let affine supported multiple inputs

4 files changed, 123 insertions, 8 deletions

diff --git a/nerv/examples/lmptb/tnn/init.lua b/nerv/examples/lmptb/tnn/init.lua
index a069527..a7a377e 100644
--- a/nerv/examples/lmptb/tnn/init.lua
+++ b/nerv/examples/lmptb/tnn/init.lua
@@ -43,5 +43,6 @@ end
 
 nerv.include('tnn.lua')
 nerv.include('layersT/softmax_ce_t.lua')
+nerv.include('layers/elem_mul.lua')
 nerv.include('layers/gate_fff.lua')
 nerv.include('layer_dag_t.lua')
diff --git a/nerv/examples/lmptb/tnn/layers/elem_mul.lua b/nerv/examples/lmptb/tnn/layers/elem_mul.lua
new file mode 100644
index 0000000..c809d3e
--- /dev/null
+++ b/nerv/examples/lmptb/tnn/layers/elem_mul.lua
@@ -0,0 +1,38 @@
+local ElemMulLayer = nerv.class('nerv.ElemMulLayer', 'nerv.Layer')
+
+function ElemMulLayer:__init(id, global_conf, layer_conf)
+    self.id = id
+    self.dim_in = layer_conf.dim_in
+    self.dim_out = layer_conf.dim_out
+    self.gconf = global_conf
+ 
+    self:check_dim_len(2, 1) -- Element-multiply input[1] and input[2]
+end
+
+function ElemMulLayer:init(batch_size)
+    if self.dim_in[1] ~= self.dim_in[2] or 
+        self.dim_in[1] ~= self.dim_out[1] then
+        nerv.error("dim_in and dim_out mismatch for ElemMulLayer")
+    end
+end
+
+function ElemMulLayer:batch_resize(batch_size)
+    --do nothing
+end
+
+function ElemMulLayer:propagate(input, output)
+    output[1]:mul_elem(input[1], input[2])
+end
+
+function ElemMulLayer:back_propagate(bp_err, next_bp_err, input, output)
+    next_bp_err[1]:mul_elem(bp_err[1], input[2]) 
+    next_bp_err[2]:mul_elem(bp_err[1], input[1])
+end
+
+function ElemMulLayer:update(bp_err, input, output)
+    --do nothing
+end
+
+function ElemMulLayer:get_params()
+    return nerv.ParamRepo({})
+end
diff --git a/nerv/examples/lmptb/tnn/layersT/lstm.lua b/nerv/examples/lmptb/tnn/layersT/lstm.lua
new file mode 100644
index 0000000..0da1f38
--- /dev/null
+++ b/nerv/examples/lmptb/tnn/layersT/lstm.lua
@@ -0,0 +1,56 @@
+local LSTMLayerT = nerv.class('nerv.LSTMLayerT', 'nerv.LayerT')
+
+function LSTMLayerT:__init(id, global_conf, layer_conf)
+    self.id = id
+    self.dim_in = layer_conf.dim_in
+    self.dim_out = layer_conf.dim_out
+    self.gconf = global_conf
+
+    --prepare a DAGLayerT to hold the lstm structure
+    local paramRepo = nerv.ParamRepo()
+    local layers = {
+        ["nerv.IndRecurrentLayer"] = {
+            ["recurrentL1"] = recurrentLconfig, 
+        }}
+
+    self:check_dim_len(1, 1) -- exactly one input and one output
+end
+
+function LSTMLayerT:init(batch_size)
+    if self.ltp.trans:ncol() ~= self.bp.trans:ncol() then
+        nerv.error("mismatching dimensions of linear transform and bias paramter")
+    end
+    if self.dim_in[1] ~= self.ltp.trans:nrow() then
+        nerv.error("mismatching dimensions of linear transform parameter and input")
+    end
+    if self.dim_out[1] ~= self.ltp.trans:ncol() then
+        nerv.error("mismatching dimensions of linear transform parameter and output")
+    end
+    self.ltp_grad = self.ltp.trans:create()
+    self.ltp:train_init()
+    self.bp:train_init()
+end
+
+function LSTMLayerT:batch_resize(batch_size)
+    -- do nothing
+end
+
+function AffineLayer:update(bp_err, input, output)
+    self.ltp:update_by_err_input(bp_err[1], input[1])
+    self.bp:update_by_gradient(bp_err[1]:colsum())
+end
+
+function AffineLayer:propagate(input, output)
+    -- apply linear transform
+    output[1]:mul(input[1], self.ltp.trans, 1.0, 0.0, 'N', 'N')
+    -- add bias
+    output[1]:add_row(self.bp.trans, 1.0)
+end
+
+function AffineLayer:back_propagate(bp_err, next_bp_err, input, output)
+    next_bp_err[1]:mul(bp_err[1], self.ltp.trans, 1.0, 0.0, 'N', 'T')
+end
+
+function AffineLayer:get_params()
+    return nerv.ParamRepo({self.ltp, self.bp})
+end
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index e24a0c6..d56fcb8 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -64,25 +64,35 @@ function AffineLayer:__init(id, global_conf, layer_conf)
     self.dim_in = layer_conf.dim_in
     self.dim_out = layer_conf.dim_out
     self.ltp = self:find_param("ltp", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[1], self.dim_out[1]}) --layer_conf.ltp
-    self.bp = self:find_param("bp", layer_conf, global_conf, nerv.BiasParam, {1, self.dim_out[1]})--layer_conf.bp
+    for i = 2, #self.dim_in do
+        self["ltp" .. i] = self:find_param("ltp" .. i, layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[i], self.dim_out[1]}) 
+    end
+    self.bp = self:find_param("bp", layer_conf, global_conf, nerv.BiasParam, {1, self.dim_out[1]}) --layer_conf.bp
     self.gconf = global_conf
-    self:check_dim_len(1, 1) -- exactly one input and one output
-    -- self.direct_update = layer_conf.direct_update or global_conf.direct_update
+    self:check_dim_len(-1, 1) -- exactly one output, allow multiple inputs
 end
 
 function AffineLayer:init(batch_size)
     if self.ltp.trans:ncol() ~= self.bp.trans:ncol() then
         nerv.error("mismatching dimensions of linear transform and bias paramter")
     end
+    self.bp:train_init()
     if self.dim_in[1] ~= self.ltp.trans:nrow() then
         nerv.error("mismatching dimensions of linear transform parameter and input")
     end
     if self.dim_out[1] ~= self.ltp.trans:ncol() then
         nerv.error("mismatching dimensions of linear transform parameter and output")
     end
-    self.ltp_grad = self.ltp.trans:create()
     self.ltp:train_init()
-    self.bp:train_init()
+    for i = 2, #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 AffineLayer:batch_resize(batch_size)
@@ -91,20 +101,30 @@ end
 
 function AffineLayer:update(bp_err, input, output)
     self.ltp:update_by_err_input(bp_err[1], input[1])
+    for i = 2, #self.dim_in do
+        self["ltp" .. i]:update_by_err_input(bp_err[1], input[i])
+    end
     self.bp:update_by_gradient(bp_err[1]:colsum())
 end
 
 function AffineLayer:propagate(input, output)
-    -- apply linear transform
     output[1]:mul(input[1], self.ltp.trans, 1.0, 0.0, 'N', 'N')
-    -- add bias
+    for i = 2, #self.dim_in do
+        output[1]:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N')
+    end
     output[1]:add_row(self.bp.trans, 1.0)
 end
 
 function AffineLayer:back_propagate(bp_err, next_bp_err, input, output)
     next_bp_err[1]:mul(bp_err[1], self.ltp.trans, 1.0, 0.0, 'N', 'T')
+    for i = 2, #self.dim_in do
+        next_bp_err[i]:mul(bp_err[1], self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T')
+    end
 end
 
 function AffineLayer:get_params()
-    return nerv.ParamRepo({self.ltp, self.bp})
+    local pr = nerv.ParamRepo({self.ltp, self.bp})
+    for i = 2, #self.dim_in do
+        pr:add(self["ltp" .. i].id, self["ltp" .. i])
+    end
 end