patch.diff

patch.diff

diff --git a/src/mlpack/methods/ann/layer/layer_traits.hpp b/src/mlpack/methods/ann/layer/layer_traits.hpp
index ff4fbf2d3..546dcc9ec 100644
--- a/src/mlpack/methods/ann/layer/layer_traits.hpp
+++ b/src/mlpack/methods/ann/layer/layer_traits.hpp
@@ -98,7 +98,7 @@ HAS_MEM_FUNC(InputHeight, HasInputHeight);
 
 // This gives us a HasRho<T, U> type (where U is a function pointer) we
 // can use with SFINAE to catch when a type has a Rho() function.
-HAS_MEM_FUNC(InputHeight, HasRho);
+HAS_MEM_FUNC(Rho, HasRho);
 
 } // namespace ann
 } // namespace mlpack
diff --git a/src/mlpack/methods/ann/layer/lstm_impl.hpp b/src/mlpack/methods/ann/layer/lstm_impl.hpp
index 8db1f13e7..fad405269 100644
--- a/src/mlpack/methods/ann/layer/lstm_impl.hpp
+++ b/src/mlpack/methods/ann/layer/lstm_impl.hpp
@@ -75,6 +75,10 @@ template<typename eT>
 void LSTM<InputDataType, OutputDataType>::Forward(
     arma::Mat<eT>&& input, arma::Mat<eT>&& output)
 {
+  // std::cout << "rho: " << rho << std::endl;
+
+  // exit(0);
+
   if (!deterministic)
   {
     cellParameter.push_back(prevCell);
diff --git a/src/mlpack/methods/ann/rnn.hpp b/src/mlpack/methods/ann/rnn.hpp
index 4117b4f1d..0fb9af215 100644
--- a/src/mlpack/methods/ann/rnn.hpp
+++ b/src/mlpack/methods/ann/rnn.hpp
@@ -197,6 +197,11 @@ class RNN
   template<typename Archive>
   void Serialize(Archive& ar, const unsigned int /* version */);
 
+  //! Get the maximum number of steps to backpropagate through time (BPTT).
+  size_t Rho() const { return rho; }
+  //! Modify the maximum number of steps to backpropagate through time (BPTT).
+  size_t& Rho() { return rho; }
+
  private:
   // Helper functions.
   /**
@@ -246,6 +251,9 @@ class RNN
   //! Number of steps to backpropagate through time (BPTT).
   size_t rho;
 
+  //! Number of steps to backpropagate through time (BPTT) at the previous step.
+  size_t prevRho;
+
   //! Instantiated outputlayer used to evaluate the network.
   OutputLayerType outputLayer;
 
diff --git a/src/mlpack/methods/ann/rnn_impl.hpp b/src/mlpack/methods/ann/rnn_impl.hpp
index 758ac9fbb..0235174c8 100644
--- a/src/mlpack/methods/ann/rnn_impl.hpp
+++ b/src/mlpack/methods/ann/rnn_impl.hpp
@@ -23,6 +23,7 @@
 #include "visitor/gradient_set_visitor.hpp"
 #include "visitor/gradient_visitor.hpp"
 #include "visitor/weight_set_visitor.hpp"
+#include "visitor/rho_set_visitor.hpp"
 
 namespace mlpack {
 namespace ann /** Artificial Neural Network. */ {
@@ -35,6 +36,7 @@ RNN<OutputLayerType, InitializationRuleType>::RNN(
     OutputLayerType outputLayer,
     InitializationRuleType initializeRule) :
     rho(rho),
+    prevRho(0),
     outputLayer(outputLayer),
     initializeRule(initializeRule),
     inputSize(0),
@@ -176,6 +178,17 @@ template<typename OutputLayerType, typename InitializationRuleType>
 void RNN<OutputLayerType, InitializationRuleType>::SinglePredict(
     const arma::mat& predictors, arma::mat& results)
 {
+  // Update the input length if the backpropagate through time parameter
+  // changed.
+  if (prevRho != rho)
+  {
+    for (size_t i = 1; i < network.size(); ++i)
+      boost::apply_visitor(RhoSetVisitor(rho), network[i]);
+
+    inputSize = predictors.n_elem / rho;
+    prevRho = rho;
+  }
+
   for (size_t seqNum = 0; seqNum < rho; ++seqNum)
   {
     currentInput = predictors.rows(seqNum * inputSize,
@@ -208,10 +221,16 @@ double RNN<OutputLayerType, InitializationRuleType>::Evaluate(
   arma::mat target = arma::mat(responses.colptr(i), responses.n_rows,
       1, false, true);
 
-  if (!inputSize)
+  // Update the input length and target length if the
+  // backpropagate through time parameter changed.
+  if (prevRho != rho)
   {
+    for (size_t i = 1; i < network.size(); ++i)
+      boost::apply_visitor(RhoSetVisitor(rho), network[i]);
+
     inputSize = input.n_elem / rho;
     targetSize = target.n_elem / rho;
+    prevRho = rho;
   }
 
   double performance = 0;
diff --git a/src/mlpack/methods/ann/visitor/CMakeLists.txt b/src/mlpack/methods/ann/visitor/CMakeLists.txt
index f7b3d6ee2..f63b6da40 100644
--- a/src/mlpack/methods/ann/visitor/CMakeLists.txt
+++ b/src/mlpack/methods/ann/visitor/CMakeLists.txt
@@ -37,6 +37,8 @@ set(SOURCES
   reset_visitor_impl.hpp
   reward_set_visitor.hpp
   reward_set_visitor_impl.hpp
+  rho_set_visitor.hpp
+  rho_set_visitor_impl.hpp
   save_output_parameter_visitor.hpp
   save_output_parameter_visitor_impl.hpp
   set_input_height_visitor.hpp
diff --git a/src/mlpack/tests/augmented_rnns_tasks_test.cpp b/src/mlpack/tests/augmented_rnns_tasks_test.cpp
index d9f063472..3dab6758c 100644
--- a/src/mlpack/tests/augmented_rnns_tasks_test.cpp
+++ b/src/mlpack/tests/augmented_rnns_tasks_test.cpp
@@ -267,10 +267,7 @@ BOOST_AUTO_TEST_CASE(LSTMBaselineTest)
 
   model.Add<IdentityLayer<> >();
   model.Add<Linear<> >(inputSize, 20);
-
-  LayerTypes lstm = new LSTM<>(20, 7, maxRho);
-  model.Add(lstm);
-
+  model.Add<LSTM<> >(20, 7, maxRho);
   model.Add<Linear<> >(7, outputSize);
   model.Add<SigmoidLayer<> >();
 
@@ -300,6 +297,11 @@ BOOST_AUTO_TEST_CASE(LSTMBaselineTest)
         std::cout << response.t() << std::endl;
       }*/
 
+      // Update the sequence length. Actually we only have to do this if the
+      // length changed, but for convenience we set the value every time and
+      // the model can check if it really changed.
+      model.Rho() = predictor.n_elem;
+
       model.Train(predictor, response, opt);
     }
   }
@@ -309,6 +311,10 @@ BOOST_AUTO_TEST_CASE(LSTMBaselineTest)
   arma::field<arma::colvec> modelOutput(testSize);
   for (size_t example = 0; example < testSize; ++example) {
     arma::colvec softOutput;
+
+    // Update the sequence length for the prediction part.
+    model.Rho() = testPredictor.at(example).n_elem;
+
     model.Predict(
       testPredictor.at(example),
       softOutput);
@@ -317,13 +323,19 @@ BOOST_AUTO_TEST_CASE(LSTMBaselineTest)
       modelOutput.at(example).at(i) =
         (modelOutput.at(example).at(i)) < 0.5 ? 0 : 1;
     }
-    /*std::cout << "Input:\n";
+
+    // std::cout << "prediction:" << modelOutput.at(example).t() << std::endl;
+    // std::cout << "target:" << modelOutput.at(example).t() << std::endl;
+    std::cout << "Input:\n";
     std::cout << testPredictor.at(example).t() << std::endl;
     std::cout << "Model output:\n";
     std::cout << modelOutput.at(example).t() << std::endl;
     std::cout << "True output:\n";
-    std::cout << testResponse.at(example).t() << std::endl;*/
+    std::cout << testResponse.at(example).t() << std::endl;
+
+    std::cout << "=======================================\n";
   }
+
   std::cout << "Final score: "
        << SequencePrecision<arma::colvec>(testResponse, modelOutput)
        << "\n";