change concept of ParamRepo; provide generalized param update; code clean-up; #25 #26 #27 #29

19 files changed, 286 insertions, 175 deletions

diff --git a/Makefile b/Makefile
index 448e003..8f1d491 100644
--- a/Makefile
+++ b/Makefile
@@ -12,8 +12,8 @@ LUA_LIBS := matrix/init.lua io/init.lua nerv.lua \
 			nn/init.lua nn/layer_repo.lua nn/param_repo.lua nn/layer_dag.lua \
 			io/sgd_buffer.lua
 INCLUDE := -I build/luajit-2.0/include/luajit-2.0/ -DLUA_USE_APICHECK
-#CUDA_BASE := /usr/local/cuda-6.5
-CUDA_BASE := /usr/local/cuda-5.0
+CUDA_BASE := /usr/local/cuda-6.5
+#CUDA_BASE := /usr/local/cuda-5.0
 CUDA_INCLUDE := -I $(CUDA_BASE)/include/
 INCLUDE += $(CUDA_INCLUDE)
 LDFLAGS := -L$(CUDA_BASE)/lib64/  -Wl,-rpath=$(CUDA_BASE)/lib64/ -lcudart -lcublas
diff --git a/examples/asr_trainer.lua b/examples/asr_trainer.lua
index 05d770f..a5727be 100644
--- a/examples/asr_trainer.lua
+++ b/examples/asr_trainer.lua
@@ -1,50 +1,58 @@
 function build_trainer(ifname)
-    local param_repo = make_param_repo(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 crit = get_criterion_layer(sublayer_repo)
     local network = get_network(layer_repo)
+    local input_order = get_input_order()
     local iterative_trainer = function (prefix, scp_file, bp)
         gconf.randomize = bp
         -- build buffer
-        local buffer = make_buffer(make_reader(scp_file, layer_repo))
+        local buffer = make_buffer(make_readers(scp_file, layer_repo))
         -- initialize the network
         network:init(gconf.batch_size)
         gconf.cnt = 0
+        err_input = {nerv.CuMatrixFloat(256, 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(crit)
+                print_stat(sublayer_repo)
+                nerv.CuMatrix.print_profile()
+                nerv.CuMatrix.clear_profile()
                 gconf.cnt = 0
+                -- break
             end
+            local input = {}
 --            if gconf.cnt == 100 then break end
-
-            input = {data.main_scp, data.phone_state}
-            output = {}
-            err_input = {}
+            for i, id in ipairs(input_order) do
+                if data[id] == nil then
+                    nerv.error("input data %s not found", id)
+                end
+                table.insert(input, data[id])
+            end
+            local output = {nerv.CuMatrixFloat(256, 1)}
             err_output = {input[1]:create()}
             network:propagate(input, output)
             if bp then
-                network:back_propagate(err_output, err_input, input, output)
+                network:back_propagate(err_input, err_output, input, output)
                 network:update(err_input, input, output)
             end
             -- collect garbage in-time to save GPU memory
             collectgarbage("collect")
         end
-        print_stat(crit)
+        print_stat(sublayer_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(crit))
-            cf = nerv.ChunkFile(fname, "w")
-            for i, p in ipairs(network:get_params()) do
-                cf:write_chunk(p)
-            end
-            cf:close()
+                            prefix, get_accuracy(sublayer_repo))
+            network:get_params():export(fname, nil)
         end
-        return get_accuracy(crit)
+        return get_accuracy(sublayer_repo)
     end
     return iterative_trainer
 end
@@ -73,7 +81,7 @@ for i = 1, max_iter do
     local accu_new = trainer(
                         string.format("%s_%s_iter_%d_lr%f_tr%.3f",
                             string.gsub(
-                                (string.gsub(pf0, "(.*/)(.*)", "%2")),
+                                (string.gsub(pf0[1], "(.*/)(.*)", "%2")),
                                 "(.*)%..*", "%1"),
                             os.date("%Y%m%d%H%M%S"),
                             i, gconf.lrate,
diff --git a/examples/swb_baseline.lua b/examples/swb_baseline.lua
index 28cc6d5..8b7e01a 100644
--- a/examples/swb_baseline.lua
+++ b/examples/swb_baseline.lua
@@ -6,14 +6,10 @@ gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9,
         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",
-        global_transf = "/slfs1/users/mfy43/swb_global_transf.nerv",
-        initialized_param = "/slfs1/users/mfy43/swb_init.nerv",
+        initialized_param = {"/slfs1/users/mfy43/swb_init.nerv",
+                "/slfs1/users/mfy43/swb_global_transf.nerv"},
         debug = false}
 
-function make_param_repo(param_file)
-    return nerv.ParamRepo({param_file, gconf.global_transf})
-end
-
 function make_sublayer_repo(param_repo)
     return nerv.LayerRepo(
     {
@@ -60,7 +56,7 @@ function make_sublayer_repo(param_repo)
         },
         ["nerv.SoftmaxCELayer"] =
         {
-            criterion = {{}, {dim_in = {3001, 1}, dim_out = {}, compressed = true}}
+            ce_crit = {{}, {dim_in = {3001, 1}, dim_out = {1}, compressed = true}}
         }
     }, param_repo, gconf)
 end
@@ -82,7 +78,7 @@ function make_layer_repo(sublayer_repo, param_repo)
                 }
             }},
             main = {{}, {
-                dim_in = {429, 1}, dim_out = {},
+                dim_in = {429, 1}, dim_out = {1},
                 sub_layers = sublayer_repo,
                 connections = {
                     ["<input>[1]"] = "affine0[1]",
@@ -100,8 +96,9 @@ function make_layer_repo(sublayer_repo, param_repo)
                     ["sigmoid5[1]"] = "affine6[1]",
                     ["affine6[1]"] = "sigmoid6[1]",
                     ["sigmoid6[1]"] = "affine7[1]",
-                    ["affine7[1]"] = "criterion[1]",
-                    ["<input>[2]"] = "criterion[2]"
+                    ["affine7[1]"] = "ce_crit[1]",
+                    ["<input>[2]"] = "ce_crit[2]",
+                    ["ce_crit[1]"] = "<output>[1]"
                 }
             }}
         }
@@ -109,55 +106,61 @@ function make_layer_repo(sublayer_repo, param_repo)
 end
 
 function get_criterion_layer(sublayer_repo)
-    return sublayer_repo:get_layer("criterion")
+    return sublayer_repo:get_layer("ce_crit")
 end
 
 function get_network(layer_repo)
     return layer_repo:get_layer("main")
 end
 
-function make_reader(scp_file, layer_repo)
-    return nerv.TNetReader(gconf,
-        {
-            id = "main_scp",
-            scp_file = scp_file,
-            conf_file = gconf.htk_conf,
-            frm_ext = gconf.frm_ext,
-            mlfs = {
-                phone_state = {
-                    file = "/slfs1/users/mfy43/swb_ivec/ref.mlf",
-                    format = "map",
-                    format_arg = "/slfs1/users/mfy43/swb_ivec/dict",
-                    dir = "*/",
-                    ext = "lab"
-                }
-            },
-            global_transf = layer_repo:get_layer("global_transf")
-        })
+function make_readers(scp_file, layer_repo)
+    return {
+                {reader = nerv.TNetReader(gconf,
+                    {
+                        id = "main_scp",
+                        scp_file = scp_file,
+                        conf_file = gconf.htk_conf,
+                        frm_ext = gconf.frm_ext,
+                        mlfs = {
+                            phone_state = {
+                                file = "/slfs1/users/mfy43/swb_ivec/ref.mlf",
+                                format = "map",
+                                format_arg = "/slfs1/users/mfy43/swb_ivec/dict",
+                                dir = "*/",
+                                ext = "lab"
+                            }
+                        },
+                        global_transf = layer_repo:get_layer("global_transf")
+                    }),
+                data = {main_scp = 429, phone_state = 1}}
+            }
 end
 
-function make_buffer(reader, buffer)
+function make_buffer(readers)
     return nerv.SGDBuffer(gconf,
         {
             buffer_size = gconf.buffer_size,
             randomize = gconf.randomize,
-            readers = {
-                { reader = reader,
-                  data = {main_scp = 429, phone_state = 1}}
-            }
+            readers = readers
         })
 end
 
-function get_accuracy(crit)
-    return crit.total_correct / crit.total_frames * 100
+function get_input_order()
+    return {"main_scp", "phone_state"}
+end
+
+function get_accuracy(sublayer_repo)
+    local ce_crit = sublayer_repo:get_layer("ce_crit")
+    return ce_crit.total_correct / ce_crit.total_frames * 100
 end
 
-function print_stat(crit)
+function print_stat(sublayer_repo)
+    local ce_crit = sublayer_repo:get_layer("ce_crit")
     nerv.info("*** training stat begin ***")
-    nerv.utils.printf("cross entropy:\t\t%.8f\n", crit.total_ce)
-    nerv.utils.printf("correct:\t\t%d\n", crit.total_correct)
-    nerv.utils.printf("frames:\t\t\t%d\n", crit.total_frames)
-    nerv.utils.printf("err/frm:\t\t%.8f\n", crit.total_ce / crit.total_frames)
-    nerv.utils.printf("accuracy:\t\t%.3f%%\n", get_accuracy(crit))
+    nerv.printf("cross entropy:\t\t%.8f\n", ce_crit.total_ce)
+    nerv.printf("correct:\t\t%d\n", ce_crit.total_correct)
+    nerv.printf("frames:\t\t\t%d\n", ce_crit.total_frames)
+    nerv.printf("err/frm:\t\t%.8f\n", ce_crit.total_ce / ce_crit.total_frames)
+    nerv.printf("accuracy:\t\t%.3f%%\n", get_accuracy(sublayer_repo))
     nerv.info("*** training stat end ***")
 end
diff --git a/examples/test_dnn_layers.lua b/examples/test_dnn_layers.lua
index bf81f7b..64c0dec 100644
--- a/examples/test_dnn_layers.lua
+++ b/examples/test_dnn_layers.lua
@@ -69,8 +69,8 @@ for i = 0, 3 do
     print(err_output1[1])
     print("err_output2")
     print(err_output2[1])
-    nerv.utils.printf("cross entropy: %.8f\n", sm.total_ce)
-    nerv.utils.printf("frames: %.8f\n", sm.total_frames)
+    nerv.printf("cross entropy: %.8f\n", sm.total_ce)
+    nerv.printf("frames: %.8f\n", sm.total_frames)
 end
 print("linear")
 print(af.ltp.trans)
diff --git a/examples/test_nn_lib.lua b/examples/test_nn_lib.lua
index 6fdbd67..5444810 100644
--- a/examples/test_nn_lib.lua
+++ b/examples/test_nn_lib.lua
@@ -144,17 +144,17 @@ for data in buffer.get_data, buffer do
     main:back_propagate(err_output, err_input, input, output)
     main:update(err_input, input, output)
 
---    nerv.utils.printf("cross entropy: %.8f\n", sm.total_ce)
---    nerv.utils.printf("correct: %d\n", sm.total_correct)
---    nerv.utils.printf("frames: %d\n", sm.total_frames)
---    nerv.utils.printf("err/frm: %.8f\n", sm.total_ce / sm.total_frames)
---    nerv.utils.printf("accuracy: %.8f\n", sm.total_correct / sm.total_frames)
+--    nerv.printf("cross entropy: %.8f\n", sm.total_ce)
+--    nerv.printf("correct: %d\n", sm.total_correct)
+--    nerv.printf("frames: %d\n", sm.total_frames)
+--    nerv.printf("err/frm: %.8f\n", sm.total_ce / sm.total_frames)
+--    nerv.printf("accuracy: %.8f\n", sm.total_correct / sm.total_frames)
     collectgarbage("collect")
 end
-nerv.utils.printf("cross entropy: %.8f\n", sm.total_ce)
-nerv.utils.printf("correct: %d\n", sm.total_correct)
-nerv.utils.printf("accuracy: %.3f%%\n", sm.total_correct / sm.total_frames * 100)
-nerv.utils.printf("writing back...\n")
+nerv.printf("cross entropy: %.8f\n", sm.total_ce)
+nerv.printf("correct: %d\n", sm.total_correct)
+nerv.printf("accuracy: %.3f%%\n", sm.total_correct / sm.total_frames * 100)
+nerv.printf("writing back...\n")
 cf = nerv.ChunkFile("output.nerv", "w")
 for i, p in ipairs(main:get_params()) do
     print(p)
diff --git a/io/sgd_buffer.lua b/io/sgd_buffer.lua
index 381b863..f4f7dfe 100644
--- a/io/sgd_buffer.lua
+++ b/io/sgd_buffer.lua
@@ -41,7 +41,7 @@ function SGDBuffer:saturate()
                 buff.data:copy_from(buff.leftover, 0, lrow)
                 buff.leftover = nil
             end
-            nerv.utils.printf("leftover: %d\n", lrow)
+            nerv.printf("leftover: %d\n", lrow)
             reader.tail = lrow
             reader.has_leftover = false
         end
diff --git a/layer/affine.lua b/layer/affine.lua
index 2cd7acb..00cbcfb 100644
--- a/layer/affine.lua
+++ b/layer/affine.lua
@@ -3,13 +3,35 @@ local LinearTransParam = nerv.class('nerv.LinearTransParam', 'nerv.MatrixParam')
 local BiasParam = nerv.class('nerv.BiasParam', 'nerv.MatrixParam')
 local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer')
 
-function MatrixParam:read(pcdata)
+function MatrixParam:read(handle)
     self.trans = self.gconf.cumat_type.new_from_host(
-                    nerv.MMatrixFloat.load(pcdata))
+                    nerv.MMatrixFloat.load(handle))
 end
 
-function MatrixParam:write(pfhandle)
-    self.trans:new_to_host():save(pfhandle)
+function MatrixParam:write(handle)
+    self.trans:new_to_host():save(handle)
+end
+
+function MatrixParam:train_init()
+    self.correction = self.trans:create()
+    self.correction:fill(0)
+end
+
+function MatrixParam:update(gradient)
+    local gconf = self.gconf
+    self.correction:add(self.correction, gradient, gconf.momentum, 1.0)
+    -- momentum gain
+    local mmt_gain = 1.0 / (1.0 - gconf.momentum);
+    local n = self.gconf.batch_size * mmt_gain
+    -- perform update
+    self.trans:add(self.trans, self.correction, 1.0, -gconf.lrate / n)
+end
+
+function LinearTransParam:update(gradient)
+    MatrixParam.update(self, gradient)
+    local gconf = self.gconf
+    -- weight decay
+    self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost)
 end
 
 function AffineLayer:__init(id, global_conf, layer_conf)
@@ -20,9 +42,10 @@ function AffineLayer:__init(id, global_conf, layer_conf)
     self.dim_out = layer_conf.dim_out
     self.gconf = global_conf
     self:check_dim_len(1, 1) -- exactly one input and one output
+    self.direct_update = layer_conf.direct_update
 end
 
-function AffineLayer:init()
+function AffineLayer:init(batch_size)
     if self.ltp.trans:ncol() ~= self.bp.trans:ncol() then
         nerv.error("mismatching dimensions of linear transform and bias paramter")
     end
@@ -32,32 +55,24 @@ function AffineLayer:init()
     if self.dim_out[1] ~= self.ltp.trans:ncol() then
         nerv.error("mismatching dimensions of linear transform parameter and output")
     end
-
-    -- linear transform correction
-    self.ltc = self.ltp.trans:create()
-    self.ltc:fill(0)
-    -- bias correction
-    self.bc = self.bp.trans:create()
-    self.bc:fill(0)
+    self.ltp_grad = self.ltp.trans:create()
+    self.ltp:train_init()
+    self.bp:train_init()
 end
 
 function AffineLayer:update(bp_err, input, output)
-    local ltp = self.ltp.trans
-    local bp = self.bp.trans
-    local ltc = self.ltc
-    local bc = self.bc
-    local gconf = self.gconf
-    -- momentum gain
-    local mmt_gain = 1.0 / (1.0 - gconf.momentum);
-    local n = input[1]:nrow() * mmt_gain
-    -- update corrections (accumulated errors)
-    ltc:mul(input[1], bp_err[1], 1.0, gconf.momentum, 'T', 'N')
-    bc:add(bc, bp_err[1]:colsum(), gconf.momentum, 1.0)
-    -- perform update
-    ltp:add(ltp, ltc, 1.0, -gconf.lrate / n)
-    bp:add(bp, bc, 1.0, -gconf.lrate / n)
-    -- weight decay
-    ltp:add(ltp, ltp, 1.0, -gconf.lrate * gconf.wcost)
+    if self.direct_update then
+        self.ltp.correction:mul(input[1], bp_err[1], 1.0, gconf.momentum, 'T', 'N')
+        -- momentum gain
+        local mmt_gain = 1.0 / (1.0 - gconf.momentum);
+        local n = self.gconf.batch_size * mmt_gain
+        -- perform update
+        self.ltp.trans:add(self.ltp.trans, self.ltp.correction, 1.0, -gconf.lrate / n)
+    else
+        self.ltp_grad:mul(input[1], bp_err[1], 1.0, 0.0, 'T', 'N')
+        self.ltp:update(self.ltp_grad)
+    end
+    self.bp:update(bp_err[1]:colsum())
 end
 
 function AffineLayer:propagate(input, output)
@@ -67,10 +82,10 @@ function AffineLayer:propagate(input, output)
     output[1]:add_row(self.bp.trans, 1.0)
 end
 
-function AffineLayer:back_propagate(next_bp_err, bp_err, input, output)
+function AffineLayer:back_propagate(bp_err, next_bp_err, input, output)
     next_bp_err[1]:mul(bp_err[1], self.ltp.trans, 1.0, 0.0, 'N', 'T')
 end
 
 function AffineLayer:get_params()
-    return {self.ltp, self.bp}
+    return nerv.ParamRepo({self.ltp, self.bp})
 end
diff --git a/layer/bias.lua b/layer/bias.lua
index 8cd326b..c99274d 100644
--- a/layer/bias.lua
+++ b/layer/bias.lua
@@ -24,5 +24,5 @@ function BiasLayer:propagate(input, output)
 end
 
 function BiasLayer:get_params()
-    return {self.bias}
+    return nerv.ParamRepo({self.bias})
 end
diff --git a/layer/combiner.lua b/layer/combiner.lua
index 75e47e2..7bd7617 100644
--- a/layer/combiner.lua
+++ b/layer/combiner.lua
@@ -7,9 +7,15 @@ function CombinerLayer:__init(id, global_conf, layer_conf)
     self.dim_out = layer_conf.dim_out
     self.gconf = global_conf
     self:check_dim_len(#self.lambda, -1)
+    if #self.dim_in < 1 then
+        nerv.error("no input specified")
+    end
+    if #self.dim_out < 1 then
+        nerv.error("no output specified")
+    end
 end
 
-function CombinerLayer:init()
+function CombinerLayer:init(batch_size)
     local dim = self.dim_in[1]
     for i = 2, #self.dim_in do
         if self.dim_in[i] ~= dim then
@@ -21,6 +27,7 @@ function CombinerLayer:init()
             nerv.error("mismatching dimensions of inputs/outputs")
         end
     end
+    self.sum = self.gconf.cumat_type(batch_size, dim)
 end
 
 function CombinerLayer:update(bp_err, input, output)
@@ -32,24 +39,21 @@ function CombinerLayer:propagate(input, output)
         output[1]:add(output[1], input[i], 1.0, self.lambda[i])
     end
     for i = 2, #self.dim_out do
-        output[i]:copy_fromd(output[1]) 
+        output[i]:copy_fromd(output[1])
     end
 end
 
-function CombinerLayer:back_propagate(next_bp_err, bp_err, input, output)
-    local sum = bp_err[1]:create()
-    sum:fill(0)
-    for i = 1, #self.dim_out do
+function CombinerLayer:back_propagate(bp_err, next_bp_err, input, output)
+    local sum = self.sum
+    sum:copy_fromd(bp_err[1])
+    for i = 2, #self.dim_out do
         sum:add(sum, bp_err[i], 1.0, 1.0)
     end
     for i = 1, #self.dim_in do
-        local scale = nerv.CuMatrixFloat(sum:nrow(), 1)
-        scale:fill(self.lambda[i])
-        next_bp_err[i]:copy_fromd(sum)
-        next_bp_err[i]:scale_rows_by_col(scale)
+        next_bp_err[i]:add(next_bp_err[i], sum, 0.0, self.lambda[i])
     end
 end
 
 function CombinerLayer:get_params()
-    return {}
+    return nerv.ParamRepo({})
 end
diff --git a/layer/init.lua b/layer/init.lua
index 169427d..e39af94 100644
--- a/layer/init.lua
+++ b/layer/init.lua
@@ -15,11 +15,15 @@ function Param:set_info(info)
     self.info = info
 end
 
-function Param:read(pfhandle)
+function Param:read(handle)
     nerv.error_method_not_implemented()
 end
 
-function Param:write(pfhandle)
+function Param:write(handle)
+    nerv.error_method_not_implemented()
+end
+
+function Param:update(gradient)
     nerv.error_method_not_implemented()
 end
 
@@ -29,7 +33,7 @@ function Layer:__init(id, global_conf, layer_conf)
     nerv.error_method_not_implemented()
 end
 
-function Layer:init()
+function Layer:init(batch_size)
     nerv.error_method_not_implemented()
 end
 
@@ -41,7 +45,7 @@ function Layer:propagate(input, output)
     nerv.error_method_not_implemented()
 end
 
-function Layer:back_propagate(next_bp_err, bp_err, input, output)
+function Layer:back_propagate(bp_err, next_bp_err, input, output)
     nerv.error_method_not_implemented()
 end
 
diff --git a/layer/mse.lua b/layer/mse.lua
index da5b24d..9a97add 100644
--- a/layer/mse.lua
+++ b/layer/mse.lua
@@ -8,12 +8,16 @@ function MSELayer:__init(id, global_conf, layer_conf)
     self:check_dim_len(2, -1)
 end
 
-function MSELayer:init()
+function MSELayer:init(batch_size)
     if self.dim_in[1] ~= self.dim_in[2] then
         nerv.error("mismatching dimensions of previous network output and labels")
     end
+    self.scale = 1 / self.dim_in[1]
     self.total_mse = 0.0
     self.total_frames = 0
+    self.mse = self.gconf.cumat_type(batch_size, self.dim_in[1])
+    self.mse_sum = self.gconf.cumat_type(batch_size, 1)
+    self.diff = self.mse:create()
 end
 
 function MSELayer:update(bp_err, input, output)
@@ -21,32 +25,28 @@ function MSELayer:update(bp_err, input, output)
 end
 
 function MSELayer:propagate(input, output)
-    local mse = input[1]:create()
+    local mse = self.mse
+    local mse_sum = self.mse_sum
     mse:add(input[1], input[2], 1.0, -1.0)
-    self.diff = mse:create()
     self.diff:copy_fromd(mse)
     mse:mul_elem(mse, mse)
-    mse = mse:rowsum(mse)
-    local scale = nerv.CuMatrixFloat(mse:nrow(), 1)
-    scale:fill(1 / input[1]:ncol())
-    mse:scale_rows_by_col(scale)
+    mse_sum:add(mse_sum, mse:rowsum(mse), 0.0, self.scale)
     if output[1] ~= nil then
-        output[1]:copy_fromd(mse)
+        output[1]:copy_fromd(mse_sum)
     end
-    self.total_mse = self.total_mse + mse:colsum()[0]
-    self.total_frames = self.total_frames + mse:nrow()
+    self.total_mse = self.total_mse + mse_sum:colsum()[0]
+    self.total_frames = self.total_frames + mse_sum:nrow()
 end
 
 -- NOTE: must call propagate before back_propagate
-function MSELayer:back_propagate(next_bp_err, bp_err, input, output)
+function MSELayer:back_propagate(bp_err, next_bp_err, input, output)
     local nbe = next_bp_err[1]
-    nbe:copy_fromd(self.diff)
-    self.diff = nil
+    nbe:add(nbe, self.diff, 0.0, 2 * self.scale)
     if bp_err[1] ~= nil then
         nbe:scale_rows_by_col(bp_err[1])
     end
 end
 
 function MSELayer:get_params()
-    return {}
+    return nerv.ParamRepo({})
 end
diff --git a/layer/sigmoid.lua b/layer/sigmoid.lua
index dd10fb9..dfd09eb 100644
--- a/layer/sigmoid.lua
+++ b/layer/sigmoid.lua
@@ -22,10 +22,10 @@ function SigmoidLayer:propagate(input, output)
     output[1]:sigmoid(input[1])
 end
 
-function SigmoidLayer:back_propagate(next_bp_err, bp_err, input, output)
+function SigmoidLayer:back_propagate(bp_err, next_bp_err, input, output)
     next_bp_err[1]:sigmoid_grad(bp_err[1], output[1])
 end
 
 function SigmoidLayer:get_params()
-    return {}
+    return nerv.ParamRepo({})
 end
diff --git a/layer/softmax_ce.lua b/layer/softmax_ce.lua
index 7888540..daf891e 100644
--- a/layer/softmax_ce.lua
+++ b/layer/softmax_ce.lua
@@ -12,13 +12,15 @@ function SoftmaxCELayer:__init(id, global_conf, layer_conf)
     self:check_dim_len(2, -1) -- two inputs: nn output and label
 end
 
-function SoftmaxCELayer:init()
+function SoftmaxCELayer:init(batch_size)
     if not self.compressed and (self.dim_in[1] ~= self.dim_in[2]) then
         nerv.error("mismatching dimensions of previous network output and labels")
     end
     self.total_ce = 0.0
     self.total_correct = 0
     self.total_frames = 0
+    self.softmax = self.gconf.cumat_type(batch_size, self.dim_in[1])
+    self.ce = self.softmax:create()
 end
 
 function SoftmaxCELayer:update(bp_err, input, output)
@@ -26,12 +28,11 @@ function SoftmaxCELayer:update(bp_err, input, output)
 end
 
 function SoftmaxCELayer:propagate(input, output)
-    local soutput = input[1]:create()  -- temporary value for calc softmax
-    self.soutput = soutput
-    local classified = soutput:softmax(input[1])
-    local ce = soutput:create()
-    ce:log_elem(soutput)
+    local softmax = self.softmax
+    local ce = self.ce
+    local classified = softmax:softmax(input[1])
     local label = input[2]
+    ce:log_elem(softmax)
     if self.compressed then
         label = label:decompress(input[1]:ncol())
     end
@@ -42,26 +43,26 @@ function SoftmaxCELayer:propagate(input, output)
     end
     -- add total ce
     self.total_ce = self.total_ce - ce:colsum()[0]
-    self.total_frames = self.total_frames + soutput:nrow()
+    self.total_frames = self.total_frames + softmax:nrow()
     -- TODO: add colsame for uncompressed label
     if self.compressed then
         self.total_correct = self.total_correct + classified:colsame(input[2])[0]
     end
 end
 
-function SoftmaxCELayer:back_propagate(next_bp_err, bp_err, input, output)
+function SoftmaxCELayer:back_propagate(bp_err, next_bp_err, input, output)
     -- softmax output - label
     local label = input[2]
     if self.compressed then
         label = label:decompress(input[1]:ncol())
     end
     local nbe = next_bp_err[1]
-    nbe:add(self.soutput, label, 1.0, -1.0)
+    nbe:add(self.softmax, label, 1.0, -1.0)
     if bp_err[1] ~= nil then
         nbe:scale_rows_by_col(bp_err[1])
     end
 end
 
 function SoftmaxCELayer:get_params()
-    return {}
+    return nerv.ParamRepo({})
 end
diff --git a/layer/window.lua b/layer/window.lua
index 3a093f4..4e9a3b1 100644
--- a/layer/window.lua
+++ b/layer/window.lua
@@ -24,5 +24,5 @@ function WindowLayer:propagate(input, output)
 end
 
 function WindowLayer:get_params()
-    return {self.window}
+    return nerv.ParamRepo({self.window})
 end
diff --git a/nerv.lua b/nerv.lua
index 467d926..a69dda6 100644
--- a/nerv.lua
+++ b/nerv.lua
@@ -1,20 +1,35 @@
 require 'libnerv'
-nerv.utils = require 'pl.utils'
 
 function nerv.error(fmt, ...)
-    error(nerv.utils.printf("[nerv] internal error: " .. fmt .. "\n", ...))
+    error(nerv.printf("[nerv] internal error: " .. fmt .. "\n", ...))
 end
 
 function nerv.error_method_not_implemented()
     nerv.error("method not implemented");
 end
 
+function nerv.printf(fmt, ...)
+    io.write(string.format(fmt, ...))
+end
+
+function nerv.mesg_with_timestamp(fmt, ...)
+    nerv.printf(
+        string.format("(%s)[nerv] info: %s\n",
+            os.date("%H:%M:%S %F"), fmt), ...)
+end
+
 function nerv.info(fmt, ...)
-    nerv.utils.printf(
+    nerv.printf(
         string.format("(%s)[nerv] info: %s\n",
             os.date("%H:%M:%S %F"), fmt), ...)
 end
 
+function nerv.warning(fmt, ...)
+    nerv.printf(
+        string.format("(%s)[nerv] warning: %s\n",
+            os.date("%H:%M:%S %F"), fmt), ...)
+end
+
 -- Torch C API wrapper
 function nerv.class(tname, parenttname)
 
@@ -77,8 +92,20 @@ function table.tostring(tbl)
   return "{" .. table.concat(result, ",") .. "}"
 end
 
-function nerv.get_type(typename)
-    return assert(loadstring("return " .. typename))()
+function nerv.get_type(tname)
+    return assert(loadstring("return " .. tname))()
+end
+
+function nerv.is_type(obj, tname)
+    local mt0 = nerv.getmetatable(tname)
+    local mt = getmetatable(obj)
+    while mt do
+        if mt == mt0 then
+            return true
+        end
+        mt = getmetatable(mt)
+    end
+    return false
 end
 
 require 'matrix.init'
diff --git a/nn/layer_dag.lua b/nn/layer_dag.lua
index 2dda7c9..8e30216 100644
--- a/nn/layer_dag.lua
+++ b/nn/layer_dag.lua
@@ -85,13 +85,14 @@ function DAGLayer:__init(id, global_conf, layer_conf)
         end
     end
 
+    -- topology sort
     local queue = {}
     local l = 1
     local r = 1
     for id, ref in pairs(layers) do
         if ref.in_deg == 0 then
             table.insert(queue, ref)
-            nerv.utils.printf("adding source layer: %s\n", id)
+            nerv.info("adding source layer: %s", id)
             r = r + 1
         end
     end
@@ -111,13 +112,13 @@ function DAGLayer:__init(id, global_conf, layer_conf)
         end
     end
     for i = 1, #queue do
-        nerv.utils.printf("queued layer: %s\n", queue[i].layer.id)
+        nerv.info("enqueued layer: %s", queue[i].layer.id)
     end
 
     for id, ref in pairs(layers) do
         -- check wether the graph is connected
         if ref.visited == false then
-            nerv.utils.printf("warning: layer %s is ignored\n", id)
+            nerv.warning("layer %s is ignored", id)
         end
     end
 
@@ -131,7 +132,7 @@ function DAGLayer:__init(id, global_conf, layer_conf)
     self.gconf = global_conf
 end
 
-function DAGLayer:init(batch_size) -- topology sort
+function DAGLayer:init(batch_size)
     for i, conn in ipairs(self.parsed_conn) do
         local _, output_dim
         local ref_from, port_from, ref_to, port_to
@@ -160,7 +161,7 @@ function DAGLayer:init(batch_size) -- topology sort
             end
         end
         -- initialize sub layers
-        ref.layer:init()
+        ref.layer:init(batch_size)
     end
     for i = 1, #self.dim_in do
         if self.inputs[i] == nil then
@@ -227,7 +228,7 @@ function DAGLayer:propagate(input, output)
     end
 end
 
-function DAGLayer:back_propagate(next_bp_err, bp_err, input, output)
+function DAGLayer:back_propagate(bp_err, next_bp_err, input, output)
     self:set_err_outputs(next_bp_err)
     self:set_err_inputs(bp_err)
     self:set_inputs(input)
@@ -235,16 +236,14 @@ function DAGLayer:back_propagate(next_bp_err, bp_err, input, output)
     for i = #self.queue, 1, -1 do
         local ref = self.queue[i]
         -- print(ref.layer.id)
-        ref.layer:back_propagate(ref.err_outputs, ref.err_inputs, ref.inputs, ref.outputs)
+        ref.layer:back_propagate(ref.err_inputs, ref.err_outputs, ref.inputs, ref.outputs)
     end
 end
 
 function DAGLayer:get_params()
-    local res = {}
+    local param_repos = {}
     for id, ref in pairs(self.queue) do