Just come back, let's merge the new master

Merge branch 'master' into txh18/rnnlm

44 files changed, 1174 insertions, 722 deletions

diff --git a/.gitmodules b/.gitmodules
index 1432de9..9f556c5 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,12 +1,6 @@
 [submodule "luajit-2.0"]
 	path = luajit-2.0
 	url = http://luajit.org/git/luajit-2.0.git
-[submodule "nerv-speech"]
-	path = nerv-speech
-	url = https://github.com/Determinant/nerv-speech.git
-[submodule "speech"]
-	path = speech
-	url = https://github.com/Determinant/nerv-speech.git
 [submodule "luarocks"]
 	path = luarocks
 	url = https://github.com/keplerproject/luarocks.git
diff --git a/Makefile b/Makefile
index fa888c3..664a83b 100644
--- a/Makefile
+++ b/Makefile
@@ -9,6 +9,7 @@ luarocks:
 install:
 	cd nerv; $(PREFIX)/bin/luarocks make
 speech:
+	cd speech/speech_utils; $(PREFIX)/bin/luarocks make
 	cd speech/htk_io; $(PREFIX)/bin/luarocks make
 clean:
 	cd nerv && make clean
diff --git a/README.md b/README.md
index 10d531c..8c21bd9 100644
--- a/README.md
+++ b/README.md
@@ -8,10 +8,12 @@ First make sure you have __lua__ and __CUDA__ installed on your computer.
 __Nerv__ is currently developed via github.You can download and make __Nerv__ by doing the following:
 ```
 cd ~
-git clone https://github.com/Determinant/nerv.git
+git clone https://github.com/Nerv-SJTU/nerv.git
 cd nerv
+git clone https://github.com/Nerv-SJTU/nerv-speech.git speech
 git submodule init && git submodule update
 make
+make speech
 ```
 The `git submodule` command is for the __luajit__ repository inside __Nerv__.  
 Now, you can try to run some example scripts.  
diff --git a/embedding_example/.gitignore b/embedding_example/.gitignore
new file mode 100644
index 0000000..8e68213
--- /dev/null
+++ b/embedding_example/.gitignore
@@ -0,0 +1,2 @@
+main
+main.o
diff --git a/embedding_example/Makefile b/embedding_example/Makefile
new file mode 100644
index 0000000..73287f4
--- /dev/null
+++ b/embedding_example/Makefile
@@ -0,0 +1,20 @@
+CFLAG += -I ../install/include/luajit-2.0/  -I ../install/include/nerv/
+LDFLAG += -L../install/lib/ -lluajit-5.1 -Wl,-rpath=../install/lib/ -lluaT -lnervcore
+GCC := gcc
+
+.PHONY: FORCE
+
+all: main FORCE
+clean:
+	-rm -f *.o
+	-rm main
+
+FORCE: ../install/bin/luarocks
+	echo "#!/bin/bash" > run.sh
+	$< path >> run.sh
+	echo "./main" >> run.sh
+main: main.o
+	$(GCC) -o $@ $<  $(LDFLAG)
+
+main.o: main.c
+	$(GCC) $(CFLAG) -o $@ $< -c
diff --git a/embedding_example/main.c b/embedding_example/main.c
new file mode 100644
index 0000000..8856d58
--- /dev/null
+++ b/embedding_example/main.c
@@ -0,0 +1,93 @@
+#include "lua.h"
+#include "lauxlib.h"
+#include "lualib.h"
+#include "matrix/matrix.h"
+#include "common.h"
+#include "luaT/luaT.h"
+#include <stdio.h>
+
+const char *nerv_matrix_host_float_tname = "nerv.MMatrixFloat";
+const char *input_name = "_nerv_embed_input";
+const char *output_name = "_nerv_embed_output";
+extern Matrix *nerv_matrix_host_float_create(long nrow, long ncol, Status *status);
+extern void nerv_matrix_host_float_data_retain(Matrix *self);
+extern void nerv_matrix_host_float_data_free(Matrix *self, Status *status);
+
+lua_State *L;
+Matrix *input, *output;
+Status status;
+
+void setup_nerv() {
+    L = lua_open();
+    luaL_openlibs(L);
+    luaL_loadfile(L, "setup_nerv.lua");
+    /* network configuration */
+    lua_pushstring(L, "../nerv/examples/swb_baseline.lua");
+    if (lua_pcall(L, 1, LUA_MULTRET, 0))
+    {
+        printf("%s\n", luaL_checkstring(L, 1));
+        exit(1);
+    }
+    /* lua stack now: input width, output width, propagator */
+    input = nerv_matrix_host_float_create(1, luaL_checkinteger(L, 1), &status);
+    NERV_LUA_CHECK_STATUS(L, status);
+    output = nerv_matrix_host_float_create(1, luaL_checkinteger(L, 2), &status);
+    NERV_LUA_CHECK_STATUS(L, status);
+    /* add reference to avoid gc */
+    luaT_pushudata(L, output, nerv_matrix_host_float_tname);
+    luaT_pushudata(L, input, nerv_matrix_host_float_tname);
+    lua_setfield(L, LUA_GLOBALSINDEX, input_name);
+    lua_setfield(L, LUA_GLOBALSINDEX, output_name);
+}
+
+
+void propagate(float for_fun) {
+    int i, j;
+    printf("ok: %d\n", lua_gettop(L));
+    lua_pushvalue(L, 3);
+    /* lua stack now: input width, output width, propagator, propagator */
+    for (i = 0; i < input->nrow; i++) /* nrow is actually 1 */
+    {
+        float *nerv_row = (float *)((char *)input->data.f + i * input->stride);
+        for (j = 0; j < input->ncol; j++)
+        {
+            nerv_row[j] = j * for_fun;
+        }
+    }
+    lua_getfield(L, LUA_GLOBALSINDEX, input_name);
+    lua_getfield(L, LUA_GLOBALSINDEX, output_name);
+    /* lua stack now: input width, output width, propagator, propagator, input, output */
+    if (lua_pcall(L, 2, 0, 0)) /* call propagator with two parameters */
+    {
+        printf("%s\n", luaL_checkstring(L, -1));
+        exit(-1);
+    }
+    /* lua stack now: input width, output width, propagator */
+    printf("## output: %ld %ld ##\n", output->nrow, output->ncol);
+    for (i = 0; i < output->nrow; i++) /* nrow is actually 1 */
+    {
+        float *nerv_row = (float *)((char *)output->data.f + i * output->stride);
+        for (j = 0; j < output->ncol; j++)
+        {
+            printf("%.8f ", nerv_row[j]);
+        }
+    }
+}
+
+void teardown_nerv() {
+    lua_pushnil(L);
+    lua_pushnil(L);
+    lua_setfield(L, LUA_GLOBALSINDEX, input_name);
+    lua_setfield(L, LUA_GLOBALSINDEX, output_name);
+    lua_gc(L, LUA_GCCOLLECT, 0);
+}
+
+int main() {
+    setup_nerv();
+    propagate(1.0);
+    propagate(2.0);
+    propagate(2.0);
+    propagate(3.0);
+    teardown_nerv();
+    return 0;
+}
diff --git a/embedding_example/run.sh b/embedding_example/run.sh
new file mode 100755
index 0000000..e919263
--- /dev/null
+++ b/embedding_example/run.sh
@@ -0,0 +1,4 @@
+#!/bin/bash
+export LUA_PATH='/home/slhome/mfy43/.luarocks/share/lua/5.1/?.lua;/home/slhome/mfy43/.luarocks/share/lua/5.1/?/init.lua;/home/slhome/mfy43/nerv/install/share/lua/5.1/?.lua;/home/slhome/mfy43/nerv/install/share/lua/5.1/?/init.lua;./?.lua;/usr/local/share/luajit-2.0.4/?.lua;/usr/local/share/lua/5.1/?.lua;/usr/local/share/lua/5.1/?/init.lua'
+export LUA_CPATH='/home/slhome/mfy43/.luarocks/lib/lua/5.1/?.so;/home/slhome/mfy43/nerv/install/lib/lua/5.1/?.so;./?.so;/usr/local/lib/lua/5.1/?.so;/usr/local/lib/lua/5.1/loadall.so'
+./main
diff --git a/embedding_example/setup_nerv.lua b/embedding_example/setup_nerv.lua
new file mode 100644
index 0000000..d80c306
--- /dev/null
+++ b/embedding_example/setup_nerv.lua
@@ -0,0 +1,25 @@
+local k,l,_=pcall(require,"luarocks.loader") _=k and l.add_context("nerv","scm-1")
+require 'nerv'
+local arg = {...}
+dofile(arg[1])
+local param_repo = nerv.ParamRepo()
+param_repo:import(gconf.initialized_param, nil, gconf)
+local layer_repo = make_layer_repo(param_repo)
+local network = get_decode_network(layer_repo)
+local global_transf = get_global_transf(layer_repo)
+local batch_size = 1
+network:init(batch_size)
+
+function propagator(input, output)
+    local transformed = nerv.speech_utils.global_transf(
+                            gconf.cumat_type.new_from_host(input),
+                            global_transf, 0, 0, gconf) -- preprocessing
+    local gpu_input = transformed
+    local gpu_output = nerv.CuMatrixFloat(output:nrow(), output:ncol())
+    network:propagate({gpu_input}, {gpu_output})
+    gpu_output:copy_toh(output)
+    -- collect garbage in-time to save GPU memory
+    collectgarbage("collect")
+end
+
+return network.dim_in[1], network.dim_out[1], propagator
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 |"
+