disable batch when sequence training

3 files changed, 15 insertions, 12 deletions

diff --git a/kaldi_seq/Makefile b/kaldi_seq/Makefile
index 7e2dd2e..1232c5a 100644
--- a/kaldi_seq/Makefile
+++ b/kaldi_seq/Makefile
@@ -23,18 +23,21 @@ build: $(OBJ_DIR) $(OBJ_SUBDIR) $(OBJS)
 install: $(LUA_DIR) $(LUA_SUBDIR) $(LIBS)
 
 include $(KDIR)/src/kaldi.mk
-KL := -L/home/slhome/ymz09/mylibs/ -L$(KDIR)/tools/openfst-1.3.4/lib/ $(KDIR)/src/feat/kaldi-feat.a $(KDIR)/src/matrix/kaldi-matrix.a $(KDIR)/src/base/kaldi-base.a $(KDIR)/src/util/kaldi-util.a $(KDIR)/src/hmm/kaldi-hmm.a $(KDIR)/src/tree/kaldi-tree.a $(KDIR)/src/nnet/kaldi-nnet.a $(KDIR)/src/cudamatrix/kaldi-cudamatrix.a $(KDIR)/src/lat/kaldi-lat.a $(KDIR)/src/hmm/kaldi-hmm.a $(KDIR)/src/tree/kaldi-tree.a $(KDIR)/src/matrix/kaldi-matrix.a $(KDIR)/src/util/kaldi-util.a $(KDIR)/src/base/kaldi-base.a -lcblas -llapack_atlas -lfst -lcudart -lcublas -DHAVE_CUDA
+
+KL1 := -rdynamic -Wl,-rpath=$(KDIR)/tools/openfst/lib -L/usr/local/cuda/lib64 -Wl,-rpath,/usr/local/cuda/lib64 -Wl,-rpath=$(KDIR)/src/lib -L.  -L$(KDIR)/src/nnet/   -L$(KDIR)/src/cudamatrix/   -L$(KDIR)/src/lat/   -L$(KDIR)/src/hmm/   -L$(KDIR)/src/tree/   -L$(KDIR)/src/matrix/   -L$(KDIR)/src/util/   -L$(KDIR)/src/base/   $(KDIR)/src/nnet//libkaldi-nnet.so $(KDIR)/src/cudamatrix//libkaldi-cudamatrix.so $(KDIR)/src/lat//libkaldi-lat.so $(KDIR)/src/hmm//libkaldi-hmm.so $(KDIR)/src/tree//libkaldi-tree.so $(KDIR)/src/matrix//libkaldi-matrix.so $(KDIR)/src/util//libkaldi-util.so $(KDIR)/src/base//libkaldi-base.so   -L$(KDIR)/tools/openfst/lib -lfst /usr/lib/liblapack.so /usr/lib/libcblas.so /usr/lib/libatlas.so /usr/lib/libf77blas.so -lm -lpthread -ldl -lcublas -lcudart   -lkaldi-nnet   -lkaldi-cudamatrix   -lkaldi-lat   -lkaldi-hmm   -lkaldi-tree   -lkaldi-matrix   -lkaldi-util   -lkaldi-base
+
+KL2 := -msse -msse2 -Wall -pthread -DKALDI_DOUBLEPRECISION=0 -DHAVE_POSIX_MEMALIGN -Wno-sign-compare -Wno-unused-local-typedefs -Winit-self -DHAVE_EXECINFO_H=1 -rdynamic -DHAVE_CXXABI_H -DHAVE_ATLAS -I$(KDIR)/tools/ATLAS/include -I$(KDIR)/tools/openfst/include -Wno-sign-compare -g  -fPIC -I/usr/local/cuda/include -L/usr/local/cuda/lib64 -DKALDI_NO_EXPF
 
 $(OBJ_DIR) $(LUA_DIR) $(OBJ_SUBDIR) $(LUA_SUBDIR):
 	-mkdir -p $@
 $(LUA_DIR)/%.lua: %.lua
 	cp $< $@
 $(LIBS): $(OBJ_DIR)/src/kaldi_mpe.o $(OBJ_DIR)/src/kaldi_mmi.o $(OBJ_DIR)/init.o $(OBJ_DIR)/src/init.o
-	gcc -shared -fPIC -o $@ $(OBJ_DIR)/src/kaldi_mpe.o $(OBJ_DIR)/src/kaldi_mmi.o $(OBJ_DIR)/init.o $(OBJ_DIR)/src/init.o -lstdc++ -Wl,-rpath=$(LIB_PATH) -L$(LIB_PATH) -lnervcore -lluaT $(KL)
+	gcc -shared -fPIC -o $@ $(OBJ_DIR)/src/kaldi_mpe.o $(OBJ_DIR)/src/kaldi_mmi.o $(OBJ_DIR)/init.o $(OBJ_DIR)/src/init.o -lstdc++ -Wl,-rpath=$(LIB_PATH) -L$(LIB_PATH) -lnervcore -lluaT $(KL1)
 $(OBJ_DIR)/src/kaldi_mpe.o: src/kaldi_mpe.cpp
-	g++ -o $@ -c $< -DHAVE_ATLAS $(KALDIINCLUDE) -g -fPIC $(INCLUDE) -DKALDI_DOUBLEPRECISION=0 -msse2 -DHAVE_POSIX_MEMALIGN
+	g++ -o $@ -c $< $(KALDIINCLUDE) -g -fPIC $(INCLUDE) $(KL2)
 $(OBJ_DIR)/src/kaldi_mmi.o: src/kaldi_mmi.cpp
-	g++ -o $@ -c $< -DHAVE_ATLAS $(KALDIINCLUDE) -g -fPIC $(INCLUDE) -DKALDI_DOUBLEPRECISION=0 -msse2 -DHAVE_POSIX_MEMALIGN
+	g++ -o $@ -c $< $(KALDIINCLUDE) -g -fPIC $(INCLUDE) $(KL2)
 $(OBJ_DIR)/%.o: %.c
 	gcc -o $@ -c $< -g $(INCLUDE) -fPIC
 clean:
diff --git a/kaldi_seq/src/kaldi_mmi.cpp b/kaldi_seq/src/kaldi_mmi.cpp
index a64abd0..ea9b4f1 100644
--- a/kaldi_seq/src/kaldi_mmi.cpp
+++ b/kaldi_seq/src/kaldi_mmi.cpp
@@ -164,12 +164,11 @@ extern "C" {
 
         // Select the GPU
 #if HAVE_CUDA == 1
-        CuDevice::Instantiate().SelectGpuId(use_gpu);
+        CuDevice::Instantiate().SelectGpuId(mmi->use_gpu);
 #endif
 
         // Read the class-frame-counts, compute priors
         mmi->log_prior = new PdfPrior(prior_opts);
-        PdfPrior &log_prior = *(mmi->log_prior);
 
         // Read transition model
         mmi->trans_model = new TransitionModel;
diff --git a/kaldi_seq/src/kaldi_mpe.cpp b/kaldi_seq/src/kaldi_mpe.cpp
index 8cdf010..60384e2 100644
--- a/kaldi_seq/src/kaldi_mpe.cpp
+++ b/kaldi_seq/src/kaldi_mpe.cpp
@@ -205,12 +205,11 @@ extern "C" {
 
         // Select the GPU
 #if HAVE_CUDA == 1
-        CuDevice::Instantiate().SelectGpuId(use_gpu);
+        CuDevice::Instantiate().SelectGpuId(mpe->use_gpu);
 #endif
 
         // Read the class-frame-counts, compute priors
         mpe->log_prior = new PdfPrior(prior_opts);
-        PdfPrior &log_prior = *(mpe->log_prior);
 
         // Read transition model
         mpe->trans_model = new TransitionModel;
@@ -256,7 +255,7 @@ extern "C" {
             return 0;
         }
 
-        assert(sizeof(BaseFloat) == sizeof(float));
+        //assert(sizeof(BaseFloat) == sizeof(float));
         // 1) get the features, numerator alignment
         mpe->ref_ali = mpe->ref_ali_reader->Value(utt);
         long mat_nrow = mat->nrow, mat_ncol = mat->ncol;
@@ -309,8 +308,9 @@ extern "C" {
     NervMatrix * calc_diff_mpe(KaldiMPE * mpe, NervMatrix * mat, const char * key)
     {
         std::string utt(key);
-        assert(sizeof(BaseFloat) == sizeof(float));
+        //assert(sizeof(BaseFloat) == sizeof(float));
 
+        CuMatrix<BaseFloat> nnet_diff;
         kaldi::Matrix<BaseFloat> nnet_out_h;
         nnet_out_h.Resize(mat->nrow, mat->ncol, kUndefined);
 
@@ -327,9 +327,10 @@ extern "C" {
         PdfPriorOptions &prior_opts = *(mpe->prior_opts);
         if (prior_opts.class_frame_counts != "") {
             CuMatrix<BaseFloat> nnet_out;
-            nnet_out.Resize(mat->nrow, mat->ncol, kUndefined);
+            nnet_out.Resize(nnet_out_h.NumRows(), nnet_out_h.NumCols(), kUndefined);
             nnet_out.CopyFromMat(nnet_out_h);
             mpe->log_prior->SubtractOnLogpost(&nnet_out);
+            nnet_out_h.Resize(nnet_out.NumRows(), nnet_out.NumCols(), kUndefined);
             nnet_out.CopyToMat(&nnet_out_h);
             nnet_out.Resize(0,0);
         }
@@ -353,7 +354,6 @@ extern "C" {
         }
 
         // 6) convert the Posterior to a matrix,
-        CuMatrix<BaseFloat> nnet_diff;
         PosteriorToMatrixMapped(post, *(mpe->trans_model), &nnet_diff);
         nnet_diff.Scale(-1.0); // need to flip the sign of derivative,
 
@@ -366,6 +366,7 @@ extern "C" {
             << " frames,"
             << " diff-range(" << nnet_diff.Min() << "," << nnet_diff.Max() << ")";
 
+        nnet_out_h.Resize(nnet_diff.NumRows(), nnet_diff.NumCols(), kUndefined);
         nnet_diff.CopyToMat(&nnet_out_h);
         nnet_diff.Resize(0,0); // release GPU memory,