Merge pull request #6 from yimmon/master

support batch_resize (sequence training related layers are in kaldi_seq)

15 files changed, 539 insertions, 5 deletions

diff --git a/nerv/Makefile b/nerv/Makefile
index b5d26bd..b449f82 100644
--- a/nerv/Makefile
+++ b/nerv/Makefile
@@ -31,7 +31,7 @@ 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/softmax.lua \
-			layer/window.lua layer/bias.lua layer/combiner.lua layer/mse.lua layer/affine_recurrent.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 
 
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/layer/affine.lua b/nerv/layer/affine.lua
index b81b2a4..015ec3f 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -60,6 +60,10 @@ function AffineLayer:init(batch_size)
     self.bp:train_init()
 end
 
+function AffineLayer:batch_resize(batch_size)
+    -- do nothing
+end
+
 function AffineLayer:update(bp_err, input, output)
     if self.direct_update then
         self.ltp.correction:mul(input[1], bp_err[1], 1.0, gconf.momentum, 'T', 'N')
diff --git a/nerv/layer/affine_recurrent.lua b/nerv/layer/affine_recurrent.lua
index 59d259c..92d98e2 100644
--- a/nerv/layer/affine_recurrent.lua
+++ b/nerv/layer/affine_recurrent.lua
@@ -37,6 +37,10 @@ function Recurrent:init(batch_size)
     self.bp:train_init()
 end
 
+function Recurrent:batch_resize(batch_size)
+    -- do nothing
+end
+
 function Recurrent:update(bp_err, input, output)
     if (self.direct_update == true) then
         local ltp_hh = self.ltp_hh.trans
diff --git a/nerv/layer/bias.lua b/nerv/layer/bias.lua
index c99274d..7e9fd46 100644
--- a/nerv/layer/bias.lua
+++ b/nerv/layer/bias.lua
@@ -18,6 +18,10 @@ function BiasLayer:init()
     end
 end
 
+function BiasLayer:batch_resize(batch_size)
+    -- do nothing
+end
+
 function BiasLayer:propagate(input, output)
     output[1]:copy_fromd(input[1])
     output[1]:add_row(self.bias.trans, 1.0)
diff --git a/nerv/layer/combiner.lua b/nerv/layer/combiner.lua
index 7bd7617..1bcfdfb 100644
--- a/nerv/layer/combiner.lua
+++ b/nerv/layer/combiner.lua
@@ -30,6 +30,12 @@ function CombinerLayer:init(batch_size)
     self.sum = self.gconf.cumat_type(batch_size, dim)
 end
 
+function CombinerLayer:batch_resize(batch_size)
+    if self.sum:nrow() ~= batch_size then
+        self.sum = self.gconf.cumat_type(batch_size, self.dim_in[1])
+    end
+end
+
 function CombinerLayer:update(bp_err, input, output)
 end
 
diff --git a/nerv/layer/mse.lua b/nerv/layer/mse.lua
index 2516998..0ee3080 100644
--- a/nerv/layer/mse.lua
+++ b/nerv/layer/mse.lua
@@ -20,6 +20,14 @@ function MSELayer:init(batch_size)
     self.diff = self.mse:create()
 end
 
+function MSELayer:batch_resize(batch_size)
+    if self.mse:nrow() ~= batch_resize then
+        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
+end
+
 function MSELayer:update(bp_err, input, output)
     -- no params, therefore do nothing
 end
diff --git a/nerv/layer/sigmoid.lua b/nerv/layer/sigmoid.lua
index dfd09eb..0a8bcdc 100644
--- a/nerv/layer/sigmoid.lua
+++ b/nerv/layer/sigmoid.lua
@@ -14,6 +14,10 @@ function SigmoidLayer:init()
     end
 end
 
+function SigmoidLayer:batch_resize(batch_size)
+    -- do nothing
+end
+
 function SigmoidLayer:update(bp_err, input, output)
     -- no params, therefore do nothing
 end
diff --git a/nerv/layer/softmax.lua b/nerv/layer/softmax.lua
index e979ebf..4205b66 100644
--- a/nerv/layer/softmax.lua
+++ b/nerv/layer/softmax.lua
@@ -14,6 +14,10 @@ function SoftmaxLayer:init(batch_size)
     end
 end
 
+function SoftmaxLayer:batch_resize(batch_size)
+    -- do nothing
+end
+
 function SoftmaxLayer:update(bp_err, input, output)
     -- no params, therefore do nothing
 end
diff --git a/nerv/layer/softmax_ce.lua b/nerv/layer/softmax_ce.lua
index f878a2f..9071e86 100644
--- a/nerv/layer/softmax_ce.lua
+++ b/nerv/layer/softmax_ce.lua
@@ -23,6 +23,13 @@ function SoftmaxCELayer:init(batch_size)
     self.ce = self.softmax:create()
 end
 
+function SoftmaxCELayer:batch_resize(batch_size)
+    if self.softmax:nrow() ~= batch_resize then
+        self.softmax = self.gconf.cumat_type(batch_size, self.dim_in[1])
+        self.ce = self.softmax:create()
+    end
+end
+
 function SoftmaxCELayer:update(bp_err, input, output)
     -- no params, therefore do nothing
 end
diff --git a/nerv/layer/window.lua b/nerv/layer/window.lua
index 4e9a3b1..8eed352 100644
--- a/nerv/layer/window.lua
+++ b/nerv/layer/window.lua
@@ -18,6 +18,10 @@ function WindowLayer:init()
     end
 end
 
+function WindowLayer:batch_resize(batch_size)
+    -- do nothing
+end
+
 function WindowLayer:propagate(input, output)
     output[1]:copy_fromd(input[1])
     output[1]:scale_rows_by_row(self.window.trans)
diff --git a/nerv/lib/matrix/generic/matrix.c b/nerv/lib/matrix/generic/matrix.c
index 6cb3dc0..4246751 100644
--- a/nerv/lib/matrix/generic/matrix.c
+++ b/nerv/lib/matrix/generic/matrix.c
@@ -4,6 +4,7 @@
 /* FIXME: malloc failure detection */
 
 void nerv_matrix_(data_free)(Matrix *self, Status *status) {
+    if(*self->data_ref == 0) return; /* FIXME: repeat free memory */
     assert(*self->data_ref > 0);
     if (--(*self->data_ref) == 0)
     {
diff --git a/nerv/nn/layer_dag.lua b/nerv/nn/layer_dag.lua
index f69d31c..73bb77d 100644
--- a/nerv/nn/layer_dag.lua
+++ b/nerv/nn/layer_dag.lua
@@ -79,7 +79,7 @@ function DAGLayer:__init(id, global_conf, layer_conf)
             end
 
             table.insert(parsed_conn,
-                {{ref_from, port_from}, {ref_to, port_to}})
+            {{ref_from, port_from}, {ref_to, port_to}})
             table.insert(ref_from.next_layers, ref_to) -- add edge
             ref_to.in_deg = ref_to.in_deg + 1          -- increase the in-degree of the target layer
         end
@@ -140,8 +140,11 @@ function DAGLayer:init(batch_size)
         ref_from, port_from = unpack(conn[1])
         ref_to, port_to = unpack(conn[2])
         _, output_dim = ref_from.layer:get_dim()
-        local mid = self.gconf.cumat_type(batch_size,
-                                        output_dim[port_from])
+        local dim = 1
+        if output_dim[port_from] > 0 then
+            dim = output_dim[port_from]
+        end
+        local mid = self.gconf.cumat_type(batch_size, dim)
         local err_mid = mid:create()
 
         ref_from.outputs[port_from] = mid
@@ -176,6 +179,33 @@ function DAGLayer:init(batch_size)
     end
 end
 
+function DAGLayer:batch_resize(batch_size)
+    self.gconf.batch_size = batch_size
+
+    for i, conn in ipairs(self.parsed_conn) do
+        local _, output_dim
+        local ref_from, port_from, ref_to, port_to
+        ref_from, port_from = unpack(conn[1])
+        ref_to, port_to = unpack(conn[2])
+        _, output_dim = ref_from.layer:get_dim()
+
+        if ref_from.outputs[port_from]:nrow() ~= batch_size and output_dim[port_from] > 0 then
+            local mid = self.gconf.cumat_type(batch_size, output_dim[port_from])
+            local err_mid = mid:create()
+
+            ref_from.outputs[port_from] = mid
+            ref_to.inputs[port_to] = mid
+
+            ref_from.err_inputs[port_from] = err_mid
+            ref_to.err_outputs[port_to] = err_mid
+        end
+    end
+    for id, ref in pairs(self.layers) do
+        ref.layer:batch_resize(batch_size)
+    end
+    collectgarbage("collect")
+end
+
 function DAGLayer:set_inputs(input)
     for i = 1, #self.dim_in do
         if input[i] == nil then
@@ -228,11 +258,13 @@ end
 function DAGLayer:propagate(input, output)
     self:set_inputs(input)
     self:set_outputs(output)
+    local ret = false
     for i = 1, #self.queue do
         local ref = self.queue[i]
         -- print(ref.layer.id)
-        ref.layer:propagate(ref.inputs, ref.outputs)
+        ret = ref.layer:propagate(ref.inputs, ref.outputs)
     end
+    return ret
 end
 
 function DAGLayer:back_propagate(bp_err, next_bp_err, input, output)