Add marian style tie-embs

src/amun/cpu/dl4mt/model.cpp

@@ -6,10 +6,6 @@ namespace amunmt {
 namespace CPU {
 namespace dl4mt {
 
-Weights::Embeddings::Embeddings(const NpzConverter& model, const std::string &key)
-  : E_(model[key])
-{}
-
 Weights::Embeddings::Embeddings(const NpzConverter& model, const std::vector<std::pair<std::string, bool>> keys)
   : E_(model.getFirstOfMany(keys))
 {}
@@ -67,8 +63,10 @@ Weights::DecSoftmax::DecSoftmax(const NpzConverter& model)
   B2_(model("ff_logit_prev_b", true)),
   W3_(model["ff_logit_ctx_W"]),
   B3_(model("ff_logit_ctx_b", true)),
-  W4_(model.getFirstOfMany({std::pair<std::string, bool>(std::string("ff_logit_W"), false),
+  W4_(model.getFirstOfMany({std::pair<std::string, bool>(
-             std::make_pair(std::string("Wemb_dec"), true)})),
+             std::string("ff_logit_W"), false),
+             std::make_pair(std::string("Wemb_dec"), true),
+             std::make_pair(std::string("Wemb"), true)})),
   B4_(model("ff_logit_b", true)),
   Gamma_0_(model["ff_logit_l1_gamma0"]),
   Gamma_1_(model["ff_logit_l1_gamma1"]),
@@ -78,12 +76,15 @@ Weights::DecSoftmax::DecSoftmax(const NpzConverter& model)
 //////////////////////////////////////////////////////////////////////////////
 
 Weights::Weights(const NpzConverter& model, size_t)
-: encEmbeddings_(model, "Wemb"),
+: encEmbeddings_(model, std::vector<std::pair<std::string, bool>>({
+                         std::make_pair(std::string("Wemb"), false),
+                         std::make_pair(std::string("Wemb_dec"), false)})),
   encForwardGRU_(model, {"encoder_W", "encoder_b", "encoder_U", "encoder_Wx", "encoder_bx",
                          "encoder_Ux", "encoder_gamma1", "encoder_gamma2"}),
   encBackwardGRU_(model, {"encoder_r_W", "encoder_r_b", "encoder_r_U", "encoder_r_Wx",
                           "encoder_r_bx", "encoder_r_Ux", "encoder_r_gamma1", "encoder_r_gamma2"}),
-  decEmbeddings_(model, std::vector<std::pair<std::string, bool>>({std::make_pair(std::string("Wemb_dec"), false),
+  decEmbeddings_(model, std::vector<std::pair<std::string, bool>>({
+                         std::make_pair(std::string("Wemb_dec"), false),
                          std::make_pair(std::string("Wemb"), false)})),
   decInit_(model),
   decGru1_(model, {"decoder_W", "decoder_b", "decoder_U", "decoder_Wx", "decoder_bx", "decoder_Ux",

src/amun/cpu/dl4mt/model.h

@@ -16,7 +16,6 @@ struct Weights {
   //////////////////////////////////////////////////////////////////////////////
 
   struct Embeddings {
-    Embeddings(const NpzConverter& model, const std::string &key);
     Embeddings(const NpzConverter& model, const std::vector<std::pair<std::string, bool>> keys);
 
     const mblas::Matrix E_;

src/amun/cpu/nematus/model.cpp

@@ -60,10 +60,6 @@ std::string Weights::Transition::name(const std::string& prefix, std::string nam
   return prefix + name + infix + "_drt_" + std::to_string(index) + suffix;
 }
 
-Weights::Embeddings::Embeddings(const NpzConverter& model, const std::string &key)
-  : E_(model[key])
-{}
-
 Weights::Embeddings::Embeddings(const NpzConverter& model, const std::vector<std::pair<std::string, bool>> keys)
   : E_(model.getFirstOfMany(keys))
 {}
@@ -143,7 +139,8 @@ Weights::DecSoftmax::DecSoftmax(const NpzConverter& model)
     W3_(model["ff_logit_ctx_W"]),
     B3_(model("ff_logit_ctx_b", true)),
     W4_(model.getFirstOfMany({std::make_pair(std::string("ff_logit_W"), false),
-                              std::make_pair(std::string("Wemb_dec"), true)})),
+                              std::make_pair(std::string("Wemb_dec"), true),
+                              std::make_pair(std::string("Wemb"), true)})),
     B4_(model("ff_logit_b", true)),
     lns_1_(model["ff_logit_lstm_ln_s"]),
     lns_2_(model["ff_logit_prev_ln_s"]),
@@ -156,7 +153,9 @@ Weights::DecSoftmax::DecSoftmax(const NpzConverter& model)
 //////////////////////////////////////////////////////////////////////////////
 
 Weights::Weights(const NpzConverter& model, size_t)
-  : encEmbeddings_(model, "Wemb"),
+  : encEmbeddings_(model, std::vector<std::pair<std::string, bool>>(
+          {std::make_pair(std::string("Wemb"), false),
+           std::make_pair(std::string("Wemb_dec"), false)})),
     decEmbeddings_(model, std::vector<std::pair<std::string, bool>>(
           {std::make_pair(std::string("Wemb_dec"), false),
            std::make_pair(std::string("Wemb"), false)})),

src/amun/cpu/nematus/model.h

@@ -49,7 +49,6 @@ struct Weights {
   };
 
   struct Embeddings {
-    Embeddings(const NpzConverter& model, const std::string &key);
     Embeddings(const NpzConverter& model, const std::vector<std::pair<std::string, bool>> keys);
 
     const mblas::Matrix E_;

src/amun/gpu/dl4mt/model.cu

@@ -5,7 +5,8 @@ namespace GPU {
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 Weights::EncEmbeddings::EncEmbeddings(const NpzConverter& model)
-: E_(model.get("Wemb", true))
+: E_(model.getFirstOfMany({std::make_pair("Wemb", false),
+                           std::make_pair("Wemb_dec", false)}, true))
 {}
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -141,7 +142,8 @@ Weights::DecSoftmax::DecSoftmax(const NpzConverter& model)
   W3_(model.get("ff_logit_ctx_W", true)),
   B3_(model.get("ff_logit_ctx_b", true, true)),
   W4_(model.getFirstOfMany({std::make_pair(std::string("ff_logit_W"), false),
-                            std::make_pair(std::string("Wemb_dec"), true)}, true)),
+                            std::make_pair(std::string("Wemb_dec"), true),
+                            std::make_pair(std::string("Wemb"), true)}, true)),
   B4_(model.get("ff_logit_b", true, true)),
   Gamma_0_(model.get("ff_logit_l1_gamma0", false)),
   Gamma_1_(model.get("ff_logit_l1_gamma1", false)),

src/marian

@@ -1 +1 @@
-Subproject commit f0f06468e1633af503b81eaa8b1fbca881b95257
+Subproject commit 7f3802e97f08e91382c6609cd772c4e80488b4d8