
|
* along with this program. If not, see <https://www.gnu.org/licenses/>.
|
|
*/
|
|
#include <MNIST.h>
|
|
#include <blt/fs/loader.h>
|
|
#include <blt/std/memory.h>
|
|
#include <blt/std/memory_util.h>
|
|
#include <variant>
|
|
#include <blt/iterator/iterator.h>
|
|
#include <dlib/dnn.h>
|
|
#include <dlib/data_io.h>
|
|
|
|
namespace fp
|
|
{
|
|
class idx_file_t
|
|
{
|
|
template <typename T>
|
|
using mk_v = std::vector<T>;
|
|
using vec_t = std::variant<mk_v<blt::u8>, mk_v<blt::i8>, mk_v<blt::u16>, mk_v<blt::u32>, mk_v<blt::f32>, mk_v<blt::f64>>;
|
|
|
|
public:
|
|
explicit idx_file_t(const std::string& path)
|
|
{
|
|
std::ifstream file{path, std::ios::in | std::ios::binary};
|
|
|
|
using char_type = std::ifstream::char_type;
|
|
char_type magic_arr[4];
|
|
file.read(magic_arr, 4);
|
|
BLT_ASSERT(magic_arr[0] == 0 && magic_arr[1] == 0);
|
|
|
|
blt::u8 dims = magic_arr[3];
|
|
blt::size_t total_size = 1;
|
|
|
|
for (blt::i32 i = 0; i < dims; i++)
|
|
{
|
|
char_type dim_arr[4];
|
|
file.read(dim_arr, 4);
|
|
blt::u32 dim;
|
|
blt::mem::fromBytes(dim_arr, dim);
|
|
dimensions.push_back(dim);
|
|
total_size *= dim;
|
|
}
|
|
|
|
switch (magic_arr[2])
|
|
{
|
|
// unsigned char
|
|
case 0x08:
|
|
data = mk_v<blt::u8>{};
|
|
read_data<blt::u8>(file, total_size);
|
|
break;
|
|
// signed char
|
|
case 0x09:
|
|
data = mk_v<blt::i8>{};
|
|
read_data<blt::i8>(file, total_size);
|
|
break;
|
|
// short
|
|
case 0x0B:
|
|
data = mk_v<blt::u16>{};
|
|
read_data<blt::u16>(file, total_size);
|
|
reverse_data<blt::u16>();
|
|
break;
|
|
// int
|
|
case 0x0C:
|
|
data = mk_v<blt::u32>{};
|
|
read_data<blt::u32>(file, total_size);
|
|
reverse_data<blt::u32>();
|
|
break;
|
|
// float
|
|
case 0x0D:
|
|
data = mk_v<blt::f32>{};
|
|
read_data<blt::f32>(file, total_size);
|
|
reverse_data<blt::f32>();
|
|
break;
|
|
// double
|
|
case 0x0E:
|
|
data = mk_v<blt::f64>{};
|
|
read_data<blt::f64>(file, total_size);
|
|
reverse_data<blt::f64>();
|
|
break;
|
|
default:
|
|
BLT_ERROR("Unspported idx file type!");
|
|
}
|
|
if (file.eof())
|
|
{
|
|
BLT_ERROR("EOF reached. It's unlikely your file was read correctly!");
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
[[nodiscard]] const std::vector<T>& get_data_as() const
|
|
{
|
|
return std::get<mk_v<T>>(data);
|
|
}
|
|
|
|
template <typename T>
|
|
std::vector<blt::span<T>> get_as_spans() const
|
|
{
|
|
std::vector<blt::span<T>> spans;
|
|
|
|
blt::size_t total_size = data_size(1);
|
|
|
|
for (blt::size_t i = 0; i < dimensions[0]; i++)
|
|
{
|
|
auto& array = std::get<mk_v<T>>(data);
|
|
spans.push_back({&array[i * total_size], total_size});
|
|
}
|
|
|
|
return spans;
|
|
}
|
|
|
|
[[nodiscard]] const std::vector<blt::u32>& get_dimensions() const
|
|
{
|
|
return dimensions;
|
|
}
|
|
|
|
[[nodiscard]] blt::size_t data_size(const blt::size_t starting_dimension = 0) const
|
|
{
|
|
blt::size_t total_size = 1;
|
|
for (const auto d : blt::iterate(dimensions).skip(starting_dimension))
|
|
total_size *= d;
|
|
return total_size;
|
|
}
|
|
|
|
private:
|
|
template <typename T>
|
|
void read_data(std::ifstream& file, blt::size_t total_size)
|
|
{
|
|
auto& array = std::get<mk_v<T>>(data);
|
|
array.resize(total_size);
|
|
file.read(reinterpret_cast<char*>(array.data()), static_cast<std::streamsize>(total_size) * sizeof(T));
|
|
}
|
|
|
|
template <typename T>
|
|
void reverse_data()
|
|
{
|
|
auto& array = std::get<mk_v<T>>(data);
|
|
for (auto& v : array)
|
|
blt::mem::reverse(v);
|
|
}
|
|
|
|
std::vector<blt::u32> dimensions;
|
|
vec_t data;
|
|
};
|
|
|
|
class image_t
|
|
{
|
|
public:
|
|
static constexpr blt::u32 target_size = 10;
|
|
using data_iterator = std::vector<dlib::matrix<blt::u8>>::const_iterator;
|
|
using label_iterator = std::vector<blt::u64>::const_iterator;
|
|
|
|
image_t(const idx_file_t& image_data, const idx_file_t& label_data): samples(image_data.get_dimensions()[0]),
|
|
input_size(image_data.data_size(1))
|
|
{
|
|
BLT_ASSERT_MSG(samples == label_data.get_dimensions()[0],
|
|
("Mismatch in data sample sizes! " + std::to_string(samples) + " vs " + std::to_string(label_data.get_dimensions()[0])).
|
|
c_str());
|
|
auto& image_array = image_data.get_data_as<blt::u8>();
|
|
auto& label_array = label_data.get_data_as<blt::u8>();
|
|
|
|
for (const auto label : label_array)
|
|
image_labels.push_back(label);
|
|
|
|
const auto row_length = image_data.get_dimensions()[2];
|
|
const auto number_of_rows = image_data.get_dimensions()[1];
|
|
|
|
for (blt::u32 i = 0; i < samples; i++)
|
|
{
|
|
dlib::matrix<blt::u8> mat(number_of_rows, row_length);
|
|
for (blt::u32 y = 0; y < number_of_rows; y++)
|
|
{
|
|
for (blt::u32 x = 0; x < row_length; x++)
|
|
{
|
|
mat(x, y) = image_array[i * input_size + y * row_length + x];
|
|
}
|
|
}
|
|
data.push_back(mat);
|
|
}
|
|
}
|
|
|
|
[[nodiscard]] const std::vector<dlib::matrix<blt::u8>>& get_image_data() const
|
|
{
|
|
return data;
|
|
}
|
|
|
|
[[nodiscard]] const std::vector<blt::u64>& get_image_labels() const
|
|
{
|
|
return image_labels;
|
|
}
|
|
|
|
private:
|
|
blt::u32 samples;
|
|
blt::u32 input_size;
|
|
std::vector<dlib::matrix<blt::u8>> data;
|
|
std::vector<blt::u64> image_labels;
|
|
};
|
|
|
|
struct batch_stats_t
|
|
{
|
|
blt::u64 batch_size;
|
|
|
|
};
|
|
|
|
struct network_stats_t
|
|
{
|
|
};
|
|
|
|
template<typename NetworkType>
|
|
batch_stats_t test_batch(NetworkType& network, image_t::data_iterator begin, image_t::data_iterator end, image_t::label_iterator lbegin)
|
|
{
|
|
batch_stats_t stats;
|
|
|
|
|
|
|
|
return stats;
|
|
}
|
|
|
|
template <typename NetworkType>
|
|
void test_network(NetworkType& network)
|
|
{
|
|
const idx_file_t test_images{"../problems/mnist/t10k-images.idx3-ubyte"};
|
|
const idx_file_t test_labels{"../problems/mnist/t10k-labels.idx1-ubyte"};
|
|
|
|
const auto test_samples = test_images.get_dimensions()[0];
|
|
|
|
const image_t test_image{test_images, test_labels};
|
|
|
|
const auto predicted_labels = network(test_image.get_image_data());
|
|
int num_right = 0;
|
|
int num_wrong = 0;
|
|
for (size_t i = 0; i < test_image.get_image_data().size(); ++i)
|
|
{
|
|
if (predicted_labels[i] == test_image.get_image_labels()[i])
|
|
++num_right;
|
|
else
|
|
++num_wrong;
|
|
}
|
|
std::cout << "testing num_right: " << num_right << std::endl;
|
|
std::cout << "testing num_wrong: " << num_wrong << std::endl;
|
|
std::cout << "testing accuracy: " << num_right / static_cast<double>(num_right + num_wrong) << std::endl;
|
|
}
|
|
|
|
template <typename NetworkType>
|
|
network_stats_t train_network(const std::string& ident, NetworkType& network)
|
|
{
|
|
const idx_file_t train_images{"../problems/mnist/train-images.idx3-ubyte"};
|
|
const idx_file_t train_labels{"../problems/mnist/train-labels.idx1-ubyte"};
|
|
|
|
const image_t train_image{train_images, train_labels};
|
|
|
|
network_stats_t stats;
|
|
|
|
dlib::dnn_trainer trainer(network);
|
|
trainer.set_learning_rate(0.01);
|
|
trainer.set_min_learning_rate(0.00001);
|
|
trainer.set_mini_batch_size(128);
|
|
trainer.be_verbose();
|
|
|
|
trainer.set_synchronization_file("mnist_sync_" + ident, std::chrono::seconds(20));
|
|
|
|
blt::size_t epochs = 0;
|
|
blt::ptrdiff_t epoch_pos = 0;
|
|
for (; epochs < trainer.getmax_epochs() && trainer.get_learning_rate() >= trainer.get_min_learning_rate(); epochs++)
|
|
{
|
|
auto& data = train_image.get_image_data();
|
|
auto& labels = train_image.get_image_labels();
|
|
for (; epoch_pos < data.size() && trainer.get_learning_rate() >= trainer.get_min_learning_rate(); epoch_pos += trainer.
|
|
get_mini_batch_size())
|
|
{
|
|
auto begin = epoch_pos;
|
|
auto end = std::min(epoch_pos + trainer.get_mini_batch_size(), data.size());
|
|
|
|
if (end - begin <= 0)
|
|
break;
|
|
|
|
trainer.train_one_step(train_image.get_image_data().begin() + begin,
|
|
data.begin() + end, labels.begin() + begin);
|
|
}
|
|
epoch_pos = 0;
|
|
trainer.wait_for_thread_to_pause();
|
|
}
|
|
|
|
// trainer.train(train_image.get_image_data(), train_image.get_image_labels());
|
|
|
|
network.clean();
|
|
dlib::serialize("mnist_network_" + ident + ".dat") << network;
|
|
|
|
const std::vector<unsigned long> predicted_labels = network(train_image.get_image_data());
|
|
int num_right = 0;
|
|
int num_wrong = 0;
|
|
// And then let's see if it classified them correctly.
|
|
for (size_t i = 0; i < train_image.get_image_data().size(); ++i)
|
|
{
|
|
if (predicted_labels[i] == train_image.get_image_labels()[i])
|
|
++num_right;
|
|
else
|
|
++num_wrong;
|
|
}
|
|
std::cout << "training num_right: " << num_right << std::endl;
|
|
std::cout << "training num_wrong: " << num_wrong << std::endl;
|
|
std::cout << "training accuracy: " << num_right / static_cast<double>(num_right + num_wrong) << std::endl;
|
|
|
|
return stats;
|
|
}
|
|
|
|
template <typename NetworkType>
|
|
NetworkType load_network(const std::string& ident)
|
|
{
|
|
NetworkType network{};
|
|
dlib::deserialize("mnist_network_" + ident + ".dat") >> network;
|
|
return network;
|
|
}
|
|
|
|
void run_mnist(int argc, const char** argv)
|
|
{
|
|
using namespace dlib;
|
|
|
|
// using net_type = loss_multiclass_log<
|
|
// fc<10,
|
|
// relu<fc<84,
|
|
// relu<fc<120,
|
|
// max_pool<2,2,2,2,relu<con<16,5,5,1,1,
|
|
// max_pool<2,2,2,2,relu<con<6,5,5,1,1,
|
|
// input<matrix<blt::u8>>>>>>>>>>>>>>;
|
|
|
|
using net_type = loss_multiclass_log<
|
|
fc<10,
|
|
sig<fc<84,
|
|
sig<fc<120,
|
|
input<matrix<blt::u8>>>>>>>>;
|
|
}
|
|
}
|