add script to convert nerv param file to TNet format

2 files changed, 61 insertions, 1 deletions

diff --git a/htk_io/tools/nerv_to_tnet.lua b/htk_io/tools/nerv_to_tnet.lua
new file mode 100644
index 0000000..c0ac76b
--- /dev/null
+++ b/htk_io/tools/nerv_to_tnet.lua
@@ -0,0 +1,60 @@
+-- usage: nerv config_file nerv_param_input tnet_output
+
+dofile(arg[1])
+param_repo = nerv.ParamRepo()
+param_repo:import({arg[2], gconf.initialized_param[2]}, nil, gconf)
+layer_repo = make_layer_repo(param_repo)
+f = assert(io.open(arg[3], "w"))
+
+function print_tnet_matrix(cumat)
+    local strs = {}
+    collectgarbage()
+    if cumat:nrow() == 1 then
+        local mat = nerv.MMatrixFloat(1, cumat:ncol())
+        cumat:copy_toh(mat)
+        table.insert(strs, string.format("v %d\n", mat:ncol()))
+        for j = 0, mat:ncol() - 1 do
+            table.insert(strs, string.format("%.8f ", mat[0][j]))
+        end
+        table.insert(strs, "\n")
+        f:write(table.concat(strs))
+    else
+        cumat = cumat:trans()
+        local mat = nerv.MMatrixFloat(cumat:nrow(), cumat:ncol())
+        cumat:copy_toh(mat)
+        table.insert(strs, string.format("m %d %d\n", mat:nrow(), mat:ncol()))
+        for i = 0, mat:nrow() - 1 do
+            local row = mat[i]
+            for j = 0, mat:ncol() - 1 do
+                table.insert(strs, string.format("%.8f ", row[j]))
+            end
+            table.insert(strs, "\n")
+            f:write(table.concat(strs))
+            strs = {}
+        end
+    end
+end
+local lnames = {"affine0", "sigmoid0",
+                "affine1", "sigmoid1",
+                "affine2", "sigmoid2",
+                "affine3", "sigmoid3",
+                "affine4", "sigmoid4",
+                "affine5", "sigmoid5",
+                "affine6", "sigmoid6",
+                "affine7", "ce_crit"}
+for i, name in ipairs(lnames) do
+    local layer = layer_repo:get_layer(name)
+    local layer_type = layer.__typename
+    if layer_type == "nerv.AffineLayer" then
+        f:write(string.format("<biasedlinearity> %d %d\n", layer.dim_out[1], layer.dim_in[1]))
+        print_tnet_matrix(layer.ltp.trans)
+        print_tnet_matrix(layer.bp.trans)
+    elseif layer_type == "nerv.SigmoidLayer" then
+        f:write(string.format("<sigmoid> %d %d\n", layer.dim_out[1], layer.dim_in[1]))
+    elseif layer_type == "nerv.SoftmaxCELayer" then
+        f:write(string.format("<softmax> %d %d\n", layer.dim_in[1], layer.dim_in[1]))
+    else
+        nerv.error("unknown layer type %s", layer_type)
+    end
+end
+f:close()
diff --git a/htk_io/tools/tnet_to_nerv.c b/htk_io/tools/tnet_to_nerv.c
index f781236..5774819 100644
--- a/htk_io/tools/tnet_to_nerv.c
+++ b/htk_io/tools/tnet_to_nerv.c
@@ -41,7 +41,7 @@ int main() {
                 fprintf(fout, "%16d", 0);
                 fprintf(fout, "{type=\"nerv.BiasParam\",id=\"affine%d_bp\"}\n",
                         cnt);
-                fprintf(fout, "1 %d\n", nrow, ncol);
+                fprintf(fout, "1 %d\n", ncol);
                 for (j = 0; j < ncol; j++)
                     fprintf(fout, "%.8f ", mat[0][j]);
                 fprintf(fout, "\n");