void load_and_apply_model(std::ifstream& model_in, const std::string& model_file,
const StringVector& input_files, const NERConfiguration& ner_config,
bool running_text, bool eval_mode, const std::string& output_format)
{
std::cerr << "Loading model '" << model_file << "'\n";
SimpleLinearCRFModel<ORDER> crf_model(model_in,true);
model_info(crf_model);
// Construct the applier
CRFApplier<ORDER> crf_applier(crf_model,ner_config);
/// Construct the outputter object
NEROneWordPerLineOutputter one_word_per_line_outputter(std::cout);
NERAnnotationOutputter ner_annotation_outputter(std::cout);
JSONOutputter json_outputter(std::cout);
NEROutputterBase* outputter = &one_word_per_line_outputter;
if (output_format == "JSON" || output_format == "json")
outputter = &json_outputter;
//if (!force_tsv_output && ner_config.input_is_running_text())
// outputter = &json_outputter; // &ner_annotation_outputter;
//std::cerr << "Reading in NE lists:";
//load_clue_lists(crf_applier);
//std::cerr << std::endl;
for (unsigned i = 0; i < input_files.size(); ++i) {
std::cerr << "Processing input file '" << input_files[i] << "'" << std::endl;
std::ifstream test_data_in(input_files[i].c_str());
if (!test_data_in) {
std::cerr << "crf-apply: Error opening file '" << input_files[i] << "'" << std::endl;
continue;
}
time_t t0 = clock();
outputter->prolog();
if (eval_mode) {
EvaluationInfo e = crf_applier.evaluation_of(test_data_in,*outputter,running_text);
std::cerr << " Accuracy: " << (e.accuracy()*100) << "%\n";
std::cerr << " Precision: " << (e.precision()*100) << "%\n";
std::cerr << " Recall: " << (e.recall()*100) << "%\n";
}
else {
crf_applier.apply_to(test_data_in,*outputter,running_text);
}
outputter->epilog();
unsigned t = clock() - t0;
std::cerr << " Processed " << crf_applier.processed_tokens() << " tokens in "
<< crf_applier.processed_sequences() << " sequences in " << t << "ms ";
if (t > 0)
std::cerr << "(" << (1000* crf_applier.processed_tokens() / float(t)) << " tokens/s)\n";
else
std::cerr << std::endl;
} // for i
}
int main(int argc, char** argv) {
auto param = InitMnistApps(argc, argv);
param.num_gpus = 2;
int num_gpu = param.num_gpus;
MinervaSystem& ms = MinervaSystem::Instance();
cout << param << endl;
size_t train_data_len = 28 * 28 * param.mb_size / num_gpu; // img size = 28x28
size_t train_label_len = 10 * param.mb_size / num_gpu; // 10 classes
size_t test_data_len = 28 * 28 * param.num_tests; // img size = 28x28
size_t test_label_len = 10 * param.num_tests; // 10 classes
vector<uint64_t> gpus;
for(int i = 0; i < num_gpu; ++i) {
gpus.push_back(ms.CreateGpuDevice(i));
}
MnistMlpAlgo cnn_algo(param);
cnn_algo.Init();
ifstream train_data_in(param.train_data_file.c_str(), ios::binary);
ifstream train_label_in(param.train_label_file.c_str(), ios::binary);
ifstream test_data_in(param.test_data_file.c_str(), ios::binary);
ifstream test_label_in(param.test_label_file.c_str(), ios::binary);
cout << "Start training:" << endl;
for (int epoch = 0; epoch < param.num_epochs; ++epoch) {
train_data_in.clear();
train_data_in.seekg(2 * sizeof(int), ios::beg);
train_label_in.clear();
train_label_in.seekg(2 * sizeof(int), ios::beg);
cout << "Epoch #" << epoch << endl;
for (int mb = 0; mb < param.num_mb; ++mb) {
for (int i = 0; i < num_gpu; ++i) {
ms.SetDevice(gpus[i]); // switch GPU
shared_ptr<float> data_ptr, label_ptr;
tie(data_ptr, label_ptr) = GetNextBatch(train_data_in, train_label_in, train_data_len, train_label_len);
NArray predict = cnn_algo.FF(data_ptr, false);
NArray label = cnn_algo.BP(label_ptr, false);
if (mb % 20 == 0) {
cout << "GPU #" << i << " ";
PrintAccuracy(predict, label, param);
}
}
ms.SetDevice(gpus[0]); // update is on GPU #0
cnn_algo.Update();
}
// Testing
ms.SetDevice(gpus[0]); // test is on GPU #0
cout << "Testing:" << endl;
test_data_in.clear();
test_data_in.seekg(2 * sizeof(int), ios::beg);
test_label_in.clear();
test_label_in.seekg(2 * sizeof(int), ios::beg);
shared_ptr<float> data_ptr, label_ptr;
tie(data_ptr, label_ptr) = GetNextBatch(test_data_in, test_label_in, test_data_len, test_label_len);
NArray predict = cnn_algo.FF(data_ptr, true);
NArray label = cnn_algo.BP(label_ptr, true);
PrintAccuracy(predict, label, param, true);
}
cout << "Training finished" << endl;
return 0;
}