support for networks with single hidden layer

(set num_hidden to 0; num_output_embeddings will be size of remaining layer)

5 files changed, 112 insertions, 43 deletions

diff --git a/src/model.cpp b/src/model.cpp
index 3611975..589a52e 100644
--- a/src/model.cpp
+++ b/src/model.cpp
@@ -20,10 +20,18 @@ namespace nplm
         int output_embedding_dimension)
 {
     input_layer.resize(input_vocab_size, input_embedding_dimension, ngram_size-1);
-    first_hidden_linear.resize(num_hidden, input_embedding_dimension*(ngram_size-1));
-    first_hidden_activation.resize(num_hidden);
-    second_hidden_linear.resize(output_embedding_dimension, num_hidden);
-    second_hidden_activation.resize(output_embedding_dimension);
+    if (num_hidden == 0) {
+        first_hidden_linear.resize(output_embedding_dimension, input_embedding_dimension*(ngram_size-1));
+        first_hidden_activation.resize(output_embedding_dimension);
+        second_hidden_linear.resize(1,1);
+        second_hidden_activation.resize(1);
+    }
+    else {
+        first_hidden_linear.resize(num_hidden, input_embedding_dimension*(ngram_size-1));
+        first_hidden_activation.resize(num_hidden);
+        second_hidden_linear.resize(output_embedding_dimension, num_hidden);
+        second_hidden_activation.resize(output_embedding_dimension);
+    }
     output_layer.resize(output_vocab_size, output_embedding_dimension);
     this->ngram_size = ngram_size;
     this->input_vocab_size = input_vocab_size;
@@ -48,7 +56,12 @@ void model::premultiply()
     // we can multiply them into a single linear layer *if* we are not training
     int context_size = ngram_size-1;
     Matrix<double,Dynamic,Dynamic> U = first_hidden_linear.U;
-    first_hidden_linear.U.resize(num_hidden, input_vocab_size * context_size);
+    if (num_hidden == 0) {
+        first_hidden_linear.U.resize(output_embedding_dimension, input_vocab_size * context_size);
+    }
+    else {
+        first_hidden_linear.U.resize(num_hidden, input_vocab_size * context_size);
+    }
     for (int i=0; i<context_size; i++)
         first_hidden_linear.U.middleCols(i*input_vocab_size, input_vocab_size) = U.middleCols(i*input_embedding_dimension, input_embedding_dimension) * input_layer.W->transpose();
     input_layer.W->resize(1,1); // try to save some memory
diff --git a/src/neuralLM.h b/src/neuralLM.h
index 0b8b72f..f451e8a 100644
--- a/src/neuralLM.h
+++ b/src/neuralLM.h
@@ -177,13 +177,19 @@ public:
 	if (normalization)
 	{
 	    Eigen::Matrix<double,Eigen::Dynamic,1> scores(shared->output_vocab.size());
-	    prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
+            if (prop.skip_hidden)
+                prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, scores);
+            else
+                prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
 	    double logz = logsum(scores.col(0));
 	    log_prob = weight * (scores(output, 0) - logz);
 	}
 	else
 	{
-	    log_prob = weight * prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, output, 0);
+            if (prop.skip_hidden)
+                log_prob = weight * prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, output, 0);
+            else
+                log_prob = weight * prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, output, 0);
 	}
 	stop_timer(3);
 
@@ -216,7 +222,10 @@ public:
 	if (normalization)
 	{
 	    Eigen::Matrix<double,Eigen::Dynamic,Eigen::Dynamic> scores(shared->output_vocab.size(), ngram.cols());
-	    prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
+            if (prop.skip_hidden)
+                prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, scores);
+            else
+                prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
 
 	    // And softmax and loss
 	    Matrix<double,Dynamic,Dynamic> output_probs(shared->nn.output_vocab_size, ngram.cols());
@@ -233,7 +242,10 @@ public:
 	    for (int j=0; j<ngram.cols(); j++)
 	    {
 	        int output = ngram(ngram_size-1, j);
-	        log_probs(0, j) = weight * prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, output, j);
+                if (prop.skip_hidden)
+                    log_probs(0, j) = weight * prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, output, j);
+                else
+                    log_probs(0, j) = weight * prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, output, j);
 	    }
 	}
     }
diff --git a/src/propagator.h b/src/propagator.h
index 0619de6..c52a6a9 100644
--- a/src/propagator.h
+++ b/src/propagator.h
@@ -23,6 +23,7 @@ public:
     Node<Linear_layer> second_hidden_linear_node;
     Node<Activation_function> second_hidden_activation_node;
     Node<Output_word_embeddings> output_layer_node;
+    bool skip_hidden;
 
 public:
     propagator () : minibatch_size(0), pnn(0) { }
@@ -39,6 +40,7 @@ public:
 	output_layer_node(const_cast<Output_word_embeddings*>(&nn.output_layer), minibatch_size),
 	minibatch_size(minibatch_size)
     {
+        skip_hidden = (nn.num_hidden == 0);
     }
 
     // This must be called if the underlying model is resized.
@@ -82,12 +84,14 @@ public:
 	stop_timer(1);
     
 
+        if (!skip_hidden) {
 	start_timer(2);
 	second_hidden_linear_node.param->fProp(first_hidden_activation_node.fProp_matrix,
 					       second_hidden_linear_node.fProp_matrix);
 	second_hidden_activation_node.param->fProp(second_hidden_linear_node.fProp_matrix,
 						   second_hidden_activation_node.fProp_matrix);
 	stop_timer(2);
+        }
 
 	// The propagation stops here because the last layer is very expensive.
     }
@@ -106,9 +110,16 @@ public:
 	stop_timer(7);
 	
 	start_timer(8);
+        if (skip_hidden) {
+            output_layer_node.param->computeGradient(first_hidden_activation_node.fProp_matrix,
+                                                    output,
+                                                    learning_rate, momentum);
+        }
+        else {
 	output_layer_node.param->computeGradient(second_hidden_activation_node.fProp_matrix,
 						 output,
 						 learning_rate, momentum);
+        }
 	stop_timer(8);
 
 	bPropRest(data, learning_rate, momentum, L2_reg);
@@ -130,9 +141,16 @@ public:
 	
 
 	start_timer(8);
-	output_layer_node.param->computeGradient(second_hidden_activation_node.fProp_matrix,
-						 samples, weights,
-						 learning_rate, momentum);
+        if (skip_hidden) {
+            output_layer_node.param->computeGradient(first_hidden_activation_node.fProp_matrix,
+                                                    samples, weights,
+                                                    learning_rate, momentum);
+        }
+        else {
+            output_layer_node.param->computeGradient(second_hidden_activation_node.fProp_matrix,
+                                                    samples, weights,
+                                                    learning_rate, momentum);
+        }
 	stop_timer(8);
 
 	bPropRest(data, learning_rate, momentum, L2_reg);
@@ -145,33 +163,46 @@ private:
     {
 	// Second hidden layer
 
-        start_timer(9);
-	second_hidden_activation_node.param->bProp(output_layer_node.bProp_matrix,
-						   second_hidden_activation_node.bProp_matrix,
-						   second_hidden_linear_node.fProp_matrix,
-						   second_hidden_activation_node.fProp_matrix);
-
-	second_hidden_linear_node.param->bProp(second_hidden_activation_node.bProp_matrix,
-					       second_hidden_linear_node.bProp_matrix);
-	stop_timer(9);
-
-	start_timer(10);
-	second_hidden_linear_node.param->computeGradient(second_hidden_activation_node.bProp_matrix,
-							 first_hidden_activation_node.fProp_matrix,
-							 learning_rate, momentum, L2_reg);
-	stop_timer(10);
-
-	// First hidden layer
-
-	start_timer(11);
-	first_hidden_activation_node.param->bProp(second_hidden_linear_node.bProp_matrix,
-						  first_hidden_activation_node.bProp_matrix,
-						  first_hidden_linear_node.fProp_matrix,
-						  first_hidden_activation_node.fProp_matrix);
-
-	first_hidden_linear_node.param->bProp(first_hidden_activation_node.bProp_matrix,
-					      first_hidden_linear_node.bProp_matrix);
-	stop_timer(11);
+        if (skip_hidden) {
+            start_timer(9);
+            first_hidden_activation_node.param->bProp(output_layer_node.bProp_matrix,
+                                                    first_hidden_activation_node.bProp_matrix,
+                                                    first_hidden_linear_node.fProp_matrix,
+                                                    first_hidden_activation_node.fProp_matrix);
+
+            first_hidden_linear_node.param->bProp(first_hidden_activation_node.bProp_matrix,
+                                                first_hidden_linear_node.bProp_matrix);
+            stop_timer(9);
+        }
+        else {
+            start_timer(9);
+            second_hidden_activation_node.param->bProp(output_layer_node.bProp_matrix,
+                                                    second_hidden_activation_node.bProp_matrix,
+                                                    second_hidden_linear_node.fProp_matrix,
+                                                    second_hidden_activation_node.fProp_matrix);
+
+            second_hidden_linear_node.param->bProp(second_hidden_activation_node.bProp_matrix,
+                                                second_hidden_linear_node.bProp_matrix);
+            stop_timer(9);
+
+            start_timer(10);
+            second_hidden_linear_node.param->computeGradient(second_hidden_activation_node.bProp_matrix,
+                                                            first_hidden_activation_node.fProp_matrix,
+                                                            learning_rate, momentum, L2_reg);
+            stop_timer(10);
+
+            // First hidden layer
+
+            start_timer(11);
+            first_hidden_activation_node.param->bProp(second_hidden_linear_node.bProp_matrix,
+                                                    first_hidden_activation_node.bProp_matrix,
+                                                    first_hidden_linear_node.fProp_matrix,
+                                                    first_hidden_activation_node.fProp_matrix);
+
+            first_hidden_linear_node.param->bProp(first_hidden_activation_node.bProp_matrix,
+                                                first_hidden_linear_node.bProp_matrix);
+            stop_timer(11);
+        }
 	
 	start_timer(12);
 	first_hidden_linear_node.param->computeGradient(first_hidden_activation_node.bProp_matrix,
diff --git a/src/testNeuralNetwork.cpp b/src/testNeuralNetwork.cpp
index f20fff9..1b4820e 100644
--- a/src/testNeuralNetwork.cpp
+++ b/src/testNeuralNetwork.cpp
@@ -101,7 +101,10 @@ int main (int argc, char *argv[])
 	prop.fProp(minibatch.topRows(myParam.ngram_size-1));
 
 	// Do full forward prop through output word embedding layer
-	prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
+        if (prop.skip_hidden)
+            prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, scores);
+        else
+            prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
 
 	// And softmax and loss
 	double minibatch_log_likelihood;
diff --git a/src/trainNeuralNetwork.cpp b/src/trainNeuralNetwork.cpp
index a088e16..57323d9 100644
--- a/src/trainNeuralNetwork.cpp
+++ b/src/trainNeuralNetwork.cpp
@@ -442,7 +442,11 @@ int main(int argc, char** argv)
 
 		// Final forward propagation step (sparse)
 		start_timer(4);
-		prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix,
+                if (prop.skip_hidden)
+                    prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix,
+                                                    minibatch_samples, scores);
+                else
+                    prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix,
 						    minibatch_samples, scores);
 		stop_timer(4);
 
@@ -491,7 +495,10 @@ int main(int argc, char** argv)
 	    {
 	        ///// Standard log-likelihood
 	        start_timer(4);
-		prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
+                if (prop.skip_hidden)
+                    prop.output_layer_node.param->fProp(prop.first_hidden_activation_node.fProp_matrix, scores);
+                else
+                    prop.output_layer_node.param->fProp(prop.second_hidden_activation_node.fProp_matrix, scores);
 		stop_timer(4);
 
 		double minibatch_log_likelihood;
@@ -566,7 +573,10 @@ int main(int argc, char** argv)
 
 		// Do full forward prop through output word embedding layer
 		start_timer(4);
-		prop_validation.output_layer_node.param->fProp(prop_validation.second_hidden_activation_node.fProp_matrix, scores);
+                if (prop_validation.skip_hidden)
+                    prop_validation.output_layer_node.param->fProp(prop_validation.first_hidden_activation_node.fProp_matrix, scores);
+                else
+                    prop_validation.output_layer_node.param->fProp(prop_validation.second_hidden_activation_node.fProp_matrix, scores);
 		stop_timer(4);
 
 		// And softmax and loss. Be careful of short minibatch