Just come back, let's merge the new master

Merge branch 'master' into txh18/rnnlm

35 files changed, 1026 insertions, 715 deletions

diff --git a/nerv/Makefile b/nerv/Makefile
index 022e2fb..b449f82 100644
--- a/nerv/Makefile
+++ b/nerv/Makefile
@@ -30,14 +30,14 @@ LUAT_OBJS := $(addprefix $(OBJ_DIR)/,$(LUAT_OBJS))
 OBJS := $(CORE_OBJS) $(NERV_OBJS) $(LUAT_OBJS)
 LIBS := $(INST_LIBDIR)/libnerv.so $(LIB_PATH)/libnervcore.so $(LIB_PATH)/libluaT.so
 LUA_LIBS := matrix/init.lua io/init.lua init.lua \
-			layer/init.lua layer/affine.lua layer/sigmoid.lua layer/softmax_ce.lua \
-			layer/window.lua layer/bias.lua layer/combiner.lua layer/mse.lua layer/affine_recurrent.lua\
+			layer/init.lua layer/affine.lua layer/sigmoid.lua layer/softmax_ce.lua layer/softmax.lua \
+			layer/window.lua layer/bias.lua layer/combiner.lua layer/mse.lua layer/affine_recurrent.lua \
 			nn/init.lua nn/layer_repo.lua nn/param_repo.lua nn/layer_dag.lua \
 			io/sgd_buffer.lua 
 
 INCLUDE := -I $(LUA_INCDIR) -DLUA_USE_APICHECK
-CUDA_BASE := /usr/local/cuda-6.5
-#CUDA_BASE := /usr/local/cuda-5.0
+#CUDA_BASE := /usr/local/cuda-7.0
+CUDA_BASE := /usr/local/cuda
 CUDA_INCLUDE := -I $(CUDA_BASE)/include/
 INCLUDE += $(CUDA_INCLUDE)
 
@@ -66,7 +66,7 @@ $(LIB_PATH)/libluaT.so: $(LUAT_OBJS)
 $(INST_LIBDIR)/libnerv.so: $(NERV_OBJS) $(LIB_PATH)/libnervcore.so $(LIB_PATH)/libluaT.so
 	gcc -shared -o $@ $(NERV_OBJS) $(LDFLAGS) -Wl,-rpath=$(LIB_PATH) -L$(LIB_PATH) -lnervcore -lluaT
 
-$(OBJ_DIR)/matrix/cumatrix.o: matrix/generic/cumatrix.c matrix/generic/matrix.c matrix/generic/cukernel.cu
+$(OBJ_DIR)/matrix/cumatrix.o: matrix/generic/cumatrix.c matrix/generic/matrix.c
 $(OBJ_DIR)/matrix/mmatrix.o: matrix/generic/mmatrix.c matrix/generic/matrix.c
 
 $(OBJ_DIR)/lib/matrix/cumatrix.o: lib/matrix/generic/cumatrix.c lib/matrix/generic/matrix.c lib/matrix/generic/cukernel.cu
diff --git a/nerv/examples/asr_trainer.lua b/nerv/examples/asr_trainer.lua
index 4fa4096..69cfeed 100644
--- a/nerv/examples/asr_trainer.lua
+++ b/nerv/examples/asr_trainer.lua
@@ -1,9 +1,9 @@
 function build_trainer(ifname)
     local param_repo = nerv.ParamRepo()
     param_repo:import(ifname, nil, gconf)
-    local sublayer_repo = make_sublayer_repo(param_repo)
-    local layer_repo = make_layer_repo(sublayer_repo, param_repo)
+    local layer_repo = make_layer_repo(param_repo)
     local network = get_network(layer_repo)
+    local global_transf = get_global_transf(layer_repo)
     local input_order = get_input_order()
     local iterative_trainer = function (prefix, scp_file, bp)
         gconf.randomize = bp
@@ -12,28 +12,41 @@ function build_trainer(ifname)
         -- initialize the network
         network:init(gconf.batch_size)
         gconf.cnt = 0
-        err_input = {nerv.CuMatrixFloat(256, 1)}
+        err_input = {nerv.CuMatrixFloat(gconf.batch_size, 1)}
         err_input[1]:fill(1)
         for data in buffer.get_data, buffer do
             -- prine stat periodically
             gconf.cnt = gconf.cnt + 1
             if gconf.cnt == 1000 then
-                print_stat(sublayer_repo)
+                print_stat(layer_repo)
                 nerv.CuMatrix.print_profile()
                 nerv.CuMatrix.clear_profile()
                 gconf.cnt = 0
                 -- break
             end
             local input = {}
---            if gconf.cnt == 100 then break end
-            for i, id in ipairs(input_order) do
+--            if gconf.cnt == 1000 then break end
+            for i, e in ipairs(input_order) do
+                local id = e.id
                 if data[id] == nil then
                     nerv.error("input data %s not found", id)
                 end
-                table.insert(input, data[id])
+                local transformed
+                if e.global_transf then
+                    transformed = nerv.speech_utils.global_transf(data[id],
+                                        global_transf,
+                                        gconf.frm_ext or 0, 0,
+                                        gconf)
+                else
+                    transformed = data[id]
+                end
+                table.insert(input, transformed)
+            end
+            local output = {nerv.CuMatrixFloat(gconf.batch_size, 1)}
+            err_output = {}
+            for i = 1, #input do
+                table.insert(err_output, input[i]:create())
             end
-            local output = {nerv.CuMatrixFloat(256, 1)}
-            err_output = {input[1]:create()}
             network:propagate(input, output)
             if bp then
                 network:back_propagate(err_input, err_output, input, output)
@@ -42,16 +55,16 @@ function build_trainer(ifname)
             -- collect garbage in-time to save GPU memory
             collectgarbage("collect")
         end
-        print_stat(sublayer_repo)
+        print_stat(layer_repo)
         nerv.CuMatrix.print_profile()
         nerv.CuMatrix.clear_profile()
         if (not bp) and prefix ~= nil then
             nerv.info("writing back...")
             local fname = string.format("%s_cv%.3f.nerv",
-                            prefix, get_accuracy(sublayer_repo))
+                            prefix, get_accuracy(layer_repo))
             network:get_params():export(fname, nil)
         end
-        return get_accuracy(sublayer_repo)
+        return get_accuracy(layer_repo)
     end
     return iterative_trainer
 end
diff --git a/nerv/examples/mmi_chime3.lua b/nerv/examples/mmi_chime3.lua
new file mode 100644
index 0000000..6ac7f28
--- /dev/null
+++ b/nerv/examples/mmi_chime3.lua
@@ -0,0 +1,183 @@
+require 'kaldi_io'
+require 'kaldi_seq'
+gconf = {lrate = 0.00001, wcost = 0, momentum = 0.0,
+        cumat_type = nerv.CuMatrixFloat,
+        mmat_type = nerv.MMatrixFloat,
+        frm_ext = 5,
+        tr_scp = "ark,o:/slfs6/users/ymz09/kaldi/src/featbin/copy-feats scp:/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_mmi/train.scp ark:- |",
+        initialized_param = {"/slfs6/users/ymz09/nerv-project/nerv/nerv-speech/kaldi_seq/test/chime3_init_mmi.nerv",
+        "/slfs6/users/ymz09/nerv-project/nerv/nerv-speech/kaldi_seq/test/chime3_global_transf_mmi.nerv"},
+        debug = false}
+
+function make_layer_repo(param_repo)
+    local layer_repo = nerv.LayerRepo(
+    {
+        -- global transf
+        ["nerv.BiasLayer"] =
+        {
+            blayer1 = {{bias = "bias1"}, {dim_in = {440}, dim_out = {440}}},
+            blayer2 = {{bias = "bias2"}, {dim_in = {440}, dim_out = {440}}}
+        },
+        ["nerv.WindowLayer"] =
+        {
+            wlayer1 = {{window = "window1"}, {dim_in = {440}, dim_out = {440}}},
+            wlayer2 = {{window = "window2"}, {dim_in = {440}, dim_out = {440}}}
+        },
+        -- biased linearity
+        ["nerv.AffineLayer"] =
+        {
+            affine0 = {{ltp = "affine0_ltp", bp = "affine0_bp"},
+            {dim_in = {440}, dim_out = {2048}}},
+            affine1 = {{ltp = "affine1_ltp", bp = "affine1_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine2 = {{ltp = "affine2_ltp", bp = "affine2_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine3 = {{ltp = "affine3_ltp", bp = "affine3_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine4 = {{ltp = "affine4_ltp", bp = "affine4_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine5 = {{ltp = "affine5_ltp", bp = "affine5_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine6 = {{ltp = "affine6_ltp", bp = "affine6_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine7 = {{ltp = "affine7_ltp", bp = "affine7_bp"},
+            {dim_in = {2048}, dim_out = {2011}}}
+        },
+        ["nerv.SigmoidLayer"] =
+        {
+            sigmoid0 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid1 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid2 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid3 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid4 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid5 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid6 = {{}, {dim_in = {2048}, dim_out = {2048}}}
+        },
+        ["nerv.MMILayer"] =
+        {
+            mmi_crit = {{}, {dim_in = {2011, -1}, dim_out = {1},
+                        cmd = {
+                            arg = "--class-frame-counts=/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced/ali_train_pdf.counts --acoustic-scale=0.1 --lm-scale=1.0 --learn-rate=0.00001 --drop-frames=true --verbose=1",
+                            mdl = "/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_ali/final.mdl",
+                            lat = "scp:/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_denlats/lat.scp",
+                            ali = "ark:gunzip -c /slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_ali/ali.*.gz |"
+                        }
+                    }
+                }
+        },
+        ["nerv.SoftmaxLayer"] = -- softmax for decode output
+        {
+            softmax = {{}, {dim_in = {2011}, dim_out = {2011}}}
+        }
+    }, param_repo, gconf)
+
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            global_transf = {{}, {
+                dim_in = {440}, dim_out = {440},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "blayer1[1]",
+                    ["blayer1[1]"] = "wlayer1[1]",
+                    ["wlayer1[1]"] = "blayer2[1]",
+                    ["blayer2[1]"] = "wlayer2[1]",
+                    ["wlayer2[1]"] = "<output>[1]"
+                }
+            }},
+            main = {{}, {
+                dim_in = {440}, dim_out = {2011},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "affine0[1]",
+                    ["affine0[1]"] = "sigmoid0[1]",
+                    ["sigmoid0[1]"] = "affine1[1]",
+                    ["affine1[1]"] = "sigmoid1[1]",
+                    ["sigmoid1[1]"] = "affine2[1]",
+                    ["affine2[1]"] = "sigmoid2[1]",
+                    ["sigmoid2[1]"] = "affine3[1]",
+                    ["affine3[1]"] = "sigmoid3[1]",
+                    ["sigmoid3[1]"] = "affine4[1]",
+                    ["affine4[1]"] = "sigmoid4[1]",
+                    ["sigmoid4[1]"] = "affine5[1]",
+                    ["affine5[1]"] = "sigmoid5[1]",
+                    ["sigmoid5[1]"] = "affine6[1]",
+                    ["affine6[1]"] = "sigmoid6[1]",
+                    ["sigmoid6[1]"] = "affine7[1]",
+                    ["affine7[1]"] = "<output>[1]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            mmi_output = {{}, {
+                dim_in = {440, -1}, dim_out = {1},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "mmi_crit[1]",
+                    ["<input>[2]"] = "mmi_crit[2]",
+                    ["mmi_crit[1]"] = "<output>[1]"
+                }
+            }},
+            softmax_output = {{}, {
+                dim_in = {440}, dim_out = {2011},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "softmax[1]",
+                    ["softmax[1]"] = "<output>[1]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    return layer_repo
+end
+
+function get_network(layer_repo)
+    return layer_repo:get_layer("mmi_output")
+end
+
+function get_decode_network(layer_repo)
+    return layer_repo:get_layer("softmax_output")
+end
+
+function get_global_transf(layer_repo)
+    return layer_repo:get_layer("global_transf")
+end
+
+function make_readers(feature_rspecifier, layer_repo)
+    return {
+                {reader = nerv.KaldiReader(gconf,
+                    {
+                        id = "main_scp",
+                        feature_rspecifier = feature_rspecifier,
+                        frm_ext = gconf.frm_ext,
+                        global_transf = layer_repo:get_layer("global_transf"),
+                        mlfs = {}
+                    })
+                }
+            }
+end
+
+function get_input_order()
+    return {{id = "main_scp", global_transf = true},
+            {id = "key"}}
+end
+
+function get_accuracy(layer_repo)
+    return 0
+end
+
+function print_stat(layer_repo)
+    local mmi_crit = layer_repo:get_layer("mmi_crit")
+    nerv.info("*** training stat begin ***")
+    nerv.printf("frames:\t\t\t%d\n", mmi_crit.total_frames)
+    nerv.info("*** training stat end ***")
+end
diff --git a/nerv/examples/mpe_chime3.lua b/nerv/examples/mpe_chime3.lua
new file mode 100644
index 0000000..ec095b0
--- /dev/null
+++ b/nerv/examples/mpe_chime3.lua
@@ -0,0 +1,186 @@
+require 'kaldi_io'
+require 'kaldi_seq'
+gconf = {lrate = 0.00001, wcost = 0, momentum = 0.0,
+        cumat_type = nerv.CuMatrixFloat,
+        mmat_type = nerv.MMatrixFloat,
+        frm_ext = 5,
+        tr_scp = "ark,s,cs:/slfs6/users/ymz09/kaldi/src/featbin/copy-feats scp:/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_smbr/train.scp ark:- |",
+        initialized_param = {"/slfs6/users/ymz09/nerv-project/nerv/nerv-speech/kaldi_seq/test/chime3_init.nerv",
+        "/slfs6/users/ymz09/nerv-project/nerv/nerv-speech/kaldi_seq/test/chime3_global_transf.nerv"},
+        debug = false}
+
+function make_layer_repo(param_repo)
+    local layer_repo = nerv.LayerRepo(
+    {
+        -- global transf
+        ["nerv.BiasLayer"] =
+        {
+            blayer1 = {{bias = "bias1"}, {dim_in = {440}, dim_out = {440}}},
+            blayer2 = {{bias = "bias2"}, {dim_in = {440}, dim_out = {440}}}
+        },
+        ["nerv.WindowLayer"] =
+        {
+            wlayer1 = {{window = "window1"}, {dim_in = {440}, dim_out = {440}}},
+            wlayer2 = {{window = "window2"}, {dim_in = {440}, dim_out = {440}}}
+        },
+        -- biased linearity
+        ["nerv.AffineLayer"] =
+        {
+            affine0 = {{ltp = "affine0_ltp", bp = "affine0_bp"},
+            {dim_in = {440}, dim_out = {2048}}},
+            affine1 = {{ltp = "affine1_ltp", bp = "affine1_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine2 = {{ltp = "affine2_ltp", bp = "affine2_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine3 = {{ltp = "affine3_ltp", bp = "affine3_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine4 = {{ltp = "affine4_ltp", bp = "affine4_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine5 = {{ltp = "affine5_ltp", bp = "affine5_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine6 = {{ltp = "affine6_ltp", bp = "affine6_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine7 = {{ltp = "affine7_ltp", bp = "affine7_bp"},
+            {dim_in = {2048}, dim_out = {2011}}}
+        },
+        ["nerv.SigmoidLayer"] =
+        {
+            sigmoid0 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid1 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid2 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid3 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid4 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid5 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid6 = {{}, {dim_in = {2048}, dim_out = {2048}}}
+        },
+        ["nerv.MPELayer"] =
+        {
+            mpe_crit = {{}, {dim_in = {2011, -1}, dim_out = {1},
+                        cmd = {
+                            arg = "--class-frame-counts=/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced/ali_train_pdf.counts --acoustic-scale=0.1 --lm-scale=1.0 --learn-rate=0.00001 --do-smbr=true --verbose=1",
+                            mdl = "/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_ali/final.mdl",
+                            lat = "scp:/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_denlats/lat.scp",
+                            ali = "ark:gunzip -c /slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_ali/ali.*.gz |"
+                        }
+                    }
+                }
+        },
+        ["nerv.SoftmaxLayer"] = -- softmax for decode output
+        {
+            softmax = {{}, {dim_in = {2011}, dim_out = {2011}}}
+        }
+    }, param_repo, gconf)
+
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            global_transf = {{}, {
+                dim_in = {440}, dim_out = {440},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "blayer1[1]",
+                    ["blayer1[1]"] = "wlayer1[1]",
+                    ["wlayer1[1]"] = "blayer2[1]",
+                    ["blayer2[1]"] = "wlayer2[1]",
+                    ["wlayer2[1]"] = "<output>[1]"
+                }
+            }},
+            main = {{}, {
+                dim_in = {440}, dim_out = {2011},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "affine0[1]",
+                    ["affine0[1]"] = "sigmoid0[1]",
+                    ["sigmoid0[1]"] = "affine1[1]",
+                    ["affine1[1]"] = "sigmoid1[1]",
+                    ["sigmoid1[1]"] = "affine2[1]",
+                    ["affine2[1]"] = "sigmoid2[1]",
+                    ["sigmoid2[1]"] = "affine3[1]",
+                    ["affine3[1]"] = "sigmoid3[1]",
+                    ["sigmoid3[1]"] = "affine4[1]",
+                    ["affine4[1]"] = "sigmoid4[1]",
+                    ["sigmoid4[1]"] = "affine5[1]",
+                    ["affine5[1]"] = "sigmoid5[1]",
+                    ["sigmoid5[1]"] = "affine6[1]",
+                    ["affine6[1]"] = "sigmoid6[1]",
+                    ["sigmoid6[1]"] = "affine7[1]",
+                    ["affine7[1]"] = "<output>[1]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            mpe_output = {{}, {
+                dim_in = {440, -1}, dim_out = {1},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "mpe_crit[1]",
+                    ["<input>[2]"] = "mpe_crit[2]",
+                    ["mpe_crit[1]"] = "<output>[1]"
+                }
+            }},
+            softmax_output = {{}, {
+                dim_in = {440}, dim_out = {2011},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "softmax[1]",
+                    ["softmax[1]"] = "<output>[1]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    return layer_repo
+end
+
+function get_network(layer_repo)
+    return layer_repo:get_layer("mpe_output")
+end
+
+function get_decode_network(layer_repo)
+    return layer_repo:get_layer("softmax_output")
+end
+
+function get_global_transf(layer_repo)
+    return layer_repo:get_layer("global_transf")
+end
+
+function make_readers(feature_rspecifier, layer_repo)
+    return {
+                {reader = nerv.KaldiReader(gconf,
+                    {
+                        id = "main_scp",
+                        feature_rspecifier = feature_rspecifier,
+                        frm_ext = gconf.frm_ext,
+                        global_transf = layer_repo:get_layer("global_transf"),
+                        mlfs = {}
+                    })
+                }
+            }
+end
+
+function get_input_order()
+    return {{id = "main_scp", global_transf = true},
+            {id = "key"}}
+end
+
+function get_accuracy(layer_repo)
+    local mpe_crit = layer_repo:get_layer("mpe_crit")
+    return mpe_crit.total_correct / mpe_crit.total_frames * 100
+end
+
+function print_stat(layer_repo)
+    local mpe_crit = layer_repo:get_layer("mpe_crit")
+    nerv.info("*** training stat begin ***")
+    nerv.printf("correct:\t\t%d\n", mpe_crit.total_correct)
+    nerv.printf("frames:\t\t\t%d\n", mpe_crit.total_frames)
+    nerv.printf("accuracy:\t\t%.3f%%\n", get_accuracy(layer_repo))
+    nerv.info("*** training stat end ***")
+end
diff --git a/nerv/examples/seq_trainer.lua b/nerv/examples/seq_trainer.lua
new file mode 100644
index 0000000..b8ed3eb
--- /dev/null
+++ b/nerv/examples/seq_trainer.lua
@@ -0,0 +1,87 @@
+function build_trainer(ifname)
+    local param_repo = nerv.ParamRepo()
+    param_repo:import(ifname, nil, gconf)
+    local layer_repo = make_layer_repo(param_repo)
+    local network = get_network(layer_repo)
+    local global_transf = get_global_transf(layer_repo)
+    local input_order = get_input_order()
+    local iterative_trainer = function (prefix, scp_file, bp)
+        local readers = make_readers(scp_file, layer_repo)
+        -- initialize the network
+        network:init(1)
+        gconf.cnt = 0
+        for ri = 1, #readers, 1 do
+            while true do
+                local data = readers[ri].reader:get_data()
+                if data == nil then
+                    break
+                end
+                -- prine stat periodically
+                gconf.cnt = gconf.cnt + 1
+                if gconf.cnt == 1000 then
+                    print_stat(layer_repo)
+                    nerv.CuMatrix.print_profile()
+                    nerv.CuMatrix.clear_profile()
+                    gconf.cnt = 0
+                    -- break
+                end
+                local input = {}
+                --            if gconf.cnt == 1000 then break end
+                for i, e in ipairs(input_order) do
+                    local id = e.id
+                    if data[id] == nil then
+                        nerv.error("input data %s not found", id)
+                    end
+                    local transformed
+                    if e.global_transf then
+                        local batch = gconf.cumat_type(data[id]:nrow(), data[id]:ncol())
+                        batch:copy_fromh(data[id])
+                        transformed = nerv.speech_utils.global_transf(batch,
+                        global_transf,
+                        gconf.frm_ext or 0, 0,
+                        gconf)
+                    else
+                        transformed = data[id]
+                    end
+                    table.insert(input, transformed)
+                end
+                err_output = {input[1]:create()}
+                network:batch_resize(input[1]:nrow())
+                if network:propagate(input, {{}}) == true then
+                    network:back_propagate({{}}, err_output, input, {{}})
+                    gconf.batch_size = 1.0 - gconf.momentum  -- important!!!
+                    network:update({{}}, input, {{}})
+                end
+                -- collect garbage in-time to save GPU memory
+                collectgarbage("collect")
+            end
+        end
+        print_stat(layer_repo)
+        nerv.CuMatrix.print_profile()
+        nerv.CuMatrix.clear_profile()
+        if prefix ~= nil then
+            nerv.info("writing back...")
+            local fname = string.format("%s_tr%.3f.nerv",
+            prefix, get_accuracy(layer_repo))
+            network:get_params():export(fname, nil)
+        end
+        return get_accuracy(layer_repo)
+    end
+    return iterative_trainer
+end
+
+dofile(arg[1])
+
+local pf0 = gconf.initialized_param
+local trainer = build_trainer(pf0)
+
+local i = 1
+nerv.info("[NN] begin iteration %d with lrate = %.6f", i, gconf.lrate)
+local accu_tr = trainer(string.format("%s_%s_iter_%d_lr%f",
+string.gsub(
+(string.gsub(pf0[1], "(.*/)(.*)", "%2")),
+"(.*)%..*", "%1"),
+os.date("%Y%m%d%H%M%S"),
+i, gconf.lrate), gconf.tr_scp, true)
+nerv.info("[TR] training set %d: %.3f", i, accu_tr)
+
diff --git a/nerv/examples/swb_baseline.lua b/nerv/examples/swb_baseline.lua
index 7783f2a..8f72200 100644
--- a/nerv/examples/swb_baseline.lua
+++ b/nerv/examples/swb_baseline.lua
@@ -2,7 +2,9 @@ require 'htk_io'
 gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9,
         cumat_type = nerv.CuMatrixFloat,
         mmat_type = nerv.MMatrixFloat,
+        direct_update = true,
         frm_ext = 5,
+        frm_trim = 5,
         tr_scp = "/slfs1/users/mfy43/swb_ivec/train_bp.scp",
         cv_scp = "/slfs1/users/mfy43/swb_ivec/train_cv.scp",
         htk_conf = "/slfs1/users/mfy43/swb_ivec/plp_0_d_a.conf",
@@ -10,8 +12,8 @@ gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9,
                 "/slfs1/users/mfy43/swb_global_transf.nerv"},
         debug = false}
 
-function make_sublayer_repo(param_repo)
-    return nerv.LayerRepo(
+function make_layer_repo(param_repo)
+    local layer_repo = nerv.LayerRepo(
     {
         -- global transf
         ["nerv.BiasLayer"] =
@@ -54,21 +56,23 @@ function make_sublayer_repo(param_repo)
             sigmoid5 = {{}, {dim_in = {2048}, dim_out = {2048}}},
             sigmoid6 = {{}, {dim_in = {2048}, dim_out = {2048}}}
         },
-        ["nerv.SoftmaxCELayer"] =
+        ["nerv.SoftmaxCELayer"] = -- softmax + ce criterion layer for finetune output
         {
             ce_crit = {{}, {dim_in = {3001, 1}, dim_out = {1}, compressed = true}}
+        },
+        ["nerv.SoftmaxLayer"] = -- softmax for decode output
+        {
+            softmax = {{}, {dim_in = {3001}, dim_out = {3001}}}
         }
     }, param_repo, gconf)
-end
 
-function make_layer_repo(sublayer_repo, param_repo)
-    return nerv.LayerRepo(
+    layer_repo:add_layers(
     {
         ["nerv.DAGLayer"] =
         {
             global_transf = {{}, {
                 dim_in = {429}, dim_out = {429},
-                sub_layers = sublayer_repo,
+                sub_layers = layer_repo,
                 connections = {
                     ["<input>[1]"] = "blayer1[1]",
                     ["blayer1[1]"] = "wlayer1[1]",
@@ -78,8 +82,8 @@ function make_layer_repo(sublayer_repo, param_repo)
                 }
             }},
             main = {{}, {
-                dim_in = {429, 1}, dim_out = {1},
-                sub_layers = sublayer_repo,
+                dim_in = {429}, dim_out = {3001},
+                sub_layers = layer_repo,
                 connections = {
                     ["<input>[1]"] = "affine0[1]",
                     ["affine0[1]"] = "sigmoid0[1]",
@@ -96,17 +100,51 @@ function make_layer_repo(sublayer_repo, param_repo)
                     ["sigmoid5[1]"] = "affine6[1]",
                     ["affine6[1]"] = "sigmoid6[1]",
                     ["sigmoid6[1]"] = "affine7[1]",
-                    ["affine7[1]"] = "ce_crit[1]",
+                    ["affine7[1]"] = "<output>[1]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            ce_output = {{}, {
+                dim_in = {429, 1}, dim_out = {1},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "ce_crit[1]",
                     ["<input>[2]"] = "ce_crit[2]",