Merge branch 'master' of https://github.com/liuq901/nerv into liuq901-master

Conflicts:
	nerv/layer/init.lua
	nerv/nn/layer_repo.lua

10 files changed, 806 insertions, 29 deletions

diff --git a/nerv/Makefile b/nerv/Makefile
index e8bcad2..421eda0 100644
--- a/nerv/Makefile
+++ b/nerv/Makefile
@@ -42,7 +42,8 @@ LUA_LIBS := matrix/init.lua io/init.lua init.lua \
 			layer/init.lua layer/affine.lua layer/sigmoid.lua layer/tanh.lua layer/softmax_ce.lua layer/softmax.lua \
 			layer/window.lua layer/bias.lua layer/combiner.lua layer/mse.lua \
 			layer/elem_mul.lua layer/lstm.lua layer/lstm_gate.lua layer/dropout.lua layer/gru.lua \
-			nn/init.lua nn/layer_repo.lua nn/param_repo.lua nn/layer_dag.lua \
+			layer/graph.lua layer/rnn.lua layer/duplicate.lua layer/identity.lua \
+			nn/init.lua nn/layer_repo.lua nn/param_repo.lua nn/layer_dag.lua nn/network.lua \
 			io/sgd_buffer.lua \
 			tnn/init.lua tnn/sutil.lua tnn/tnn.lua
 
diff --git a/nerv/layer/duplicate.lua b/nerv/layer/duplicate.lua
new file mode 100644
index 0000000..137472b
--- /dev/null
+++ b/nerv/layer/duplicate.lua
@@ -0,0 +1,41 @@
+local DuplicateLayer = nerv.class('nerv.DuplicateLayer', 'nerv.Layer')
+
+function DuplicateLayer:__init(id, global_conf, layer_conf)
+    nerv.Layer.__init(self, id, global_conf, layer_conf)
+    self:check_dim_len(1, -1)
+    if #self.dim_out < 1 then
+        nerv.error('no output specified')
+    end
+    for i = 1, #self.dim_out do
+        if self.dim_out[i] ~= self.dim_in[1] then
+            nerv.error('mismatching dimensions of outputs')
+        end
+    end
+end
+
+function DuplicateLayer:init()
+end
+
+function DuplicateLayer:batch_resize()
+end
+
+function DuplicateLayer:propagate(input, output)
+    for i = 1, #self.dim_out do
+        output[i]:copy_from(input[1])
+        -- FIXME: use reference copy to speed up
+    end
+end
+
+function DuplicateLayer:back_propagate(bp_err, next_bp_err)
+    next_bp_err[1]:copy_from(bp_err[1])
+    for i = 2, #self.dim_out do
+        next_bp_err[1]:add(next_bp_err[1], bp_err[i], 1.0, 1.0)
+    end
+end
+
+function DuplicateLayer:update()
+end
+
+function DuplicateLayer:get_params()
+    return nerv.ParamRepo({}, self.loc_type)
+end
diff --git a/nerv/layer/graph.lua b/nerv/layer/graph.lua
new file mode 100644
index 0000000..5f42fca
--- /dev/null
+++ b/nerv/layer/graph.lua
@@ -0,0 +1,156 @@
+local GraphLayer = nerv.class('nerv.GraphLayer', 'nerv.Layer')
+
+function GraphLayer:__init(id, global_conf, layer_conf)
+    nerv.Layer.__init(self, id, global_conf, layer_conf)
+    self:graph_init(layer_conf.layer_repo, layer_conf.connections)
+end
+
+local function parse_id(str)
+    local id, port, _
+    _, _, id, port = string.find(str, "([a-zA-Z0-9_.]+)%[([0-9]+)%]")
+    if id == nil or port == nil then
+        _, _, id, port = string.find(str, "(.+)%[([0-9]+)%]")
+        if not (id == "<input>" or id == "<output>") then
+            nerv.error("wrong format of connection id")
+        end
+    end
+    port = tonumber(port)
+    return id, port
+end
+
+function GraphLayer:add_prefix(layers, connections)
+    local function ap(name)
+        return self.id .. '.' .. name
+    end
+
+    for layer_type, sublayers in pairs(layers) do
+        local tmp = {}
+        for name, layer_config in pairs(sublayers) do
+            tmp[ap(name)] = layer_config
+        end
+        layers[layer_type] = tmp
+    end
+
+    for i = 1, #connections do
+        local from, to = connections[i][1], connections[i][2]
+        if parse_id(from) ~= '<input>' then
+            connections[i][1] = ap(from)
+        end
+        if parse_id(to) ~= '<output>' then
+            connections[i][2] = ap(to)
+        end
+    end
+end
+
+function GraphLayer:discover(id, layer_repo)
+    if id == '<output>' then
+        id = '<input>'
+    end
+    local layers = self.layers
+    local ref = layers[id]
+    if ref == nil then
+        local layer = layer_repo:get_layer(id)
+        local dim_in, dim_out = layer:get_dim()
+        self.layer_num = self.layer_num + 1
+        ref = {
+            layer = layer,
+            inputs = {},
+            outputs = {},
+            dim_in = dim_in,
+            dim_out = dim_out,
+            id = self.layer_num,
+        }
+        layers[id] = ref
+    end
+    return ref
+end
+
+function GraphLayer:graph_init(layer_repo, connections)
+    local layers = {}
+    layers['<input>'] = {
+        inputs = {},
+        outputs = {},
+        dim_in = self.dim_out,
+        dim_out = self.dim_in,
+        id = 0,
+    }
+    self.layers = layers
+    self.layer_num = 0
+    self.connections = {}
+
+    -- check data dimension between connected ports
+    for _, edge in pairs(connections) do
+        local from, to, time = edge[1], edge[2], edge[3]
+        local id_from, port_from = parse_id(from)
+        local id_to, port_to = parse_id(to)
+        local ref_from = self:discover(id_from, layer_repo)
+        local ref_to = self:discover(id_to, layer_repo)
+        if ref_from.outputs[port_from] ~= nil then
+            nerv.error('%s has already been attached', from)
+        end
+        if ref_to.inputs[port_to] ~= nil then
+            nerv.error('%s has already been attached', to)
+        end
+        if ref_from.dim_out[port_from] ~= ref_to.dim_in[port_to] then
+            nerv.error('mismatching data dimension between %s and %s', from, to)
+        end
+        if ref_from.id == 0 and ref_to.id == 0 then
+            nerv.error('short-circuit connection between <input> and <output>')
+        end
+        ref_from.outputs[port_from] = true
+        ref_to.inputs[port_to] = true
+        table.insert(self.connections, {ref_from.id, port_from, ref_to.id, port_to, time})
+    end
+
+    -- check dangling ports
+    for id, ref in pairs(layers) do
+        if id ~= '<input>' then
+            for i = 1, #ref.dim_in do
+                if ref.inputs[i] == nil then
+                    nerv.error('dangling input port %d of layer %s', i, id)
+                end
+            end
+            for i = 1, #ref.dim_out do
+                if ref.outputs[i] == nil then
+                   nerv.error('dangling output port %d os layer %s', i, id)
+                end
+            end
+        end
+    end
+    for i = 1, #self.dim_in do
+        if layers['<input>'].outputs[i] == nil then
+            nerv.error('dangling port %d of layer <input>', i)
+        end
+    end
+    for i = 1, #self.dim_out do
+        if layers['<input>'].inputs[i] == nil then
+            nerv.error('dangling port %d of layer <output>', i)
+        end
+    end
+end
+
+function GraphLayer:set_attr(name, value)
+    self[name] = value
+    for id, ref in pairs(self.layers) do
+        if id ~= '<input>' then
+            ref.layer:set_attr(name, value)
+        end
+    end
+end
+
+function GraphLayer:get_sublayer(id)
+    if self.layers[id] == nil or id == '<input>' then
+        nerv.error('layer with id %s not found', id)
+    end
+    return self.layers[id].layer
+end
+
+function GraphLayer:get_params()
+    local param_repos = {}
+    for id, ref in pairs(self.layers) do
+        if id ~= '<input>' then
+            table.insert(param_repos, ref.layer:get_params())
+        end
+    end
+    return nerv.ParamRepo.merge(param_repos, self.loc_type)
+end
diff --git a/nerv/layer/gru.lua b/nerv/layer/gru.lua
index a590a67..71718d7 100644
--- a/nerv/layer/gru.lua
+++ b/nerv/layer/gru.lua
@@ -13,7 +13,7 @@ function GRULayer:__init(id, global_conf, layer_conf)
     -- prepare a DAGLayer to hold the lstm structure
     local pr = layer_conf.pr
     if pr == nil then
-        pr = nerv.ParamRepo(nil, self.loc_type)
+        pr = nerv.ParamRepo({}, self.loc_type)
     end
     
     local function ap(str)
@@ -102,7 +102,7 @@ end
 function GRULayer:bind_params()
     local pr = layer_conf.pr
     if pr == nil then
-        pr = nerv.ParamRepo(nil, self.loc_type)
+        pr = nerv.ParamRepo({}, self.loc_type)
     end
     self.lrepo:rebind(pr)
 end
diff --git a/nerv/layer/identity.lua b/nerv/layer/identity.lua
new file mode 100644
index 0000000..d56337d
--- /dev/null
+++ b/nerv/layer/identity.lua
@@ -0,0 +1,30 @@
+local IdentityLayer = nerv.class('nerv.IdentityLayer', 'nerv.Layer')
+
+function IdentityLayer:__init(id, global_conf, layer_conf)
+    nerv.Layer.__init(self, id, global_conf, layer_conf)
+    self:check_dim_len(1, 1)
+    if self.dim_in[1] ~= self.dim_out[1] then
+        nerv.error('mismatching dimensions of input and output')
+    end
+end
+
+function IdentityLayer:init()
+end
+
+function IdentityLayer:batch_resize()
+end
+
+function IdentityLayer:propagate(input, output)
+    output[1]:copy_from(input[1])
+end
+
+function IdentityLayer:back_propagate(bp_err, next_bp_err)
+    next_bp_err[1]:copy_from(bp_err[1])
+end
+
+function IdentityLayer:update()
+end
+
+function IdentityLayer:get_params()
+    return nerv.ParamRepo({}, self.loc_type)
+end
diff --git a/nerv/layer/init.lua b/nerv/layer/init.lua
index 146ad8c..475ef62 100644
--- a/nerv/layer/init.lua
+++ b/nerv/layer/init.lua
@@ -85,6 +85,14 @@ function Layer:get_dim()
     return self.dim_in, self.dim_out
 end
 
+function Layer:set_attr(name, value)
+    self[name] = value
+end
+
+function Layer:get_sublayer(id)
+    nerv.error('primitive layer does not have sublayers')
+end
+
 function Layer:find_param(plist, lconf, gconf, p_type, p_dim)
     if type(plist) == "string" then
         plist = {plist}
@@ -119,6 +127,7 @@ function Layer:find_param(plist, lconf, gconf, p_type, p_dim)
     return p
 end
 
+nerv.include('graph.lua')
 nerv.include('affine.lua')
 nerv.include('sigmoid.lua')
 nerv.include('tanh.lua')
@@ -133,6 +142,9 @@ nerv.include('lstm.lua')
 nerv.include('lstm_gate.lua')
 nerv.include('dropout.lua')
 nerv.include('gru.lua')
+nerv.include('rnn.lua')
+nerv.include('duplicate.lua')
+nerv.include('identity.lua')
 
 -- The following lines are for backward compatibility, and will be removed in
 -- the future. The use of these names are deprecated.
diff --git a/nerv/layer/lstm.lua b/nerv/layer/lstm.lua
index d4c9212..641d5dc 100644
--- a/nerv/layer/lstm.lua
+++ b/nerv/layer/lstm.lua
@@ -8,7 +8,7 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
     -- prepare a DAGLayer to hold the lstm structure
     local pr = layer_conf.pr
     if pr == nil then
-        pr = nerv.ParamRepo(nil, self.loc_type)
+        pr = nerv.ParamRepo({}, self.loc_type)
     end
     
     local function ap(str)
@@ -18,47 +18,47 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
     local dout1, dout2, dout3 = self.dim_out[1], self.dim_out[2], self.dim_out[3]
     local layers = {
         ["nerv.CombinerLayer"] = {
-            [ap("inputXDup")] = {{}, {dim_in = {din1},
+            [ap("inputXDup")] = {dim_in = {din1},
                                       dim_out = {din1, din1, din1, din1},
-                                      lambda = {1}}},
+                                      lambda = {1}},
 
-            [ap("inputHDup")] = {{}, {dim_in = {din2},
+            [ap("inputHDup")] = {dim_in = {din2},
                                       dim_out = {din2, din2, din2, din2},
-                                      lambda = {1}}},
+                                      lambda = {1}},
 
-            [ap("inputCDup")] = {{}, {dim_in = {din3},
+            [ap("inputCDup")] = {dim_in = {din3},
                                       dim_out = {din3, din3, din3},
-                                      lambda = {1}}},
+                                      lambda = {1}},
 
-            [ap("mainCDup")] = {{}, {dim_in = {din3, din3},
+            [ap("mainCDup")] = {dim_in = {din3, din3},
                                      dim_out = {din3, din3, din3},
-                                     lambda = {1, 1}}},
+                                     lambda = {1, 1}},
         },
         ["nerv.AffineLayer"] = {
-            [ap("mainAffineL")] = {{}, {dim_in = {din1, din2},
+            [ap("mainAffineL")] = {dim_in = {din1, din2},
                                         dim_out = {dout1},
-                                        pr = pr}},
+                                        pr = pr},
         },
         ["nerv.TanhLayer"] = {
-            [ap("mainTanhL")] = {{}, {dim_in = {dout1}, dim_out = {dout1}}},
-            [ap("outputTanhL")] = {{}, {dim_in = {dout1}, dim_out = {dout1}}},
+            [ap("mainTanhL")] = {dim_in = {dout1}, dim_out = {dout1}},
+            [ap("outputTanhL")] = {dim_in = {dout1}, dim_out = {dout1}},
         },
         ["nerv.LSTMGateLayer"] = {
-            [ap("forgetGateL")] = {{}, {dim_in = {din1, din2, din3},
-                                        dim_out = {din3}, pr = pr}},
-            [ap("inputGateL")] = {{}, {dim_in = {din1, din2, din3},
-                                        dim_out = {din3}, pr = pr}},
-            [ap("outputGateL")] = {{}, {dim_in = {din1, din2, din3},
-                                        dim_out = {din3}, pr = pr}},
+            [ap("forgetGateL")] = {dim_in = {din1, din2, din3},
+                                        dim_out = {din3}, pr = pr},
+            [ap("inputGateL")] = {dim_in = {din1, din2, din3},
+                                        dim_out = {din3}, pr = pr},
+            [ap("outputGateL")] = {dim_in = {din1, din2, din3},
+                                        dim_out = {din3}, pr = pr},
 
         },
         ["nerv.ElemMulLayer"] = {
-            [ap("inputGMulL")] = {{}, {dim_in = {din3, din3},
-                                        dim_out = {din3}}},
-            [ap("forgetGMulL")] = {{}, {dim_in = {din3, din3},
-                                        dim_out = {din3}}},
-            [ap("outputGMulL")] = {{}, {dim_in = {din3, din3},
-                                        dim_out = {din3}}},
+            [ap("inputGMulL")] = {dim_in = {din3, din3},
+                                        dim_out = {din3}},
+            [ap("forgetGMulL")] = {dim_in = {din3, din3},
+                                        dim_out = {din3}},
+            [ap("outputGMulL")] = {dim_in = {din3, din3},
+                                        dim_out = {din3}},
         },
     }
     
@@ -114,7 +114,7 @@ end
 function LSTMLayer:bind_params()
     local pr = layer_conf.pr
     if pr == nil then
-        pr = nerv.ParamRepo(nil, self.loc_type)
+        pr = nerv.ParamRepo({}, self.loc_type)
     end
     self.lrepo:rebind(pr)
 end
diff --git a/nerv/layer/rnn.lua b/nerv/layer/rnn.lua
new file mode 100644
index 0000000..e59cf5b
--- /dev/null
+++ b/nerv/layer/rnn.lua
@@ -0,0 +1,38 @@
+local RNNLayer = nerv.class('nerv.RNNLayer', 'nerv.GraphLayer')
+
+function RNNLayer:__init(id, global_conf, layer_conf)
+    nerv.Layer.__init(self, id, global_conf, layer_conf)
+    self:check_dim_len(1, 1)
+
+    local din = layer_conf.dim_in[1]
+    local dout = layer_conf.dim_out[1]
+
+    local pr = layer_conf.pr
+    if pr == nil then
+        pr = nerv.ParamRepo({}, self.loc_type)
+    end
+
+    local layers = {
+        ['nerv.AffineLayer'] = {
+            main = {dim_in = {din, dout}, dim_out = {dout}, pr = pr},
+        },
+        ['nerv.SigmoidLayer'] = {
+            sigmoid = {dim_in = {dout}, dim_out = {dout}},
+        },
+        ['nerv.DuplicateLayer'] = {
+            dup = {dim_in = {dout}, dim_out = {dout, dout}},
+        }
+    }
+
+    local connections = {
+        {'<input>[1]', 'main[1]', 0},
+        {'main[1]', 'sigmoid[1]', 0},
+        {'sigmoid[1]', 'dup[1]', 0},
+        {'dup[1]', 'main[2]', 1},
+        {'dup[2]', '<output>[1]', 0},
+    }
+
+    self:add_prefix(layers, connections)
+    local layer_repo = nerv.LayerRepo(layers, pr, global_conf)
+    self:graph_init(layer_repo, connections)
+end
diff --git a/nerv/nn/init.lua b/nerv/nn/init.lua
index cbaf52b..c32ea09 100644
--- a/nerv/nn/init.lua
+++ b/nerv/nn/init.lua
@@ -1,3 +1,4 @@
 nerv.include('layer_repo.lua')
 nerv.include('param_repo.lua')
 nerv.include('layer_dag.lua')
+nerv.include('network.lua')
diff --git a/nerv/nn/network.lua b/nerv/nn/network.lua
new file mode 100644
index 0000000..35e11e3
--- /dev/null
+++ b/nerv/nn/network.lua
@@ -0,0 +1,498 @@
+local network = nerv.class('nerv.Network')
+
+function network:__init(id, global_conf, network_conf)
+    self.id = id
+    self.network = network_conf.network
+    self.dim_in = self.network.dim_in
+    self.dim_out = self.network.dim_out
+    self.gconf = global_conf
+    if self.gconf.use_cpu then
+        self.mat_type = self.gconf.mmat_type
+    else
+        self.mat_type = self.gconf.cumat_type
+    end
+    self.clip = network_conf.clip
+    self.nn_act_default = network_conf.nn_act_default
+    if self.nn_act_default == nil then
+        self.nn_act_default = 0
+    end
+    self.layers = {}
+    self.input_conn = {}
+    self.output_conn = {}
+    self.socket = self:compile(self.network)
+    for i = 1, #self.dim_in do
+        local edge = self.socket.inputs[i]
+        local id, port, time = edge[1], edge[2], edge[3]
+        if self.input_conn[id][port] ~= nil then
+            nerv.error('duplicate edge')
+        end
+        self.input_conn[id][port] = {0, i, time}
+    end
+    for i = 1, #self.dim_out do
+        local edge = self.socket.outputs[i]
+        local id, port, time = edge[1], edge[2], edge[3]
+        if self.output_conn[id][port] ~= nil then
+            nerv.error('duplicate edge')
+        end
+        self.output_conn[id][port] = {0, i, time}
+    end
+    self.delay = 0
+    for i = 1, #self.layers do
+        local dim_in, _ = self.layers[i]:get_dim()
+        for j = 1, #dim_in do
+            local time = self.input_conn[i][j][3]
+            if math.abs(time) > self.delay then
+                self.delay = math.abs(time)
+            end
+        end
+    end
+end
+
+function network:compile(layer)
+    local socket = {inputs = {}, outputs = {}}
+    if not nerv.is_type(layer, 'nerv.GraphLayer') then
+        table.insert(self.layers, layer)
+        local id = #self.layers
+        self.input_conn[id] = {}
+        self.output_conn[id] = {}
+        local dim_in, dim_out = layer:get_dim()
+        for i = 1, #dim_in do
+            socket.inputs[i] = {id, i, 0}
+        end
+        for i = 1, #dim_out do
+            socket.outputs[i] = {id, i, 0}
+        end
+    else
+        local sublayer_socket = {}
+        for id, sublayer in pairs(layer.layers) do
+            if id ~= '<input>' then
+               sublayer_socket[sublayer.id] = self:compile(sublayer.layer)
+            end
+        end
+        for _, edge in pairs(layer.connections) do
+            -- id = 0 means <input> or <output>
+            local id_from, port_from = edge[1], edge[2]
+            local id_to, port_to = edge[3], edge[4]
+            local time = edge[5]
+            if id_from == 0 then
+                if socket.inputs[port_from] ~= nil then
+                    nerv.error('duplicate input socket')
+                end
+                local input = sublayer_socket[id_to].inputs[port_to]
+                local id, port, t = input[1], input[2], input[3] + time
+                socket.inputs[port_from] = {id, port, t}
+            else
+                local output = sublayer_socket[id_from].outputs[port_from]
+                local id, port, t = output[1], output[2], output[3] + time
+                if id_to == 0 then
+                    if socket.outputs[port_to] ~= nil then
+                        nerv.error('duplicate output socket')
+                    end
+                    socket.outputs[port_to] = {id, port, t}
+                else
+                    local input = sublayer_socket[id_to].inputs[port_to]
+                    local id1, port1, t1 = input[1], input[2], input[3]
+                    if self.input_conn[id1][port1] ~= nil or self.output_conn[id][port] ~= nil then
+                        nerv.error('duplicate edge')
+                    end
+                    self.input_conn[id1][port1] = {id, port, t + t1}
+                    self.output_conn[id][port] = {id1, port1, t + t1}
+                end
+            end
+        end
+    end
+    return socket
+end
+
+function network:init(batch_size, chunk_size)
+    self.batch_size = batch_size
+    self.chunk_size = chunk_size
+
+    self:topsort()
+    
+    self:make_initial_store()
+    collectgarbage('collect')
+
+    for i = 1, #self.layers do
+        self.layers[i]:init(batch_size, chunk_size)
+    end
+end
+
+function network:topsort()
+    nerv.info('network topology sort')
+    local degree = {}
+    for t = 1, self.chunk_size do
+        degree[t] = {}
+        for i = 1, #self.layers do  
+            degree[t][i] = 0
+        end
+    end
+
+    for t = 1, self.chunk_size do
+        for i = 1, #self.layers do
+            local _, dim_out = self.layers[i]:get_dim()
+            for j = 1, #dim_out do
+                if self.output_conn[i][j] ~= nil then
+                    local edge = self.output_conn[i][j]
+                    local id, time = edge[1], edge[3] + t
+                    if time >= 1 and time <= self.chunk_size and id ~= 0 then
+                        degree[time][id] = degree[time][id] + 1
+                    end
+                end
+            end
+        end
+    end
+
+    self.queue = {}
+    local l = 1
+    local r = 0
+    for t = 1, self.chunk_size do
+        for i = 1, #self.layers do
+            if degree[t][i] == 0 then
+                r = r + 1
+                self.queue[r] = {chunk = t, id = i}
+            end
+        end
+    end
+    while l<=r do
+        local t, i = self.queue[l].chunk, self.queue[l].id
+        l = l + 1
+        local _, dim_out = self.layers[i]:get_dim()
+        for j = 1, #dim_out do
+            if self.output_conn[i][j] ~= nil then
+                local edge = self.output_conn[i][j]
+                local id, time = edge[1], edge[3] + t
+                if time >= 1 and time <= self.chunk_size and id ~= 0 then
+                    degree[time][id] = degree[time][id] - 1
+                    if degree[time][id] == 0 then
+                        r = r + 1
+                        self.queue[r] = {chunk = time, id = id}
+                    end
+                end
+            end
+        end
+    end
+
+    if r ~= self.chunk_size * #self.layers then
+        nerv.error('loop detected')
+    end
+end
+
+function network:make_initial_store()
+    nerv.info('network initing storage')
+
+    -- allocate memory
+    local memory = {}
+    local err_memory = {}
+    for t = 1 - self.delay, self.chunk_size + self.delay do
+        memory[t] = {}
+        err_memory[t] = {}
+        for i = 1, #self.layers do
+            memory[t][i] = {}
+            err_memory[t][i] = {}
+            local dim_in, dim_out = self.layers[i]:get_dim()
+            for j = 1, #dim_in do
+                err_memory[t][i][j] = self.mat_type(self.batch_size, dim_in[j])
+                err_memory[t][i][j]:fill(0)
+            end
+            for j = 1, #dim_out do
+                memory[t][i][j] = self.mat_type(self.batch_size, dim_out[j])
+                memory[t][i][j]:fill(self.nn_act_default)
+            end
+        end
+        -- memory[t][0] stores network input
+        memory[t][0] = {}
+        for j = 1, #self.dim_in do
+            memory[t][0][j] = self.mat_type(self.batch_size, self.dim_in[j])
+            memory[t][0][j]:fill(self.nn_act_default)
+        end
+        -- err_memory[t][0] stores network err_input
+        err_memory[t][0] = {}
+        for j = 1, #self.dim_out do
+            err_memory[t][0][j] = self.mat_type(self.batch_size, self.dim_out[j])
+            err_memory[t][0][j]:fill(0)
+        end
+    end
+
+    -- connect memory and reference 
+    self.input = {}
+    self.output = {}
+    self.err_input = {}
+    self.err_output = {}
+    for t = 1, self.chunk_size do
+        self.input[t] = {}
+        self.output[t] = {}
+        self.err_input[t] = {}
+        self.err_output[t] = {}
+        for i = 1, #self.layers do
+            self.input[t][i] = {}
+            self.output[t][i] = {}
+            self.err_input[t][i] = {}
+            self.err_output[t][i] = {}
+            local dim_in, dim_out = self.layers[i]:get_dim()
+            for j = 1, #dim_in do
+                local edge = self.input_conn[i][j]
+                local id, port, time = edge[1], edge[2], edge[3]
+                if id ~= 0 or t - time < 1 or t - time > self.chunk_size then
+                    self.input[t][i][j] = memory[t - time][id][port]
+                end
+                if id ~= 0 then
+                    self.err_output[t][i][j] = err_memory[t][i][j]
+                end
+            end
+            for j = 1, #dim_out do
+                local edge = self.output_conn[i][j]
+                local id, port, time = edge[1], edge[2], edge[3]
+                if id ~= 0 then
+                    self.output[t][i][j] = memory[t][i][j]
+                end
+                if id ~= 0 or t + time < 1 or t + time > self.chunk_size then
+                    self.err_input[t][i][j] = err_memory[t + time][id][port]
+                end
+            end
+        end
+    end
+
+    -- check dangling reference
+    for t = 1, self.chunk_size do
+        for i = 1, #self.dim_in do
+            local edge = self.socket.inputs[i]
+            local id, port, time = edge[1], edge[2], edge[3]
+            if t + time >= 1 and t + time <= self.chunk_size then
+                if self.input[t + time][id][port] ~= nil then
+                    nerv.error('input reference not nil')
+                end
+                self.input[t + time][id][port] = true      -- just a place holder
+                if self.err_output[t + time][id][port] ~= nil then
+                    nerv.error('err_output reference not nil')
+                end
+                self.err_output[t + time][id][port] = true -- just a place holder
+            end
+        end
+        for i = 1, #self.dim_out do
+            local edge = self.socket.outputs[i]
+            local id, port, time = edge[1], edge[2], edge[3]
+            if t - time >= 1 and t - time <= self.chunk_size then
+                if self.output[t - time][id][port] ~= nil then
+                    nerv.error('output reference not nil')
+                end
+                self.output[t - time][id][port] = true     -- just a place holder
+                if self.err_input[t - time][id][port] ~= nil then
+                    nerv.error('err_output reference not nil')
+                end
+                self.err_input[t - time][id][port] = true  -- just a place holder
+            end
+        end
+    end
+    for t = 1, self.chunk_size do
+        for i = 1, #self.layers do
+            local dim_in, dim_out = self.layers[i]:get_dim()
+            for j = 1, #dim_in do
+                if self.input[t][i][j] == nil then
+                    nerv.error('input reference dangling')
+                end
+                if self.err_output[t][i][j] == nil then
+                    nerv.error('err_output reference dangling')
+                end
+            end
+            for j = 1, #dim_out do
+                if self.output[t][i][j] == nil then
+                    nerv.error('output reference dangling')
+                end
+                if self.err_input[t][i][j] == nil then
+                    nerv.error('err_input reference dangling')
+                end
+            end
+        end
+    end
+
+    -- allocate reference for legacy of previous mini-batch
+    self.legacy = {}
+    for t = 1 - self.delay, 0 do
+        self.legacy[t] = {}
+        for i = 1, #self.layers do
+            self.legacy[t][i] = {}
+            local _, dim_out = self.layers[i]:get_dim()
+            for j = 1, #dim_out do
+                self.legacy[t][i][j] = memory[t][i][j]
+            end
+        end
+    end
+end
+
+function network:set_input(input)
+    for t = 1, self.chunk_size do
+        for i = 1, #self.dim_in do
+            local edge = self.socket.inputs[i]
+            local id, port, time = edge[1], edge[2], edge[3]
+            if t + time >= 1 and t + time <= self.chunk_size then
+                self.input[t + time][id][port] = input[t][i]
+            end
+        end
+    end
+end
+
+function network:set_output(output)
+    for t = 1, self.chunk_size do
+        for i = 1, #self.dim_out do
+            local edge = self.socket.outputs[i]
+            local id, port, time = edge[1], edge[2], edge[3]
+            if t - time >= 1 and t - time <= self.chunk_size then
+                self.output[t - time][id][port] = output[t][i]
+            end
+        end
+    end
+end
+
+function network:set_err_input(err_input)
+    for t = 1, self.chunk_size do
+        for i = 1, #self.dim_out do
+            local edge = self.socket.outputs[i]
+            local id, port, time = edge[1], edge[2], edge[3]
+            if t - time >= 1 and t - time <= self.chunk_size then
+                self.err_input[t - time][id][port] = err_input[t][i]
+            end
+        end
+    end
+end
+
+function network:set_err_output(err_output)
+    for t = 1, self.chunk_size do
+        for i = 1, #self.dim_in do
+            local edge = self.socket.inputs[i]
+            local id, port, time = edge[1], edge[2], edge[3]
+            if t + time >= 1 and t + time <= self.chunk_size then
+                self.err_output[t + time][id][port] = err_output[t][i]
+            end
+        end
+    end
+end
+
+--[[
+    [info] is a table that contains information of current mini-batch. These fields must be contained:
+        [input], [output] : matrix array which stores the network input and output
+        [seq_length] : a table contains the length of every sequences
+        [new_seq]: a table contains the batch number of new sequences
+        [do_train]: a bool value indicates do train or not
+        if [do_train] is true, these fileds also must be contained:
+            [err_input], [err_output] : matrix array which stores the network err_input and err_output
+--]]
+function network:mini_batch_init(info)
+    self.info = info
+    self:set_input(self.info.input)
+    self:set_output(self.info.output)
+
+    -- calculate border
+    self.max_length = 0
+    self.border = {}
+    for i = 1, self.chunk_size do
+        self.border[i] = {}
+    end
+    for i = 1, self.batch_size do
+        if self.info.seq_length[i] > self.max_length then
+            self.max_length = self.info.seq_length[i]
+        end
+        for t = 1, self.delay do
+            local chunk = self.info.seq_length[i] + t
+            if chunk > self.chunk_size then
+                break
+            end
+            table.insert(self.border[chunk], i)
+        end
+    end
+
+    -- copy legacy
+    for t = 1 - self.delay, 0 do
+        for i = 1, #self.layers do
+            local _, dim_out = self.layers[i]:get_dim()
+            for j = 1, #dim_out do
+                if t + self.chunk_size >= 1 and self.output_conn[i][j][1] ~= 0 then
+                    self.legacy[t][i][j]:copy_from(self.output[t + self.chunk_size][i][j])
+                end
+                for k = 1, #self.info.new_seq do
+                    local batch = self.info.new_seq[k]
+                    self.legacy[t][i][j][batch - 1]:fill(self.nn_act_default)
+                end
+            end
+        end
+    end
+
+    if self.info.do_train then
+        self:set_err_input(self.info.err_input)
+        self:set_err_output(self.info.err_output)
+        
+        -- flush border gradient
+        for t = self.max_length + 1, self.max_length + self.delay do
+            if t > self.chunk_size then
+                break
+            end
+            for i = 1, #self.layers do
+                local dim_in, _ = self.layers[i]:get_dim()
+                for j = 1, #dim_in do
+                    self.err_output[t][i][j]:fill(0)
+                end
+            end
+        end
+    end
+end
+
+function network:propagate()
+    for i = 1, #self.queue do
+        local t, id = self.queue[i].chunk, self.queue[i].id
+        if t <= self.max_length then
+            self.layers[id]:propagate(self.input[t][id], self.output[t][id], t)
+        end
+        -- flush border activation
+        for j = 1, #self.border[t] do
+            local batch = self.border[t][j]
+            local _, dim_out = self.layers[id]:get_dim()
+            for k = 1, #dim_out do
+                self.output[t][id][k][batch - 1]:fill(self.nn_act_default)
+            end
+        end
+    end
+end
+
+function network:back_propagate()
+    for i = #self.queue, 1, -1 do
+        local t, id = self.queue[i].chunk, self.queue[i].id
+        if t <= self.max_length then
+            -- flush border gradient
+            for j = 1, #self.border[t] do
+                local batch = self.border[t][j]
+                local _, dim_out = self.layers[id]:get_dim()
+                for k = 1, #dim_out do
+                    self.err_input[t][id][k][batch - 1]:fill(0)
+                end
+            end
+            self.layers[id]:back_propagate(self.err_input[t][id], self.err_output[t][id], self.input[t][id], self.output[t][id], t)
+            if self.clip ~= nil then
+                local dim_in, _ = self.layers[id]:get_dim()
+                for j = 1, #dim_in do
+                    self.err_output[t][id][j]:clip(-self.clip, self.clip)
+                end
+            end
+        end
+    end
+end
+
+function network:update()
+    for i = 1, #self.queue do
+        local t, id = self.queue[i].chunk, self.queue[i].id
+        if t <= self.max_length then
+            self.layers[id]:update(self.err_input[t][id], self.input[t][id], self.output[t][id], t)
+        end
+    end
+end
+
+function network:set_attr(name, value)
+    self.network:set_attr(name, value)
+end
+
+function network:get_sublayer(id)
+    return self.network:get_sublayer(id)
+end
+
+function network:get_params()
+    return self.network:get_params()
+end