apaam/mlpack
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Added CopyTask input representation with unary representation of repeat count

Browse Source
This commit is contained in:
Konstantin Sidorov
2017-07-31 12:36:22 +03:00
parent f8365d3c09
commit f94f8d8c1c
1 changed files with 18 additions and 100 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/mlpack/tests/augmented_rnns_tasks_test.cpp
+18 -100
View File
@@ -262,111 +262,22 @@ BOOST_AUTO_TEST_CASE(AddTaskTest) {
BOOST_REQUIRE(ok);
}
/*
BOOST_AUTO_TEST_CASE(LSTMBaselineTestCopy)
{
ofstream fout;
fout.open("output.log");
fout << "Report for aligned representation.\n";
fout << "Training epochs = " << 20 << "\n";
int maxNRepeat[] = {10, 10, 10, 6, 4, 3, 3, 3, 2, 2, 2};
for (size_t maxLen = 2; maxLen <= 8; ++maxLen) {
for (size_t nRepeats = 1; nRepeats <= maxNRepeat[maxLen]; ++nRepeats) {
bool ok = true;
const size_t outputSize = 1;
const size_t inputSize = 1;
const size_t rho = 2;
const size_t maxRho = 128;
RNN<MeanSquaredError<> > model(rho);
model.Add<IdentityLayer<> >();
model.Add<Linear<> >(inputSize, 30);
model.Add<LSTM<> >(30, 15, maxRho);
model.Add<LeakyReLU<> >();
model.Add<Linear<> >(15, outputSize);
model.Add<SigmoidLayer<> >();
Adam<decltype(model)> opt(model);
CopyTask task(maxLen, nRepeats);
arma::field<arma::colvec> trainPredictor, trainResponse;
size_t trainSize = 15 + 5 * maxLen;
task.Generate(trainPredictor, trainResponse, trainSize);
size_t testSize = 15 + 5 * maxLen;
arma::field<arma::colvec> testPredictor, testResponse;
task.Generate(testPredictor, testResponse, testSize);
for (size_t epoch = 0; epoch < 20; ++epoch) {
for (size_t example = 0; example < trainPredictor.n_elem; ++example) {
size_t totSize =
trainPredictor.at(example).n_elem + trainResponse.at(example).n_elem;
arma::mat predictor = arma::zeros(totSize, 1);
predictor.col(0).rows(0,trainPredictor.at(example).n_elem-1) =
trainPredictor.at(example);
//predictor.col(1) = arma::ones(totSize);
arma::mat response = arma::zeros(totSize, 1);
response.col(0).rows(trainPredictor.at(example).n_elem,totSize-1) =
trainResponse.at(example);
model.Rho() = totSize;
model.Train(predictor, response, opt);
}
std::cerr << "Finished running training epoch #"
<< epoch+1 << "\n";
}
arma::field<arma::colvec> modelOutput(testSize);
for (size_t example = 0; example < testSize; ++example) {
arma::colvec softOutput;
size_t totSize =
testPredictor.at(example).n_elem + testResponse.at(example).n_elem;
arma::mat predictor = arma::zeros(totSize, 1);
predictor.col(0).rows(0,testPredictor.at(example).n_elem-1) =
testPredictor.at(example);
//predictor.col(1) = arma::ones(totSize);
model.Rho() = predictor.n_rows;
model.Predict(
predictor,
softOutput);
modelOutput.at(example) = softOutput.rows(
testPredictor.at(example).n_elem,
softOutput.n_rows-1);
Binarize<double>(modelOutput.at(example), modelOutput.at(example), 0.5);
// TODO Check this one!
std::cerr << "Predictor:\n"
<< predictor
<< "Model response:\n"
<< softOutput;
}
std::cerr << "Final score for ("
<< maxLen << ","
<< nRepeats << "): "
<< SequencePrecision<arma::colvec>(testResponse, modelOutput)
<< "\n";
fout << "Final score for ("
<< maxLen << ","
<< nRepeats << "): "
<< SequencePrecision<arma::colvec>(testResponse, modelOutput)
<< "\n";
fout << "Sample size = " << trainSize << "\n";
fout.flush();
}
}
fout.close();
}*/
BOOST_AUTO_TEST_CASE(LSTMBaselineTestCopyRepeatRepr)
{
ofstream fout;
fout.open("output-aug.log");
fout << "Report for augmented representation.\n";
fout << "Training epochs = " << 20 << "\n";
int maxNRepeat[] = {10, 10, 10, 6, 4, 3, 3, 3, 2, 2, 2};
for (size_t maxLen = 2; maxLen <= 8; ++maxLen) {
for (size_t nRepeats = 1; nRepeats <= maxNRepeat[maxLen]; ++nRepeats) {
bool ok = true;
const size_t outputSize = 1;
const size_t inputSize = 1;
const size_t inputSize = 2;
const size_t rho = 2;
const size_t maxRho = 128;
@@ -398,9 +309,8 @@ BOOST_AUTO_TEST_CASE(LSTMBaselineTestCopyRepeatRepr)
trainPredictor.at(example);
predictor.col(1).rows(trainPredictor.at(example).n_elem,totSize-1) =
arma::ones(totSize-trainPredictor.at(example).n_elem);
// You can't do this here...
/* predictor = predictor.t();
predictor.reshape(predictor.n_elem, 1);*/
predictor = predictor.t();
predictor.reshape(predictor.n_elem, 1);
arma::mat response = arma::zeros(totSize, 1);
response.col(0).rows(trainPredictor.at(example).n_elem,totSize-1) =
trainResponse.at(example);
@@ -422,10 +332,9 @@ BOOST_AUTO_TEST_CASE(LSTMBaselineTestCopyRepeatRepr)
assert(predictor.n_rows == totSize);
predictor.col(1).rows(testPredictor.at(example).n_elem, totSize-1) =
arma::ones(totSize-testPredictor.at(example).n_elem);
// ... but you *must* do it here.
predictor = predictor.t();
predictor.reshape(predictor.n_elem, 1);
model.Rho() = predictor.n_rows;
model.Rho() = totSize;
model.Predict(
predictor,
softOutput);
@@ -447,8 +356,17 @@ BOOST_AUTO_TEST_CASE(LSTMBaselineTestCopyRepeatRepr)
<< nRepeats << "): "
<< SequencePrecision<arma::colvec>(testResponse, modelOutput)
<< "\n";
fout << "Final score for ("
<< maxLen << ","
<< nRepeats << "): "
<< SequencePrecision<arma::colvec>(testResponse, modelOutput)
<< "\n";
fout << "Sample size = " << trainSize << "\n";
fout.flush();
}
}
fout.close();
}
arma::field<arma::colvec> binarizeAdd(arma::field<arma::colvec> data) {