void ANNetwork::saveWs(bool file) const
{
FILE* fp;
wchar_t fname[256] = L"";
for (unsigned int i = 0; i < m_layers.size(); i++) {
swprintf(fname, L"W%d.txt", i + 1);
if (file && (fp = _wfopen(fname, L"wt"))) {
const vec2D &tW = *m_layers[i]->m_tW;
for (unsigned int x = 0; x < tW.width(); x++) {
for (unsigned int y = 0; y < tW.height(); y++)
fwprintf(fp, L" %g\t", tW(y, x));
fwprintf(fp, L"\n");
}
fclose(fp);
} else {
const vec2D& tW = *m_layers[i]->m_tW;
wprintf(L"W%d %dx%d\n", i + 1, tW.width(), tW.height());
for (unsigned int x = 0; x < tW.width(); x++) {
for (unsigned int y = 0; y < tW.height(); y++)
wprintf(L" %8.4g", tW(y, x));
wprintf(L"\n");
}
}
}
}
ImpWTLF ( iterator BWT, uint64_t const rn, ::libmaus2::util::TempFileNameGenerator & rtmpgen, uint64_t const rmaxval = 0)
: n(rn)
{
if ( n )
{
uint64_t maxval = rmaxval;
for ( uint64_t i = 0; i < n; ++i )
maxval = std::max ( maxval, static_cast<uint64_t>(BWT[i]) );
uint64_t const b = ::libmaus2::math::numbits(maxval);
::libmaus2::wavelet::ImpExternalWaveletGenerator IEWG(b,rtmpgen);
for ( uint64_t i = 0; i < n; ++i )
IEWG.putSymbol(BWT[i]);
std::string const tmpfilename = rtmpgen.getFileName();
IEWG.createFinalStream(tmpfilename);
std::ifstream istr(tmpfilename.c_str(),std::ios::binary);
wt_ptr_type tW(new wt_type(istr));
W = UNIQUE_PTR_MOVE(tW);
istr.close();
remove ( tmpfilename.c_str() );
D = ::libmaus2::autoarray::AutoArray < uint64_t >((1ull<<W->getB())+1);
for ( uint64_t i = 0; i < (1ull<<W->getB()); ++i )
D [ i ] = W->rank(i,n-1);
D.prefixSums();
}
}
//////////////////////////////////////////init neuron weights///////////////////////////////////////////////////////
void ANNetwork::init_weights(unsigned int rseed) const
{
int w;
srand(rseed);
//input layer remains with w=1.0
for (unsigned int l = 0; l < m_layers.size(); l++) {
vec2D& tW = *m_layers[l]->m_tW;
for (unsigned int n = 0; n < tW.height(); n++) {
for (unsigned int i = 0; i < tW.width(); i++) {
w = 0xFFF & rand();
w -= 0x800;
tW(n, i) = (float)w / 2048.0f;
}
}
}
}
ImpWTLF (::libmaus2::huffman::RLDecoder & decoder, uint64_t const b, ::libmaus2::util::TempFileNameGenerator & rtmpgen)
: n(decoder.getN())
{
if ( n )
{
::libmaus2::wavelet::ImpExternalWaveletGenerator IEWG(b,rtmpgen);
for ( uint64_t i = 0; i < n; ++i )
IEWG.putSymbol(decoder.decode());
std::string const tmpfilename = rtmpgen.getFileName();
IEWG.createFinalStream(tmpfilename);
std::ifstream istr(tmpfilename.c_str(),std::ios::binary);
wt_ptr_type tW(new wt_type(istr));
W = UNIQUE_PTR_MOVE(tW);
istr.close();
remove ( tmpfilename.c_str() );
D = ::libmaus2::autoarray::AutoArray < uint64_t >((1ull<<W->getB())+1);
for ( uint64_t i = 0; i < (1ull<<W->getB()); ++i )
D [ i ] = W->rank(i,n-1);
D.prefixSums();
}
}
//////////////////////////////////////////////////constructor/destructor////////////////////////////////////////////////////////
ANNetwork::ANNetwork(const wchar_t* fname) : m_status(-1),
m_nrule(0.2f), m_alpha(0.7f),
m_input_vec(0), m_add_vec(0), m_mul_vec(0)
{
int res = 0;
unsigned int neurons_num = 0;
float w = 0.0f;
unsigned int layers_num = 0;
FILE* fp = _wfopen(fname, L"rt");
if (fp != 0) {
//get num of layers////////////////////////////////////
if ((res = fwscanf(fp, L"%d", &layers_num)) != 1) {
fclose(fp);
m_status = -1;
return;
}
m_neurons.resize(layers_num);
//get num of Neurons per layer/////////////////////////
for (unsigned int l = 0; l < layers_num; l++) {
if ((res = fwscanf(fp, L"%d", &neurons_num)) != 1) {
fclose(fp);
m_status = -2;
return;
} else
m_neurons[l] = neurons_num;
}
for (unsigned int l = 0; l < layers_num - 1; l++)
m_layers.push_back(new AnnLayer(m_neurons[l] + 1, m_neurons[l+1]));
//activation function for hidden/output layers/////////////////////
wchar_t function[256];
for (unsigned int l = 0; l < layers_num; l++) {
if ((res = fwscanf(fp, L"%s", function)) != 1) {
fclose(fp);
m_status = -3;
return;
} else {
if (std::wstring(function) == L"linear")
m_actfuns.push_back(AnnLayer::LINEAR);
else if (std::wstring(function) == L"sigmoid")
m_actfuns.push_back(AnnLayer::SIGMOID);
else if (std::wstring(function) == L"tanh")
m_actfuns.push_back(AnnLayer::TANH);
else
m_actfuns.push_back(AnnLayer::LINEAR);
}
}
//normalization params////////////////////////////////////////////
m_add_vec = new vec2D(1, m_neurons[0]);
m_mul_vec = new vec2D(1, m_neurons[0]);
m_input_vec = new vec2D(1, m_neurons[0]);
for (unsigned int n = 0; n < m_neurons[0]; n++) {
float add, mul;
if ((res = fwscanf(fp, L"%g %g", &add, &mul)) != 2) { //blank network file?
for (unsigned int m = 0; m < m_neurons[0]; m++) {
(*m_add_vec)[m] = 0.0f; //default add = 0
(*m_mul_vec)[m] = 1.0f; //default mult = 1
}
break;
}
(*m_add_vec)[n] = add;
(*m_mul_vec)[n] = mul;
}
//load weights/////////////////////////////////////////////////////
for (unsigned int l = 0; l < m_layers.size(); l++) { //load all weights except input layer
vec2D& tW = *m_layers[l]->m_tW;
for (unsigned int n = 0; n < tW.height(); n++) { //num of Neurons in layer
for (unsigned int i = 0; i < tW.width(); i++) { //num of inputs in Neuron
if ((res = fwscanf(fp, L"%g", &w)) != 1) { //blank network file?
fclose(fp);
m_status = 1;
init_weights((unsigned int)time(0)); //init to random values
return;
} else {
tW(n, i) = w;
}
}
}
}
fclose(fp);
m_status = 0;
} else
m_status = -1;
}
