model dir

src/CMakeLists.txt

@@ -5,6 +5,7 @@ cuda_add_library(marian_lib
   cnpy/cnpy.cpp
   exception.cpp
   expressions.cu 
+  sgd.cu
   tensor.cu	
   tensor_operators.cu
 )

src/sgd.cu

@@ -0,0 +1,68 @@
+#include "sgd.h"
+#include "thrust_functions.h"
+
+namespace marian {
+SGD::SGD(Expr& cost_func, Expr& inX, Expr& inY,
+    const std::vector<Expr*> params, float eta,
+    std::vector<float>& xData, size_t numFeatures,
+    std::vector<float>& yData, size_t numClasses,
+    size_t epochs, size_t batchSize)
+: cost_function_(&cost_func),
+  inX_(&inX),
+  inY_(&inY),
+  params_(params),
+  eta_(eta),
+  xData_(xData),
+  numFeatures_(numFeatures),
+  yData_(yData),
+  numClasses_(numClasses),
+  epochs_(epochs),
+  batchSize_(batchSize)
+{}
+
+void SGD::Run()
+{
+  size_t numExamples = xData_.size()/ numFeatures_;
+  Tensor xt({(int)batchSize_, (int)numExamples}, 0.0f);
+  Tensor yt({(int)batchSize_, (int)numClasses_}, 0.0f);
+
+  for (size_t numEpoch = 0; numEpoch < epochs_; ++numEpoch) {
+    std::cerr << "Starting epoch #" << numEpoch << std::endl;
+    size_t startId = 0;
+    size_t endId = startId + batchSize_;
+
+    while (endId < numExamples) {
+      PrepareBatch(startId, endId, xt, yt);
+      *inX_ = xt;
+      *inY_ = yt;
+
+      cost_function_->forward(batchSize_);
+      cost_function_->backward();
+
+      UpdateModel();
+
+      startId += batchSize_;
+      endId += batchSize_;
+    }
+  }
+}
+
+void SGD::PrepareBatch(size_t startId, size_t endId, Tensor& xt, Tensor& yt) {
+  std::vector<float> x(xData_.begin() + startId * numFeatures_,
+                       xData_.begin() + endId * numFeatures_);
+  std::vector<float> y(yData_.begin() + startId * numClasses_,
+                       yData_.begin() + endId * numClasses_);
+
+  xt.set(x);
+  yt.set(y);
+}
+
+void SGD::UpdateModel() {
+  for (auto& param : params_) {
+    using namespace thrust::placeholders;
+    Element(_1 = _1 - eta_ * _2, param->val(), param->grad());
+  }
+}
+
+} // namespace
+

src/sgd.h

@@ -5,6 +5,7 @@
 
 #include "expressions.h"
 #include "thrust_functions.h"
+#include "tensor_operators.h"
 
 namespace marian {
 
@@ -14,62 +15,13 @@ class SGD {
         const std::vector<Expr*> params, float eta,
         std::vector<float>& xData, size_t numFeatures,
         std::vector<float>& yData, size_t numClasses,
-        size_t epochs, size_t batchSize)
-    : cost_function_(&cost_func),
-      inX_(&inX),
-      inY_(&inY),
-      params_(params),
-      eta_(eta),
-      xData_(xData),
-      numFeatures_(numFeatures),
-      yData_(yData),
-      numClasses_(numClasses),
-      epochs_(epochs),
-      batchSize_(batchSize)
-  {}
+        size_t epochs, size_t batchSize);
 
-    void Run() {
-      size_t numExamples = xData_.size()/ numFeatures_;
-      Tensor xt({(int)batchSize_, (int)numExamples}, 0.0f);
-      Tensor yt({(int)batchSize_, (int)numClasses_}, 0.0f);
+    void Run();
 
-      for (size_t numEpoch = 0; numEpoch < epochs_; ++numEpoch) {
-        std::cerr << "Starting epoch #" << numEpoch << std::endl;
-        size_t startId = 0;
-        size_t endId = startId + batchSize_;
+    void PrepareBatch(size_t startId, size_t endId, Tensor& xt, Tensor& yt);
 
-        while (endId < numExamples) {
-          PrepareBatch(startId, endId, xt, yt);
-          *inX_ = xt;
-          *inY_ = yt;
-
-          cost_function_->forward(batchSize_);
-          cost_function_->backward();
-
-          UpdateModel();
-
-          startId += batchSize_;
-          endId += batchSize_;
-        }
-      }
-    }
-
-    void PrepareBatch(size_t startId, size_t endId, Tensor& xt, Tensor& yt) {
-      std::vector<float> x(xData_.begin() + startId * numFeatures_,
-                           xData_.begin() + endId * numFeatures_);
-      std::vector<float> y(yData_.begin() + startId * numClasses_,
-                           yData_.begin() + endId * numClasses_);
-
-      xt.set(x);
-      yt.set(y);
-    }
-
-    void UpdateModel() {
-      for (auto& param : params_) {
-        using namespace thrust::placeholders;
-        Element(_1 = _1 - eta_ * _2, param->val(), param->grad());
-      }
-    }
+    void UpdateModel();
 
   private:
     std::shared_ptr<Expr> cost_function_;

src/validate_mnist.cu

@@ -22,7 +22,7 @@ int main(int argc, char** argv) {
 
   
   std::cerr << "Loading model params...";
-  NpzConverter converter("../scripts/test_model/model.npz");
+  NpzConverter converter("../scripts/test_model_single/model.npz");
 
   std::vector<float> wData, bData;
   Shape wShape, bShape;