Merge branch 'fastnn' of github.com:Nerv-SJTU/nerv-speech into fastnnfastnn

7 files changed, 323 insertions, 27 deletions

diff --git a/htk_io/init.lua b/htk_io/init.lua
index b360b67..677d3e9 100644
--- a/htk_io/init.lua
+++ b/htk_io/init.lua
@@ -1,8 +1,9 @@
 require 'libhtkio'
 require 'speech_utils'
+require 'threads'
 local TNetReader = nerv.class("nerv.TNetReader", "nerv.DataReader")
 
-function TNetReader:__init(global_conf, reader_conf)
+function TNetReader:__init(global_conf, reader_conf, feat_repo_shareid, data_mutex_shareid)
     self.feat_id = reader_conf.id
     self.frm_ext = reader_conf.frm_ext
     self.gconf = global_conf
@@ -10,9 +11,22 @@ function TNetReader:__init(global_conf, reader_conf)
     if self.debug == nil then
         self.debug = false
     end
-    self.feat_repo = nerv.TNetFeatureRepo(reader_conf.scp_file,
+
+    if feat_repo_shareid ~= nil then
+	self.feat_repo = nerv.TNetFeatureRepo(feat_repo_shareid)
+    else
+    	self.feat_repo = nerv.TNetFeatureRepo(reader_conf.scp_file,
                                           reader_conf.conf_file,
                                           reader_conf.frm_ext)
+    end
+
+    --print(self.feat_repo)
+    self.data_mutex = nil
+    if data_mutex_shareid ~= nil then
+	self.data_mutex = threads.Mutex(data_mutex_shareid)
+    end
+	
+
     self.lab_repo = {}
     if reader_conf.mlfs then
         for id, mlf_spec in pairs(reader_conf.mlfs) do
@@ -26,7 +40,14 @@ function TNetReader:__init(global_conf, reader_conf)
 end
 
 function TNetReader:get_data()
+    if self.data_mutex ~= nil then
+	self.data_mutex:lock()
+    end
+
     if self.feat_repo:is_end() then
+	if self.data_mutex ~= nil then
+		self.data_mutex:unlock()
+	end
         return nil
     end
     local res = {}
@@ -66,6 +87,11 @@ function TNetReader:get_data()
     end
     -- move the pointer to next
     self.feat_repo:next()
+
+    if self.data_mutex ~= nil then
+        self.data_mutex:unlock()
+    end 
+
     collectgarbage("collect")
     return res
 end
diff --git a/htk_io/src/cwrapper.cpp b/htk_io/src/cwrapper.cpp
index b7ce2d5..66cde23 100644
--- a/htk_io/src/cwrapper.cpp
+++ b/htk_io/src/cwrapper.cpp
@@ -8,6 +8,7 @@
 extern "C" {
 #include "cwrapper.h"
 #include "string.h"
+#include "pthread.h"
 #include "nerv/common.h"
 
     extern Matrix *nerv_matrix_host_float_create(long nrow, long ncol, Status *status);
@@ -29,6 +30,7 @@ extern "C" {
         const char* cvn_mask;
         const char* cvg_file;
         TNet::Matrix<float> feats_host; /* KaldiLib implementation */
+	int  refcount;
     };
 
     TNetFeatureRepo *tnet_feature_repo_new(const char *p_script, const char *config, int context) {
@@ -53,6 +55,18 @@ extern "C" {
         return repo;
     }
 
+    TNetFeatureRepo *tnet_feature_repo_newWithId(long id)
+    {
+	TNetFeatureRepo *repo = (TNetFeatureRepo*)id;
+	__sync_fetch_and_add(&repo->refcount, 1);
+	return repo;
+    }
+
+    long tnet_feature_repo_id(TNetFeatureRepo *repo)
+    {
+	return (long)(repo);
+    }
+
     Matrix *tnet_feature_repo_read_utterance(TNetFeatureRepo *repo, lua_State *L, int debug) {
         Matrix *mat;                    /* nerv implementation */
         repo->feature_repo.ReadFullMatrix(repo->feats_host);
@@ -93,12 +107,19 @@ extern "C" {
     }
 
     void tnet_feature_repo_destroy(TNetFeatureRepo *repo) {
-        if (repo->cmn_mask)
-            free(repo->cmn_path);
-        if (repo->cvn_mask)
-            free(repo->cvn_path);
-        free(repo->p_deriv_win_lenghts);
-        delete repo;
+	if (NULL != repo)
+	{
+	 	if(__sync_fetch_and_add(&repo->refcount, -1) == 1)
+		{
+        		if (repo->cmn_mask)
+            			free(repo->cmn_path);
+        		if (repo->cvn_mask)
+            			free(repo->cvn_path);
+        		free(repo->p_deriv_win_lenghts);
+        		delete repo;
+			repo = NULL;
+		}
+	}
     }
 
     struct TNetLabelRepo {
@@ -114,6 +135,16 @@ extern "C" {
         return repo;
     }
 
+    TNetLabelRepo *tnet_label_repo_newWithId(long id)
+    {
+	return (TNetLabelRepo*)id;
+    }
+
+    long tnet_label_repo_id(TNetLabelRepo *repo)
+    {
+	return (long)(repo);
+    }
+
     Matrix *tnet_label_repo_read_utterance(TNetLabelRepo *repo,
                                             size_t frames,
                                             size_t sample_rate,
@@ -143,6 +174,10 @@ extern "C" {
     }
 
     void tnet_label_repo_destroy(TNetLabelRepo *repo) {
-        delete repo;
+	if (NULL != repo)	
+	{
+        	delete repo;
+		repo = NULL;
+	}
     }
 }
diff --git a/htk_io/src/cwrapper.h b/htk_io/src/cwrapper.h
index e1bce6e..44e77cf 100644
--- a/htk_io/src/cwrapper.h
+++ b/htk_io/src/cwrapper.h
@@ -16,6 +16,10 @@ extern "C" {
     void tnet_feature_repo_next(TNetFeatureRepo *repo);
     int tnet_feature_repo_is_end(TNetFeatureRepo *repo);
     void tnet_feature_repo_destroy(TNetFeatureRepo *repo);
+    TNetFeatureRepo *tnet_feature_repo_newWithId(long id); 
+    long tnet_feature_repo_id(TNetFeatureRepo *repo);
+
+
 
     typedef struct TNetLabelRepo TNetLabelRepo;
 
@@ -31,6 +35,11 @@ extern "C" {
                                             int debug);
 
     void tnet_label_repo_destroy(TNetLabelRepo *repo);
+
+    TNetLabelRepo *tnet_label_repo_newWithId(long id); 
+
+    long tnet_label_repo_id(TNetLabelRepo *repo);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/htk_io/src/init.c b/htk_io/src/init.c
index 8a1ec3b..04046e9 100644
--- a/htk_io/src/init.c
+++ b/htk_io/src/init.c
@@ -7,13 +7,37 @@ const char *nerv_tnet_label_repo_tname = "nerv.TNetLabelRepo";
 const char *nerv_matrix_host_float_tname = "nerv.MMatrixFloat";
 
 static int feat_repo_new(lua_State *L) {
-    const char *scp_file = luaL_checkstring(L, 1);
-    const char *conf = luaL_checkstring(L, 2);
-    int frm_ext = luaL_checkinteger(L, 3);
-    TNetFeatureRepo *repo = tnet_feature_repo_new(scp_file, conf, frm_ext);
+    TNetFeatureRepo *repo = NULL;
+    if(lua_gettop(L) == 1)
+    {
+	long id = luaL_checkinteger(L, 1);
+    	repo = tnet_feature_repo_newWithId(id);
+    }
+    else
+    {
+    	const char *scp_file = luaL_checkstring(L, 1);
+    	const char *conf = luaL_checkstring(L, 2);
+    	int frm_ext = luaL_checkinteger(L, 3);
+    	repo = tnet_feature_repo_new(scp_file, conf, frm_ext);
+    }
+
     luaT_pushudata(L, repo, nerv_tnet_feat_repo_tname);
     return 1;
 }
+static int feat_repo_id(lua_State *L) {
+	TNetFeatureRepo *repo = luaT_checkudata(L, 1, nerv_tnet_feat_repo_tname);
+	lua_pushinteger(L, tnet_feature_repo_id(repo));
+	return 1;
+}
+
+static int feat_repo_tostring(lua_State *L) 
+{
+        char str[128];
+       	TNetFeatureRepo *repo = luaT_checkudata(L, 1, nerv_tnet_feat_repo_tname);  
+        snprintf(str, 128, "%s <%lx>", nerv_tnet_feat_repo_tname, tnet_feature_repo_id(repo));
+        lua_pushstring(L, str);
+        return 1;
+}
 
 static int feat_repo_destroy(lua_State *L) {
     TNetFeatureRepo *repo = luaT_checkudata(L, 1, nerv_tnet_feat_repo_tname);
@@ -55,22 +79,40 @@ static const luaL_Reg feat_repo_methods[] = {
     {"cur_tag", feat_repo_current_tag},
     {"next", feat_repo_next},
     {"is_end", feat_repo_is_end},
+    {"__tostring", feat_repo_tostring},
+    {"id", feat_repo_id},
     {NULL, NULL}
 };
 
 static int label_repo_new(lua_State *L) {
-    const char *mlf_file = luaL_checkstring(L, 1);
-    const char *fmt = luaL_checkstring(L, 2);
-    const char *arg = luaL_checkstring(L, 3);
-    const char *dir = luaL_checkstring(L, 4);
-    const char *ext = luaL_checkstring(L, 5);
-    TNetLabelRepo *repo = tnet_label_repo_new(
+    TNetLabelRepo *repo = NULL;
+    if(lua_gettop(L) == 1)
+    {   
+        long id = luaL_checkinteger(L, 1); 
+        repo = tnet_label_repo_newWithId(id);
+    }   
+    else
+    {   
+    	const char *mlf_file = luaL_checkstring(L, 1);
+    	const char *fmt = luaL_checkstring(L, 2);
+    	const char *arg = luaL_checkstring(L, 3);
+    	const char *dir = luaL_checkstring(L, 4);
+    	const char *ext = luaL_checkstring(L, 5);
+    	repo = tnet_label_repo_new(
                                 mlf_file, fmt, arg,
                                 dir, ext);
+    }
     luaT_pushudata(L, repo, nerv_tnet_label_repo_tname);
     return 1;
 }
 
+static int label_repo_id(lua_State *L) {
+	TNetLabelRepo *repo = luaT_checkudata(L, 1, nerv_tnet_label_repo_tname);
+	lua_pushinteger(L, tnet_label_repo_id(repo));
+        return 1;
+}
+
+
 static int label_repo_read_utterance(lua_State *L) {
     TNetLabelRepo *repo = luaT_checkudata(L, 1, nerv_tnet_label_repo_tname);
     TNetFeatureRepo *feat_repo = luaT_checkudata(L, 2, nerv_tnet_feat_repo_tname);
@@ -95,6 +137,7 @@ static int label_repo_destroy(lua_State *L) {
 
 static const luaL_Reg label_repo_methods[] = {
     {"get_utter", label_repo_read_utterance},
+    {"id", label_repo_id},
     {NULL, NULL}
 };
 
diff --git a/htk_io/tools/nerv_to_tnet.cpp b/htk_io/tools/nerv_to_tnet.cpp
new file mode 100644
index 0000000..91be7db
--- /dev/null
+++ b/htk_io/tools/nerv_to_tnet.cpp
@@ -0,0 +1,89 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string>
+#include <cstring>
+#include <iostream>
+#include <fstream>
+
+char token[1024];
+typedef struct node
+{
+	double **mat;
+	int nrow;
+	int ncol;
+	int index;
+}node;
+
+int main(int argc, char **argv)
+{
+	if (argc != 3)
+    	{   
+        	printf("%s nerv.model.in tnet.model.out\n", argv[0]);
+    	}   
+
+    	freopen(argv[1], "r", stdin);
+    	freopen(argv[2], "w", stdout);
+	node* trans[100], *bp[100];
+	int layers = 0;
+	int i, j, k;
+	char type[128];
+	int idx = 0;
+	char *flag = "affine";
+	while (NULL != gets(token))
+    	{   
+        	int nrow, ncol;
+		if (token[0] == '{')
+		{
+			char *ss = strstr(token, flag);	
+    			if (sscanf(ss, "affine%d_%[a-z]", &idx, type) == 2)
+			{
+				node *no = new node;
+				scanf("%d %d", &nrow, &ncol);	
+				no->nrow = nrow;
+				no->ncol = ncol;
+				no->index = idx;
+				no->mat = (double **)malloc(nrow * sizeof(double *));
+				for (i = 0; i < nrow; i++)
+                		no->mat[i] = (double *)malloc(ncol * sizeof(double));
+                		for (i = 0; i < nrow; i++)
+            				for (j = 0; j < ncol; j++)
+                    			scanf("%lf", no->mat[i] + j);
+				
+				if (nrow == 1)
+					bp[idx] = no;
+				else
+					trans[idx] = no;		
+				layers++;
+			}
+		}
+	}
+	if (layers%2)
+	{
+		perror("number layers of LinearTrans and Bias is not the same.\n");
+		return -1;
+	}
+
+	layers /= 2;
+	for (i = 0; i < layers; i++)
+	{
+		printf("<biasedlinearity> %d %d\n", trans[i]->ncol, trans[i]->nrow);
+		printf("m %d %d\n", trans[i]->ncol, trans[i]->nrow);
+		for (j = 0; j < trans[i]->ncol; j++)
+		{
+			for (k = 0; k < trans[i]->nrow; k++)
+				printf("%.10lf ", trans[i]->mat[k][j]);		
+			printf("\n");
+		}
+		printf("v %d", bp[i]->ncol);	
+		for (j = 0; j < bp[i]->ncol; j++)
+			printf(" %.10lf", bp[i]->mat[0][j]);
+		printf("\n");	
+		if (i != layers - 1 && trans[i+1]->ncol < trans[i+1]->nrow)
+		printf("<sigmoid> %d %d\n", bp[i]->ncol, bp[i]->ncol);
+	}
+	printf("<softmax> %d %d\n", bp[layers - 1]->ncol, bp[layers - 1]->ncol);
+	perror("done!\n");
+
+	return 0;
+}
+
diff --git a/htk_io/tools/tnet_to_nerv.cpp b/htk_io/tools/tnet_to_nerv.cpp
index a779a25..f96781a 100644
--- a/htk_io/tools/tnet_to_nerv.cpp
+++ b/htk_io/tools/tnet_to_nerv.cpp
@@ -2,14 +2,22 @@
 #include <fstream>
 #include <string>
 #include <cstring>
-#include <cstdlib>
+#include <stdlib.h>
 char token[1024];
 char output[1024];
 double **mat;
 int main(int argc, char **argv) {
+
+    if (argc != 3)
+    {
+	printf("%s tnet.model.in nerv.model.out\n", argv[0]);
+    }
+
     std::ofstream fout;
-    fout.open(argv[1]);
-    int cnt = 0;
+    freopen(argv[1], "r", stdin);
+    fout.open(argv[2]);
+    int cnt = 0, bias = 1, win = 1;
+    long length = 0, base = 0;
     while (scanf("%s", token) != EOF)
     {
         int nrow, ncol;
@@ -19,13 +27,13 @@ int main(int argc, char **argv) {
             scanf("%d %d", &ncol, &nrow);
             scanf("%s %d %d", token, &ncol, &nrow);
             printf("%d %d\n", nrow, ncol);
-            mat = (double **)malloc(nrow * sizeof(double *));
+	    mat = (double **)malloc(nrow * sizeof(double *));
             for (i = 0; i < nrow; i++)
                 mat[i] = (double *)malloc(ncol * sizeof(double));
             for (j = 0; j < ncol; j++)
                 for (i = 0; i < nrow; i++)
                     scanf("%lf", mat[i] + j);
-            long base = fout.tellp();
+            base = fout.tellp();
             sprintf(output, "%16d", 0);
             fout << output;
             sprintf(output, "{type=\"nerv.LinearTransParam\",id=\"affine%d_ltp\"}\n",
@@ -38,10 +46,8 @@ int main(int argc, char **argv) {
                 for (j = 0; j < ncol; j++)
                     fout << mat[i][j] << " ";
                 fout << std::endl;
-                free(mat[i]);
             }
-            free(mat);
-            long length = fout.tellp() - base;
+            length = fout.tellp() - base;
             fout.seekp(base);
             sprintf(output, "[%13lu]\n", length);
             fout << output;
@@ -69,6 +75,62 @@ int main(int argc, char **argv) {
                 cnt++;
             }
         }
+	else if (strcmp(token, "<bias>") == 0)
+	{
+		scanf("%d %d", &ncol, &nrow);
+		scanf("%s %d", token, &ncol);	
+		base = fout.tellp();
+		nrow = 1;	
+		mat = (double **)malloc(nrow * sizeof(double *));
+            	for (i = 0; i < nrow; i++)
+                	mat[i] = (double *)malloc(ncol * sizeof(double));
+            	for (j = 0; j < ncol; j++)
+                	for (i = 0; i < nrow; i++)
+                    		scanf("%lf", mat[i] + j);
+                sprintf(output, "%16d", 0);
+                fout << output;
+                sprintf(output, "{type=\"nerv.MatrixParam\",id=\"bias%d\"}\n",bias);
+                fout << output;
+                sprintf(output, "1 %d\n", ncol);
+                fout << output;
+                for (j = 0; j < ncol; j++)
+                    fout << mat[0][j] << " ";
+                fout << std::endl;
+                length = fout.tellp() - base;
+                fout.seekp(base);
+                sprintf(output, "[%13lu]\n", length);
+                fout << output;
+                fout.seekp(0, std::ios_base::end);
+                bias++;
+	}
+	else if (strcmp(token, "<window>") == 0)
+        {
+                scanf("%d %d", &ncol, &nrow);
+                scanf("%s %d", token, &ncol);
+                base = fout.tellp();
+		nrow = 1;
+                mat = (double **)malloc(nrow * sizeof(double *));
+                for (i = 0; i < nrow; i++)
+                        mat[i] = (double *)malloc(ncol * sizeof(double));
+                for (j = 0; j < ncol; j++)
+                        for (i = 0; i < nrow; i++)
+                                scanf("%lf", mat[i] + j);
+                sprintf(output, "%16d", 0);
+                fout << output;
+                sprintf(output, "{type=\"nerv.MatrixParam\",id=\"window%d\"}\n",win);
+                fout << output;
+                sprintf(output, "1 %d\n", ncol);
+                fout << output;
+                for (j = 0; j < ncol; j++)
+                    fout << mat[0][j] << " ";
+                fout << std::endl;
+                length = fout.tellp() - base;
+                fout.seekp(base);
+                sprintf(output, "[%13lu]\n", length);
+                fout << output;
+                fout.seekp(0, std::ios_base::end);
+                win++;
+        }
     }
     return 0;
 }
diff --git a/speech_utils/init.lua b/speech_utils/init.lua
index f89f4fd..1bfec16 100644
--- a/speech_utils/init.lua
+++ b/speech_utils/init.lua
@@ -14,3 +14,35 @@ function nerv.speech_utils.global_transf(feat_utter, global_transf,
     collectgarbage("collect")
     return expanded
 end
+
+function nerv.speech_utils.feat_expand(feat_utter, frm_ext, gconf)
+	local rearranged
+	if frm_ext > 0 then
+        	local step = frm_ext * 2 + 1 
+        	-- expand the feature
+        	 local expanded = gconf.mmat_type(feat_utter:nrow(), feat_utter:ncol() * step)	
+		 expanded:expand_frm(feat_utter, frm_ext)
+
+		-- rearrange the feature (``transpose'' operation in TNet)
+		rearranged = gconf.mmat_type(feat_utter:nrow() - frm_ext*2, feat_utter:ncol() * step)
+        	rearranged:rearrange_frm(expanded, step, frm_ext, feat_utter:nrow() - frm_ext)
+	else
+		rearranged = feat_utter
+	end
+
+	collectgarbage("collect")
+    	return rearranged 
+end
+
+
+function nerv.speech_utils.normalize(mat, global_transf)
+    	-- prepare for transf
+    	local input = {mat}
+    	local output = {mat:create()}
+    	-- do transf
+    	global_transf:init(input[1]:nrow())
+    	global_transf:propagate(input, output)
+	
+	return output[1]
+end
+