From 89a3fa93d571f446bcd1fa69ddd35257d975c239 Mon Sep 17 00:00:00 2001
From: Qi Liu <liuq901@163.com>
Date: Mon, 9 May 2016 17:57:30 +0800
Subject: merge affine & lstm_gate & projection layer

---
 nerv/Makefile              |  4 +-
 nerv/examples/ptb/main.lua |  2 +-
 nerv/layer/affine.lua      | 67 ++++++++++++++++++++++++--------
 nerv/layer/init.lua        |  2 -
 nerv/layer/lstm.lua        |  8 ++--
 nerv/layer/lstm_gate.lua   | 97 ----------------------------------------------
 nerv/layer/lstmp.lua       | 12 ++----
 nerv/layer/projection.lua  | 70 ---------------------------------
 8 files changed, 61 insertions(+), 201 deletions(-)
 delete mode 100644 nerv/layer/lstm_gate.lua
 delete mode 100644 nerv/layer/projection.lua

diff --git a/nerv/Makefile b/nerv/Makefile
index 0d9934a..a3d0f34 100644
--- a/nerv/Makefile
+++ b/nerv/Makefile
@@ -40,9 +40,9 @@ 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/tanh.lua layer/softmax_ce.lua layer/softmax.lua \
-			layer/lstmp.lua layer/projection.lua layer/relu.lua\
+			layer/lstmp.lua layer/relu.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 \
+			layer/elem_mul.lua layer/lstm.lua layer/dropout.lua layer/gru.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/network.lua nn/trainer.lua\
 			io/frm_buffer.lua io/seq_buffer.lua
diff --git a/nerv/examples/ptb/main.lua b/nerv/examples/ptb/main.lua
index 5d1a326..7191da0 100644
--- a/nerv/examples/ptb/main.lua
+++ b/nerv/examples/ptb/main.lua
@@ -105,7 +105,7 @@ function trainer:epoch_preprocess(dataset, do_train)
     total_frame = 0
 end
 
-function trainer:mini_batch_middleprocess(cnt, info)
+function trainer:mini_batch_inprocess(cnt, info)
     for t = 1, gconf.chunk_size do
         local tmp = info.output[1][t]:new_to_host()
         for i = 1, gconf.batch_size do
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index 16250fd..8b4751c 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -87,7 +87,11 @@ local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer')
 -- @param global_conf see `self.gconf` of `nerv.Layer.__init`
 -- @param layer_conf a table providing with settings dedicated for the layer,
 -- for `layer_conf` fields that are shared by all layers, see
--- `nerv.Layer.__init`. The affine layer requires parameters to be bound, the
+-- `nerv.Layer.__init`. This fields can be specified:
+-- * `activation`: the type of the activation function layer, also known as \sigma in \sigma(Wx + b). Default value none (no activation function).
+-- * `no_bias`: a bool value indicates use bias parameter or not. Default value false.
+-- * `param_type`: a string table has the same length with `dim_in`, indicates the parameter type for every input. 'D' for diagonal weight matrix, 'N' for normal weight matrix. Default 'N' for every input.
+-- The affine layer requires parameters to be bound, the
 -- following parameter names will be looked up while binding:
 --
 -- * `ltp`: the linear transformation parameter, also known as the weight matrix, W in Wx + b
@@ -95,7 +99,11 @@ local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer')
 
 function AffineLayer:__init(id, global_conf, layer_conf)
     nerv.Layer.__init(self, id, global_conf, layer_conf)
+    self.param_type = layer_conf.param_type or table.vector(#self.dim_in, 'N')
     self:check_dim_len(-1, 1) -- exactly one output, allow multiple inputs
+    if layer_conf.activation then
+        self.activation = layer_conf.activation('', global_conf, {dim_in = {self.dim_out[1]}, dim_out = {self.dim_out[1]}})
+    end
     self:bind_params()
 end
 
@@ -108,19 +116,24 @@ function AffineLayer:bind_params()
         self["ltp" .. i] = self:find_param(pid_list, lconf, self.gconf,
                                             nerv.LinearTransParam,
                                             {self.dim_in[i], self.dim_out[1]})
+        if self.param_type[i] == 'D' then
+            self['ltp' .. i].trans:diagonalize()
+        end
         local no_update = lconf["no_update_ltp" .. i]
         if (no_update ~= nil) and no_update or lconf.no_update_all then
             self["ltp" .. i].no_update = true
         end
     end
     self.ltp = self.ltp1 -- alias of ltp1
-    self.bp = self:find_param("bp", lconf, self.gconf,
-                                nerv.BiasParam,
-                                {1, self.dim_out[1]},
-                                nerv.Param.gen_zero)
-    local no_update = lconf["no_update_bp"]
-    if (no_update ~= nil) and no_update or lconf.no_update_all then
-        self.bp.no_update = true
+    if not self.no_bias then
+       self.bp = self:find_param("bp", lconf, self.gconf,
+                                    nerv.BiasParam,
+                                    {1, self.dim_out[1]},
+                                    nerv.Param.gen_zero)
+        local no_update = lconf["no_update_bp"]
+        if (no_update ~= nil) and no_update or lconf.no_update_all then
+            self.bp.no_update = true
+        end
     end
 end
 
@@ -137,7 +150,15 @@ function AffineLayer:init(batch_size)
         end
         self["ltp" .. i]:train_init()
     end
-    self.bp:train_init()
+    if not self.no_bias then
+        self.bp:train_init()
+    end
+    if self.activation then
+        self.act_bak = self.mat_type(batch_size, self.dim_out[1])
+        self.act_bak:fill(0)
+        self.err_bak = self.mat_type(batch_size, self.dim_out[1])
+        self.err_bak:fill(0)
+    end
 end
 
 function AffineLayer:batch_resize(batch_size)
@@ -148,25 +169,39 @@ function AffineLayer:update()
     for i = 1, #self.dim_in do
         self["ltp" .. i]:update_by_err_input()
     end
-    self.bp:update_by_gradient()
+    if not self.no_bias then
+        self.bp:update_by_gradient()
+    end
 end
 
 function AffineLayer:propagate(input, output)
+    local result = self.activation and self.act_bak or output[1]
     -- apply linear transform
-    output[1]:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N')
+    result:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N')
     for i = 2, #self.dim_in do
-        output[1]:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N')
+        result:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N')
     end
     -- add bias
-    output[1]:add_row(self.bp.trans, 1.0)
+    if not self.no_bias then
+        result:add_row(self.bp.trans, 1.0)
+    end
+    if self.activation then
+        self.activation:propagate({self.act_bak}, output)
+    end
 end
 
 function AffineLayer:back_propagate(bp_err, next_bp_err, input, output)
+    if self.activation then
+        self.activation:back_propagate(bp_err, {self.err_bak}, {self.act_bak}, output)
+    end
+    local result = self.activation and self.err_bak or bp_err[1]
     for i = 1, #self.dim_in do
-        next_bp_err[i]:mul(bp_err[1], self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T')
-        self["ltp" .. i]:back_propagate_by_err_input(bp_err[1], input[i])
+        next_bp_err[i]:mul(result, self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T')
+        self["ltp" .. i]:back_propagate_by_err_input(result, input[i])
+    end
+    if not self.no_bias then
+        self.bp:back_propagate_by_gradient(result:colsum())
     end
-    self.bp:back_propagate_by_gradient(bp_err[1]:colsum())
 end
 
 function AffineLayer:get_params()
diff --git a/nerv/layer/init.lua b/nerv/layer/init.lua
index d175d02..054784b 100644
--- a/nerv/layer/init.lua
+++ b/nerv/layer/init.lua
@@ -272,13 +272,11 @@ nerv.include('combiner.lua')
 nerv.include('softmax.lua')
 nerv.include('elem_mul.lua')
 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')
-nerv.include('projection.lua')
 nerv.include('lstmp.lua')
 nerv.include('relu.lua')
 
diff --git a/nerv/layer/lstm.lua b/nerv/layer/lstm.lua
index 3de3453..e568ee8 100644
--- a/nerv/layer/lstm.lua
+++ b/nerv/layer/lstm.lua
@@ -23,16 +23,14 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
         },
         ['nerv.AffineLayer'] = {
             mainAffine = {dim_in = {din, dout}, dim_out = {dout}, pr = pr},
+            forgetGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
+            inputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
+            outputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
         },
         ['nerv.TanhLayer'] = {
             mainTanh = {dim_in = {dout}, dim_out = {dout}},
             outputTanh = {dim_in = {dout}, dim_out = {dout}},
         },
-        ['nerv.LSTMGateLayer'] = {
-            forgetGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr},
-            inputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr},
-            outputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr},
-        },
         ['nerv.ElemMulLayer'] = {
             inputGateMul = {dim_in = {dout, dout}, dim_out = {dout}},
             forgetGateMul = {dim_in = {dout, dout}, dim_out = {dout}},
diff --git a/nerv/layer/lstm_gate.lua b/nerv/layer/lstm_gate.lua
deleted file mode 100644
index 39a3ff7..0000000
--- a/nerv/layer/lstm_gate.lua
+++ /dev/null
@@ -1,97 +0,0 @@
-local LSTMGateLayer = nerv.class('nerv.LSTMGateLayer', 'nerv.Layer')
--- NOTE: this is a full matrix gate
-
-function LSTMGateLayer:__init(id, global_conf, layer_conf)
-    nerv.Layer.__init(self, id, global_conf, layer_conf)
-    self.param_type = layer_conf.param_type
-    self:check_dim_len(-1, 1) --accept multiple inputs
-    self:bind_params()
-end
-
-function LSTMGateLayer:bind_params()
-    local lconf = self.lconf
-    lconf.no_update_ltp1 = lconf.no_update_ltp1 or lconf.no_update_ltp
-    for i = 1, #self.dim_in do
-        local pid = "ltp" .. i
-        local pid_list = i == 1 and {pid, "ltp"} or pid
-        self["ltp" .. i] = self:find_param(pid_list, lconf, self.gconf,
-                                            nerv.LinearTransParam,
-                                            {self.dim_in[i], self.dim_out[1]})
-        if self.param_type[i] == 'D' then
-            self["ltp" .. i].trans:diagonalize()
-        end
-        local no_update = lconf["no_update_ltp" .. i]
-        if (no_update ~= nil) and no_update or lconf.no_update_all then
-            self["ltp" .. i].no_update = true
-        end
-    end
-    self.ltp = self.ltp1 -- alias of ltp1
-    self.bp = self:find_param("bp", lconf, self.gconf,
-                                nerv.BiasParam, {1, self.dim_out[1]},
-                                nerv.Param.gen_zero)
-    local no_update = lconf["no_update_bp"]
-    if (no_update ~= nil) and no_update or lconf.no_update_all then
-        self.bp.no_update = true
-    end
-end
-
-function LSTMGateLayer:init(batch_size)
-    if self.dim_out[1] ~= self.bp.trans:ncol() then
-        nerv.error("mismatching dimensions of linear transform and bias paramter")
-    end
-    for i = 1, #self.dim_in do
-        if self.dim_in[i] ~= self["ltp" .. i].trans:nrow() then
-            nerv.error("mismatching dimensions of linear transform parameter and input")
-        end
-        if self.dim_out[1] ~= self["ltp" .. i].trans:ncol() then
-            nerv.error("mismatching dimensions of linear transform parameter and output")
-        end
-        self["ltp" .. i]:train_init()
-    end
-    self.bp:train_init()
-    self.err_bakm = self.mat_type(batch_size, self.dim_out[1])
-end
-
-function LSTMGateLayer:batch_resize(batch_size)
-    if self.err_m:nrow() ~= batch_size then
-        self.err_bakm = self.mat_type(batch_size, self.dim_out[1])
-    end
-end
-
-function LSTMGateLayer:propagate(input, output)
-    -- apply linear transform
-    output[1]:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N')
-    for i = 2, #self.dim_in do
-        output[1]:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N')
-    end
-    -- add bias
-    output[1]:add_row(self.bp.trans, 1.0)
-    output[1]:sigmoid(output[1])
-end
-
-function LSTMGateLayer:back_propagate(bp_err, next_bp_err, input, output)
-    self.err_bakm:sigmoid_grad(bp_err[1], output[1])
-    for i = 1, #self.dim_in do
-        next_bp_err[i]:mul(self.err_bakm, self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T')
-        self["ltp" .. i]:back_propagate_by_err_input(self.err_bakm, input[i])
-    end
-    self.bp:back_propagate_by_gradient(self.err_bakm:colsum())
-end
-
-function LSTMGateLayer:update()
-    for i = 1, #self.dim_in do
-        self["ltp" .. i]:update_by_err_input()
-        if self.param_type[i] == 'D' then
-            self["ltp" .. i].trans:diagonalize()
-        end
-    end
-    self.bp:update_by_gradient()
-end
-
-function LSTMGateLayer:get_params()
-    local pr = nerv.ParamRepo({self.bp}, self.loc_type)
-    for i = 1, #self.dim_in do
-        pr:add(self["ltp" .. i])
-    end
-    return pr
-end
diff --git a/nerv/layer/lstmp.lua b/nerv/layer/lstmp.lua
index bbb2091..dc30797 100644
--- a/nerv/layer/lstmp.lua
+++ b/nerv/layer/lstmp.lua
@@ -24,24 +24,20 @@ function LSTMPLayer:__init(id, global_conf, layer_conf)
         },
         ['nerv.AffineLayer'] = {
             mainAffine = {dim_in = {din, dout}, dim_out = {dcell}, pr = pr},
+            forgetGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
+            inputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
+            outputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
+            projection = {dim_in = {dcell}, dim_out = {dout}, pr = pr, no_bias = true},
         },
         ['nerv.TanhLayer'] = {
             mainTanh = {dim_in = {dcell}, dim_out = {dcell}},
             outputTanh = {dim_in = {dcell}, dim_out = {dcell}},
         },
-        ['nerv.LSTMGateLayer'] = {
-            forgetGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr},
-            inputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr},
-            outputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr},
-        },
         ['nerv.ElemMulLayer'] = {
             inputGateMul = {dim_in = {dcell, dcell}, dim_out = {dcell}},
             forgetGateMul = {dim_in = {dcell, dcell}, dim_out = {dcell}},
             outputGateMul = {dim_in = {dcell, dcell}, dim_out = {dcell}},
         },
-        ['nerv.ProjectionLayer'] = {
-            projection = {dim_in = {dcell}, dim_out = {dout}, pr = pr},
-        },
     }
     
     local connections = {
diff --git a/nerv/layer/projection.lua b/nerv/layer/projection.lua
deleted file mode 100644
index 077125b..0000000
--- a/nerv/layer/projection.lua
+++ /dev/null
@@ -1,70 +0,0 @@
-local ProjectionLayer = nerv.class('nerv.ProjectionLayer', 'nerv.Layer')
-
---- The constructor.
-function ProjectionLayer:__init(id, global_conf, layer_conf)
-    nerv.Layer.__init(self, id, global_conf, layer_conf)
-    self:check_dim_len(-1, 1) -- exactly one output, allow multiple inputs
-    self:bind_params()
-end
-
-function ProjectionLayer:bind_params()
-    local lconf = self.lconf
-    lconf.no_update_ltp1 = lconf.no_update_ltp1 or lconf.no_update_ltp
-    for i = 1, #self.dim_in do
-        local pid = "ltp" .. i
-        local pid_list = i == 1 and {pid, "ltp"} or pid
-        self["ltp" .. i] = self:find_param(pid_list, lconf, self.gconf,
-                                            nerv.LinearTransParam,
-                                            {self.dim_in[i], self.dim_out[1]})
-        local no_update = lconf["no_update_ltp" .. i]
-        if (no_update ~= nil) and no_update or lconf.no_update_all then
-            self["ltp" .. i].no_update = true
-        end
-    end
-    self.ltp = self.ltp1 -- alias of ltp1
-end
-
-function ProjectionLayer:init(batch_size)
-    for i = 1, #self.dim_in do
-        if self.dim_in[i] ~= self["ltp" .. i].trans:nrow() then
-            nerv.error("mismatching dimensions of linear transform parameter and input")
-        end
-        if self.dim_out[1] ~= self["ltp" .. i].trans:ncol() then
-            nerv.error("mismatching dimensions of linear transform parameter and output")
-        end
-        self["ltp" .. i]:train_init()
-    end
-end
-
-function ProjectionLayer:batch_resize(batch_size)
-    -- do nothing
-end
-
-function ProjectionLayer:update()
-    for i = 1, #self.dim_in do
-        self["ltp" .. i]:update_by_err_input()
-    end
-end
-
-function ProjectionLayer:propagate(input, output)
-    -- apply linear transform
-    output[1]:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N')
-    for i = 2, #self.dim_in do
-        output[1]:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N')
-    end
-end
-
-function ProjectionLayer:back_propagate(bp_err, next_bp_err, input, output)
-    for i = 1, #self.dim_in do
-        next_bp_err[i]:mul(bp_err[1], self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T')
-        self["ltp" .. i]:back_propagate_by_err_input(bp_err[1], input[i])
-    end
-end
-
-function ProjectionLayer:get_params()
-    local pr = nerv.ParamRepo({self.ltp1}, self.loc_type)
-    for i = 2, #self.dim_in do
-        pr:add(self["ltp" .. i])
-    end
-    return pr
-end
-- 
cgit v1.2.3-70-g09d2


From 03439902dbd339cfbbc684b6fcc6b1810fa02ede Mon Sep 17 00:00:00 2001
From: Qi Liu <liuq901@163.com>
Date: Mon, 9 May 2016 20:51:10 +0800
Subject: fix bug in affine.lua

---
 nerv/layer/affine.lua | 24 +++++++++++++-----------
 nerv/nn/network.lua   |  1 -
 2 files changed, 13 insertions(+), 12 deletions(-)

diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index 8b4751c..1ac4681 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -88,7 +88,7 @@ local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer')
 -- @param layer_conf a table providing with settings dedicated for the layer,
 -- for `layer_conf` fields that are shared by all layers, see
 -- `nerv.Layer.__init`. This fields can be specified:
--- * `activation`: the type of the activation function layer, also known as \sigma in \sigma(Wx + b). Default value none (no activation function).
+-- * `activation`: the type of the activation function layer, also known as \sigma in \sigma(Wx + b). The activation function layer must gurantee not use parameter `input` in its `back_propagate` function. Default value none (no activation function).
 -- * `no_bias`: a bool value indicates use bias parameter or not. Default value false.
 -- * `param_type`: a string table has the same length with `dim_in`, indicates the parameter type for every input. 'D' for diagonal weight matrix, 'N' for normal weight matrix. Default 'N' for every input.
 -- The affine layer requires parameters to be bound, the
@@ -99,11 +99,12 @@ local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer')
 
 function AffineLayer:__init(id, global_conf, layer_conf)
     nerv.Layer.__init(self, id, global_conf, layer_conf)
-    self.param_type = layer_conf.param_type or table.vector(#self.dim_in, 'N')
     self:check_dim_len(-1, 1) -- exactly one output, allow multiple inputs
+    self.param_type = layer_conf.param_type or table.vector(#self.dim_in, 'N')
     if layer_conf.activation then
         self.activation = layer_conf.activation('', global_conf, {dim_in = {self.dim_out[1]}, dim_out = {self.dim_out[1]}})
     end
+    self.no_bias = layer_conf.no_bias
     self:bind_params()
 end
 
@@ -138,7 +139,7 @@ function AffineLayer:bind_params()
 end
 
 function AffineLayer:init(batch_size)
-    if self.dim_out[1] ~= self.bp.trans:ncol() then
+    if not self.no_bias and self.dim_out[1] ~= self.bp.trans:ncol() then
         nerv.error("mismatching dimensions of linear transform and bias paramter")
     end
     for i = 1, #self.dim_in do
@@ -154,10 +155,8 @@ function AffineLayer:init(batch_size)
         self.bp:train_init()
     end
     if self.activation then
-        self.act_bak = self.mat_type(batch_size, self.dim_out[1])
-        self.act_bak:fill(0)
-        self.err_bak = self.mat_type(batch_size, self.dim_out[1])
-        self.err_bak:fill(0)
+        self.bak_mat = self.mat_type(batch_size, self.dim_out[1])
+        self.bak_mat:fill(0)
     end
 end
 
@@ -168,6 +167,9 @@ end
 function AffineLayer:update()
     for i = 1, #self.dim_in do
         self["ltp" .. i]:update_by_err_input()
+        if self.param_type[i] == 'D' then
+            self['ltp' .. i].trans:diagonalize()
+        end
     end
     if not self.no_bias then
         self.bp:update_by_gradient()
@@ -175,7 +177,7 @@ function AffineLayer:update()
 end
 
 function AffineLayer:propagate(input, output)
-    local result = self.activation and self.act_bak or output[1]
+    local result = self.activation and self.bak_mat or output[1]
     -- apply linear transform
     result:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N')
     for i = 2, #self.dim_in do
@@ -186,15 +188,15 @@ function AffineLayer:propagate(input, output)
         result:add_row(self.bp.trans, 1.0)
     end
     if self.activation then
-        self.activation:propagate({self.act_bak}, output)
+        self.activation:propagate({result}, output)
     end
 end
 
 function AffineLayer:back_propagate(bp_err, next_bp_err, input, output)
+    local result = self.activation and self.bak_mat or bp_err[1]
     if self.activation then
-        self.activation:back_propagate(bp_err, {self.err_bak}, {self.act_bak}, output)
+        self.activation:back_propagate(bp_err, {result}, {result}, output)
     end
-    local result = self.activation and self.err_bak or bp_err[1]
     for i = 1, #self.dim_in do
         next_bp_err[i]:mul(result, self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T')
         self["ltp" .. i]:back_propagate_by_err_input(result, input[i])
diff --git a/nerv/nn/network.lua b/nerv/nn/network.lua
index bf69ccc..d0d5462 100644
--- a/nerv/nn/network.lua
+++ b/nerv/nn/network.lua
@@ -416,7 +416,6 @@ function network:make_initial_store()
             local dim_in, dim_out = self.layers[i]:get_dim()
             for j = 1, #dim_in do
                 if self.input[t][i][j] == nil then
-                    print(t,i,j,self.layers[i].id)
                     nerv.error('input reference dangling')
                 end
                 if self.err_output[t][i][j] == nil then
-- 
cgit v1.2.3-70-g09d2


From 4585970021f75d4c9e7154fc1681a80efa0f48ab Mon Sep 17 00:00:00 2001
From: Qi Liu <liuq901@163.com>
Date: Mon, 9 May 2016 21:02:57 +0800
Subject: clip on gradient now

---
 nerv/examples/trainer.lua |  1 +
 nerv/layer/affine.lua     |  4 ++++
 nerv/nn/network.lua       | 10 ----------
 nerv/nn/trainer.lua       |  2 +-
 4 files changed, 6 insertions(+), 11 deletions(-)

diff --git a/nerv/examples/trainer.lua b/nerv/examples/trainer.lua
index f6c7a5a..caed2e2 100644
--- a/nerv/examples/trainer.lua
+++ b/nerv/examples/trainer.lua
@@ -80,6 +80,7 @@ local trainer_defaults = {
 
 local options = make_options(trainer_defaults)
 local extra_opt_spec = {
+    {"clip", nil, "number"},
     {"resume-from", nil, "string"},
     {"help", "h", "boolean", default = false, desc = "show this help information"},
     {"dir", nil, "string", desc = "specify the working directory"},
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index 1ac4681..b68cf3d 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -48,6 +48,10 @@ function MatrixParam:_update(alpha, beta)
     -- momentum gain
     local mmt_gain = 1.0 / (1.0 - gconf.momentum)
     local n = gconf.batch_size * mmt_gain
+    -- clip gradient
+    if gconf.clip then
+        self.correction_acc:clip(-gconf.clip, gconf.clip)
+    end
     -- perform update
     if gconf.momentum > 0 then
         self.correction:add(self.correction, self.correction_acc, gconf.momentum, 1.0)
diff --git a/nerv/nn/network.lua b/nerv/nn/network.lua
index d0d5462..358b100 100644
--- a/nerv/nn/network.lua
+++ b/nerv/nn/network.lua
@@ -33,8 +33,6 @@ local network = nerv.class('nerv.Network')
 --
 -- * `network`: a `nerv.Layer` instance describing the structure of the network
 --   to be compiled
--- * `clip`: a `number` value indicating the cliping threshold (i.e. preserve
---   the values within [-clip, +clip])
 -- * `nn_act_default`: a `number` value indicating the value used for filling
 --   "holes" in activation values of a batch matrix (0 by default)
 
@@ -49,7 +47,6 @@ function network:__init(id, global_conf, network_conf)
     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
@@ -660,13 +657,6 @@ function network:back_propagate()
         local t, id = self.queue[i].chunk, self.queue[i].id
         if t <= self.max_length then
             self.layers[id]:back_propagate(self.err_input[t][id], self.err_output[t][id], self.input[t][id], self.output[t][id], t)
-            -- gradient clip
-            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
         -- flush border gradient
         if self.flush[t][id].timestamp == self.timestamp then
diff --git a/nerv/nn/trainer.lua b/nerv/nn/trainer.lua
index 44390ea..8357c10 100644
--- a/nerv/nn/trainer.lua
+++ b/nerv/nn/trainer.lua
@@ -25,7 +25,7 @@ function trainer:__init(gconf)
     self.input_order = self:get_input_order()
     self.network = nerv.Network('network', gconf,
                                 {network = graph,
-                                 clip = gconf.clip})
+                                 nn_act_default = gconf.nn_act_default})
     local network = self.network
     network:init(gconf.batch_size, gconf.chunk_size)
 
-- 
cgit v1.2.3-70-g09d2


From e85a4af65640aef77378275d478c1ba8b06b785e Mon Sep 17 00:00:00 2001
From: Qi Liu <liuq901@163.com>
Date: Wed, 11 May 2016 16:53:56 +0800
Subject: seq buffer support shuffle

---
 nerv/io/seq_buffer.lua | 91 +++++++++++++++++++++++++++++++++++++++++---------
 nerv/nn/trainer.lua    |  3 ++
 2 files changed, 79 insertions(+), 15 deletions(-)

diff --git a/nerv/io/seq_buffer.lua b/nerv/io/seq_buffer.lua
index 8cde1b3..b1c2a75 100644
--- a/nerv/io/seq_buffer.lua
+++ b/nerv/io/seq_buffer.lua
@@ -28,6 +28,8 @@ local SeqBuffer = nerv.class('nerv.SeqBuffer', 'nerv.DataBuffer')
 -- * `batch_size`: the number of rows for each batch matrix
 -- * `chunk_size`: the length of the BPTT context (number of batch
 --   matrices to provide upon each invocation of `get_data()`)
+-- * `buffer_size`: the number of frames to be buffered and shuffled at once (shuffle
+--   in the sequence level, not in the frame level)
 -- * `nn_act_default`: the default value to fill into the "holes" (non-data
 --   frames)
 
@@ -36,6 +38,8 @@ function SeqBuffer:__init(global_conf, buffer_conf)
 
     self.batch_size = buffer_conf.batch_size
     self.chunk_size = buffer_conf.chunk_size
+    self.buffer_size = buffer_conf.buffer_size
+    self.randomize = buffer_conf.randomize
     self.readers = {}
     for _, v in ipairs(buffer_conf.readers) do
         table.insert(self.readers, v.reader)
@@ -49,6 +53,11 @@ function SeqBuffer:__init(global_conf, buffer_conf)
     self.queue = {}
     self.head = 1
     self.tail = 0
+    self.offset = 1
+    self.buffer = {}
+    self.length = {}
+    self.index = {}
+    self.complete = false
 end
 
 function SeqBuffer:new_mini_batch()
@@ -62,25 +71,77 @@ function SeqBuffer:new_mini_batch()
     return res
 end
 
+local function random_shuffle(a)
+    for i = #a, 2, -1 do
+        local j = math.random(i)
+        a[i], a[j] = a[j], a[i]
+    end
+end
+
+function SeqBuffer:fill_buffer()
+    if self.complete then
+        return false
+    end
+    local t = os.clock()
+    self.buffer = {}
+    self.length = {}
+    local size = 0
+    while size < self.buffer_size do
+        local drow = nil
+        local data = {}
+        for i = 1, #self.readers do
+            local tmp = self.readers[i]:get_data()
+            if tmp == nil then
+                self.complete = true
+                break
+            end
+            for id, d in pairs(tmp) do
+                if drow == nil then
+                    drow = d:nrow()
+                elseif d:nrow() ~= drow then
+                    nerv.error('readers provides with inconsistent rows of data')
+                end
+                data[id] = d
+            end
+        end
+        if self.complete then
+            break
+        end
+        size = size + drow
+        table.insert(self.buffer, data)
+        table.insert(self.length, drow)
+    end
+    self.index = {}
+    for i = 1, #self.buffer do
+        self.index[i] = i
+    end
+    if self.randomize then
+        random_shuffle(self.index)
+    end
+    self.offset = 1
+    collectgarbage('collect')
+    nerv.info('%.3fs to fill the buffer', os.clock() - t)
+    return #self.buffer > 0
+end
+
+function SeqBuffer:get_buffered_data()
+    if self.offset > #self.buffer then
+        if not self:fill_buffer() then
+            return nil
+        end
+    end
+    local id = self.index[self.offset]
+    self.offset = self.offset + 1
+    return self.buffer[id], self.length[id]
+end
+
 function SeqBuffer:saturate(batch)
     if self.queue[self.head] ~= nil and self.queue[self.head].seq_length[batch] ~= 0 then
         return true
     end
-    local data = {}
-    local drow = nil
-    for i = 1, #self.readers do
-        local tmp = self.readers[i]:get_data()
-        if tmp == nil then
-            return false
-        end
-        for id, d in pairs(tmp) do
-            if drow == nil then
-                drow = d:nrow()
-            elseif d:nrow() ~= drow then
-                nerv.error('readers provides with inconsistent rows of data')
-            end
-            data[id] = d
-        end
+    local data, drow = self:get_buffered_data()
+    if data == nil then
+        return false
     end
     local offset = 0
     local head = self.head
diff --git a/nerv/nn/trainer.lua b/nerv/nn/trainer.lua
index 8357c10..a17b36c 100644
--- a/nerv/nn/trainer.lua
+++ b/nerv/nn/trainer.lua
@@ -77,9 +77,12 @@ function trainer:make_buffer(readers)
         })
     else
         return nerv.SeqBuffer(gconf, {
+            buffer_size = gconf.buffer_size,
             batch_size = gconf.batch_size,
             chunk_size = gconf.chunk_size,
+            randomize = gconf.randomize,
             readers = readers,
+            nn_act_default = gconf.nn_act_default,
         })
     end
 end
-- 
cgit v1.2.3-70-g09d2


From 313ab19b004603f10c282c5ae035e36e3437ea1f Mon Sep 17 00:00:00 2001
From: Qi Liu <liuq901@163.com>
Date: Wed, 11 May 2016 16:56:37 +0800
Subject: add doc

---
 nerv/io/seq_buffer.lua | 1 +
 1 file changed, 1 insertion(+)

diff --git a/nerv/io/seq_buffer.lua b/nerv/io/seq_buffer.lua
index b1c2a75..5c60f64 100644
--- a/nerv/io/seq_buffer.lua
+++ b/nerv/io/seq_buffer.lua
@@ -30,6 +30,7 @@ local SeqBuffer = nerv.class('nerv.SeqBuffer', 'nerv.DataBuffer')
 --   matrices to provide upon each invocation of `get_data()`)
 -- * `buffer_size`: the number of frames to be buffered and shuffled at once (shuffle
 --   in the sequence level, not in the frame level)
+-- * `randomize`: shuffle the buffer after filled if true
 -- * `nn_act_default`: the default value to fill into the "holes" (non-data
 --   frames)
 
-- 
cgit v1.2.3-70-g09d2


From 10916f721a945a5edd052ab93027413fd3c01f65 Mon Sep 17 00:00:00 2001
From: Qi Liu <liuq901@163.com>
Date: Thu, 12 May 2016 11:49:44 +0800
Subject: lstm & lstmp layer support multiple input

---
 nerv/layer/lstm.lua  | 53 ++++++++++++++++++++++++++++++++------------------
 nerv/layer/lstmp.lua | 55 +++++++++++++++++++++++++++++++++-------------------
 nerv/layer/rnn.lua   | 15 +++-----------
 3 files changed, 72 insertions(+), 51 deletions(-)

diff --git a/nerv/layer/lstm.lua b/nerv/layer/lstm.lua
index e568ee8..5d73ad2 100644
--- a/nerv/layer/lstm.lua
+++ b/nerv/layer/lstm.lua
@@ -2,9 +2,12 @@ local LSTMLayer = nerv.class('nerv.LSTMLayer', 'nerv.GraphLayer')
 
 function LSTMLayer:__init(id, global_conf, layer_conf)
     nerv.Layer.__init(self, id, global_conf, layer_conf)
-    self:check_dim_len(1, 1)
+    self:check_dim_len(-1, 1)
+    if #self.dim_in == 0 then
+        nerv.error('LSTM layer %s has no input', self.id)
+    end
 
-    local din = layer_conf.dim_in[1]
+    local din = layer_conf.dim_in
     local dout = layer_conf.dim_out[1]
 
     local pr = layer_conf.pr
@@ -17,15 +20,17 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
             mainCombine = {dim_in = {dout, dout}, dim_out = {dout}, lambda = {1, 1}},
         },
         ['nerv.DuplicateLayer'] = {
-            inputDup = {dim_in = {din}, dim_out = {din, din, din, din}},
             outputDup = {dim_in = {dout}, dim_out = {dout, dout, dout, dout, dout}},
             cellDup = {dim_in = {dout}, dim_out = {dout, dout, dout, dout, dout}},
         },
         ['nerv.AffineLayer'] = {
-            mainAffine = {dim_in = {din, dout}, dim_out = {dout}, pr = pr},
-            forgetGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
-            inputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
-            outputGate = {dim_in = {din, dout, dout}, dim_out = {dout}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
+            mainAffine = {dim_in = table.connect({dout}, din), dim_out = {dout}, pr = pr},
+            forgetGate = {dim_in = table.connect({dout, dout}, din), dim_out = {dout},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
+            inputGate = {dim_in = table.connect({dout, dout}, din), dim_out = {dout},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
+            outputGate = {dim_in = table.connect({dout, dout}, din), dim_out = {dout},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
         },
         ['nerv.TanhLayer'] = {
             mainTanh = {dim_in = {dout}, dim_out = {dout}},
@@ -37,26 +42,29 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
             outputGateMul = {dim_in = {dout, dout}, dim_out = {dout}},
         },
     }
+    for i = 1, #din do
+        layers['nerv.DuplicateLayer']['inputDup' .. i] = {dim_in = {din[i]}, dim_out = {din[i], din[i], din[i], din[i]}}
+    end
     
     local connections = {
         -- lstm input
-        {'<input>[1]', 'inputDup[1]', 0},
+        --{'<input>[1 .. n]', 'inputDup(1 .. n)[1]', 0},
 
         -- input gate
-        {'inputDup[1]', 'inputGate[1]', 0},
-        {'outputDup[1]', 'inputGate[2]', 1},
-        {'cellDup[1]', 'inputGate[3]', 1},
+        {'outputDup[1]', 'inputGate[1]', 1},
+        {'cellDup[1]', 'inputGate[2]', 1},
+        --{'inputDup(1 .. n)[1]', 'inputGate[3 .. n + 2]', 0},
 
         -- forget gate
-        {'inputDup[2]', 'forgetGate[1]', 0},
-        {'outputDup[2]', 'forgetGate[2]', 1},
-        {'cellDup[2]', 'forgetGate[3]', 1},
+        {'outputDup[2]', 'forgetGate[1]', 1},
+        {'cellDup[2]', 'forgetGate[2]', 1},
+        --{'inputDup(1 .. n)[2]', 'forgetGate[3 .. n + 2]', 0},
 
         -- lstm cell
         {'forgetGate[1]', 'forgetGateMul[1]', 0},
         {'cellDup[3]', 'forgetGateMul[2]', 1},
-        {'inputDup[3]', 'mainAffine[1]', 0},
-        {'outputDup[3]', 'mainAffine[2]', 1},
+        {'outputDup[3]', 'mainAffine[1]', 1},
+        --{'inputDup(1 .. n)[3]', 'mainAffine[2 .. n + 1]', 0},
         {'mainAffine[1]', 'mainTanh[1]', 0},
         {'inputGate[1]', 'inputGateMul[1]', 0},
         {'mainTanh[1]', 'inputGateMul[2]', 0},
@@ -65,9 +73,9 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
         {'mainCombine[1]', 'cellDup[1]', 0},
 
         -- forget gate
-        {'inputDup[4]', 'outputGate[1]', 0},
-        {'outputDup[4]', 'outputGate[2]', 1},
-        {'cellDup[4]', 'outputGate[3]', 0},
+        {'outputDup[4]', 'outputGate[1]', 1},
+        {'cellDup[4]', 'outputGate[2]', 0},
+        --{'inputDup(1 .. n)[4]', 'outputGate[2 .. n + 1]', 0},
 
         -- lstm output
         {'cellDup[5]', 'outputTanh[1]', 0},
@@ -76,6 +84,13 @@ function LSTMLayer:__init(id, global_conf, layer_conf)
         {'outputGateMul[1]', 'outputDup[1]', 0},
         {'outputDup[5]', '<output>[1]', 0},
     }
+    for i = 1, #din do
+        table.insert(connections, {'<input>[' .. i .. ']', 'inputDup' .. i .. '[1]', 0})
+        table.insert(connections, {'inputDup' .. i .. '[1]', 'inputGate[' .. (i + 2) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[2]', 'forgetGate[' .. (i + 2) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[3]', 'mainAffine[' .. (i + 1) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[4]', 'outputGate[' .. (i + 2) .. ']', 0})
+    end
 
     self:add_prefix(layers, connections)
     local layer_repo = nerv.LayerRepo(layers, pr, global_conf)
diff --git a/nerv/layer/lstmp.lua b/nerv/layer/lstmp.lua
index dc30797..49c9516 100644
--- a/nerv/layer/lstmp.lua
+++ b/nerv/layer/lstmp.lua
@@ -2,9 +2,12 @@ local LSTMPLayer = nerv.class('nerv.LSTMPLayer', 'nerv.GraphLayer')
 
 function LSTMPLayer:__init(id, global_conf, layer_conf)
     nerv.Layer.__init(self, id, global_conf, layer_conf)
-    self:check_dim_len(1, 1)
+    self:check_dim_len(-1, 1)
+    if #self.dim_in == 0 then
+        nerv.error('LSTMP layer %s has no input', self.id)
+    end
 
-    local din = layer_conf.dim_in[1]
+    local din = layer_conf.dim_in
     local dcell = layer_conf.cell_dim
     local dout = layer_conf.dim_out[1]
 
@@ -18,15 +21,17 @@ function LSTMPLayer:__init(id, global_conf, layer_conf)
             mainCombine = {dim_in = {dcell, dcell}, dim_out = {dcell}, lambda = {1, 1}},
         },
         ['nerv.DuplicateLayer'] = {
-            inputDup = {dim_in = {din}, dim_out = {din, din, din, din}},
             outputDup = {dim_in = {dout}, dim_out = {dout, dout, dout, dout, dout}},
             cellDup = {dim_in = {dcell}, dim_out = {dcell, dcell, dcell, dcell, dcell}},
         },
         ['nerv.AffineLayer'] = {
-            mainAffine = {dim_in = {din, dout}, dim_out = {dcell}, pr = pr},
-            forgetGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
-            inputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
-            outputGate = {dim_in = {din, dout, dcell}, dim_out = {dcell}, param_type = {'N', 'N', 'D'}, pr = pr, activation = nerv.SigmoidLayer},
+            mainAffine = {dim_in = table.connect({dout}, din), dim_out = {dcell}, pr = pr},
+            forgetGate = {dim_in = table.connect({dout, dcell}, din), dim_out = {dcell},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
+            inputGate = {dim_in = table.connect({dout, dcell}, din), dim_out = {dcell},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
+            outputGate = {dim_in = table.connect({dout, dcell}, din), dim_out = {dcell},
+                param_type = table.connect({'N', 'D'}, table.vector(#din, 'N')), pr = pr, activation = nerv.SigmoidLayer},
             projection = {dim_in = {dcell}, dim_out = {dout}, pr = pr, no_bias = true},
         },
         ['nerv.TanhLayer'] = {
@@ -39,26 +44,29 @@ function LSTMPLayer:__init(id, global_conf, layer_conf)
             outputGateMul = {dim_in = {dcell, dcell}, dim_out = {dcell}},
         },
     }
-    
+    for i = 1, #din do
+        layers['nerv.DuplicateLayer']['inputDup' .. i] = {dim_in = {din[i]}, dim_out = {din[i], din[i], din[i], din[i]}}
+    end
+
     local connections = {
         -- lstm input
-        {'<input>[1]', 'inputDup[1]', 0},
+        --{'<input>[1 .. n]', 'inputDup(1 .. n)[1]', 0},
 
         -- input gate
-        {'inputDup[1]', 'inputGate[1]', 0},
-        {'outputDup[1]', 'inputGate[2]', 1},
-        {'cellDup[1]', 'inputGate[3]', 1},
+        {'outputDup[1]', 'inputGate[1]', 1},
+        {'cellDup[1]', 'inputGate[2]', 1},
+        --{'inputDup(1 .. n)[1]', 'inputGate[3 .. n + 2]', 0},
 
         -- forget gate
-        {'inputDup[2]', 'forgetGate[1]', 0},
-        {'outputDup[2]', 'forgetGate[2]', 1},
-        {'cellDup[2]', 'forgetGate[3]', 1},
+        {'outputDup[2]', 'forgetGate[1]', 1},
+        {'cellDup[2]', 'forgetGate[2]', 1},
+        --{'inputDup(1 .. n)[2]', 'forgetGate[3 .. n + 2]', 0},
 
         -- lstm cell
         {'forgetGate[1]', 'forgetGateMul[1]', 0},
         {'cellDup[3]', 'forgetGateMul[2]', 1},
-        {'inputDup[3]', 'mainAffine[1]', 0},
-        {'outputDup[3]', 'mainAffine[2]', 1},
+        {'outputDup[3]', 'mainAffine[1]', 1},
+        --{'inputDup(1 .. n)[3]', 'mainAffine[2 .. n + 1]', 0},
         {'mainAffine[1]', 'mainTanh[1]', 0},
         {'inputGate[1]', 'inputGateMul[1]', 0},
         {'mainTanh[1]', 'inputGateMul[2]', 0},
@@ -67,9 +75,9 @@ function LSTMPLayer:__init(id, global_conf, layer_conf)
         {'mainCombine[1]', 'cellDup[1]', 0},
 
         -- forget gate
-        {'inputDup[4]', 'outputGate[1]', 0},
-        {'outputDup[4]', 'outputGate[2]', 1},
-        {'cellDup[4]', 'outputGate[3]', 0},
+        {'outputDup[4]', 'outputGate[1]', 1},
+        {'cellDup[4]', 'outputGate[2]', 0},
+        --{'inputDup(1 .. n)[4]', 'outputGate[2 .. n + 1]', 0},
 
         -- lstm output
         {'cellDup[5]', 'outputTanh[1]', 0},
@@ -79,6 +87,13 @@ function LSTMPLayer:__init(id, global_conf, layer_conf)
         {'projection[1]', 'outputDup[1]', 0},
         {'outputDup[5]', '<output>[1]', 0},
     }
+    for i = 1, #din do
+        table.insert(connections, {'<input>[' .. i .. ']', 'inputDup' .. i .. '[1]', 0})
+        table.insert(connections, {'inputDup' .. i .. '[1]', 'inputGate[' .. (i + 2) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[2]', 'forgetGate[' .. (i + 2) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[3]', 'mainAffine[' .. (i + 1) .. ']', 0})
+        table.insert(connections, {'inputDup' .. i .. '[4]', 'outputGate[' .. (i + 2) .. ']', 0})
+    end
 
     self:add_prefix(layers, connections)
     local layer_repo = nerv.LayerRepo(layers, pr, global_conf)
diff --git a/nerv/layer/rnn.lua b/nerv/layer/rnn.lua
index fd6e753..63e0b55 100644
--- a/nerv/layer/rnn.lua
+++ b/nerv/layer/rnn.lua
@@ -4,12 +4,7 @@ function RNNLayer:__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 == 0 then
-        nerv.error('RNN Layer %s has no input', self.id)
-    end
-
-    self.activation = layer_conf.activation
-    if self.activation == nil then
-        self.activation = 'nerv.SigmoidLayer'
+        nerv.error('RNN layer %s has no input', self.id)
     end
 
     local din = layer_conf.dim_in
@@ -22,10 +17,7 @@ function RNNLayer:__init(id, global_conf, layer_conf)
 
     local layers = {
         ['nerv.AffineLayer'] = {
-            main = {dim_in = table.connect({dout}, din), dim_out = {dout}, pr = pr},
-        },
-        [self.activation] = {
-            activation = {dim_in = {dout}, dim_out = {dout}},
+            main = {dim_in = table.connect({dout}, din), dim_out = {dout}, pr = pr, activation = layer_conf.activation},
         },
         ['nerv.DuplicateLayer'] = {
             duplicate = {dim_in = {dout}, dim_out = {dout, dout}},
@@ -33,8 +25,7 @@ function RNNLayer:__init(id, global_conf, layer_conf)
     }
 
     local connections = {
-        {'main[1]', 'activation[1]', 0},
-        {'activation[1]', 'duplicate[1]', 0},
+        {'main[1]', 'duplicate[1]', 0},
         {'duplicate[1]', 'main[1]', 1},
         {'duplicate[2]', '<output>[1]', 0},
     }
-- 
cgit v1.2.3-70-g09d2