Merge branch 'master' into marian-integration

3 files changed, 29 insertions, 24 deletions

diff --git a/README.md b/README.md
index 31ba30ca..50f56d92 100644
--- a/README.md
+++ b/README.md
@@ -146,3 +146,8 @@ The setting above uses 8 CPU threads and 4 GPU threads (2 GPUs x 2 threads). The
 ## Acknowledgements
 
 The development of Amunmt received funding from the European Union's Horizon 2020 Research and Innovation Programme under grant agreements 688139 (SUMMA; 2016-2019) and 645487 (Modern MT; 2015-2017) and the Amazon Academic Research Awards program.
+
+
+
+
+
diff --git a/src/amun/cpu/dl4mt/decoder.h b/src/amun/cpu/dl4mt/decoder.h
index 607d41d8..b1c10e97 100644
--- a/src/amun/cpu/dl4mt/decoder.h
+++ b/src/amun/cpu/dl4mt/decoder.h
@@ -182,14 +182,13 @@ class Decoder {
           auto t = blaze::forEach(T1_ + T2_ + T3_, Tanh());
 
           if(!filtered_) {
-            Probs_ = t * w_.W4_;
-            AddBiasVector<byRow>(Probs_, w_.B4_);
+            Probs = t * w_.W4_;
+            AddBiasVector<byRow>(Probs, w_.B4_);
           } else {
-            Probs_ = t * FilteredW4_;
-            AddBiasVector<byRow>(Probs_, FilteredB4_);
+            Probs = t * FilteredW4_;
+            AddBiasVector<byRow>(Probs, FilteredB4_);
           }
-          mblas::Softmax(Probs_);
-          Probs = blaze::forEach(Probs_, Log());
+          LogSoftmax(Probs);
         }
 
         void Filter(const std::vector<size_t>& ids) {
@@ -209,7 +208,6 @@ class Decoder {
         mblas::Matrix T1_;
         mblas::Matrix T2_;
         mblas::Matrix T3_;
-        mblas::Matrix Probs_;
     };
 
   public:
diff --git a/src/amun/cpu/npz_converter.h b/src/amun/cpu/npz_converter.h
index 6bd2f692..7488ead7 100644
--- a/src/amun/cpu/npz_converter.h
+++ b/src/amun/cpu/npz_converter.h
@@ -12,55 +12,55 @@ class NpzConverter {
       public:
         NpyMatrixWrapper(const cnpy::NpyArray& npy)
         : npy_(npy) {}
-        
+
         size_t size() const {
           return size1() * size2();
         }
-        
+
         float* data() const {
           return (float*)npy_.data;
         }
-        
+
         float operator()(size_t i, size_t j) const {
           return ((float*)npy_.data)[i * size2() + j];
         }
-        
+
         size_t size1() const {
           return npy_.shape[0];
         }
-        
+
         size_t size2() const {
           if(npy_.shape.size() == 1)
             return 1;
           else
-            return npy_.shape[1];        
+            return npy_.shape[1];
         }
-        
+
       private:
         const cnpy::NpyArray& npy_;
     };
-  
+
   public:
-    
+
     typedef blaze::CustomMatrix<float, blaze::unaligned,
       blaze::unpadded, blaze::rowMajor> BlazeWrapper;
-        
-    
+
+
     NpzConverter(const std::string& file)
       : model_(cnpy::npz_load(file)),
         destructed_(false) {
       }
-    
+
     ~NpzConverter() {
       if(!destructed_)
         model_.destruct();
     }
-    
+
     void Destruct() {
       model_.destruct();
       destructed_ = true;
     }
-    
+
     mblas::Matrix operator[](const std::string& key) const {
       BlazeWrapper matrix;
       auto it = model_.find(key);
@@ -69,14 +69,16 @@ class NpzConverter {
         matrix = BlazeWrapper(np.data(), np.size1(), np.size2());
       }
       else {
-        std::cerr << "Missing " << key << std::endl; 
+        if (key.find("gamma") == std::string::npos) {
+          std::cerr << "Missing " << key << std::endl;
+        }
       }
 
       mblas::Matrix ret;
       ret = matrix;
       return std::move(ret);
     }
-  
+
     mblas::Matrix operator()(const std::string& key,
                                    bool transpose) const {
       BlazeWrapper matrix;
@@ -91,7 +93,7 @@ class NpzConverter {
       ret = matrix2;
       return std::move(ret);
     }
-  
+
   private:
     cnpy::npz_t model_;
     bool destructed_;