1 files changed, 19 insertions, 24 deletions

diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index 38743aa..a05ae17 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -25,7 +25,9 @@ end
 
 function MatrixParam:train_init()
     self.correction = self.trans:create()
+    self.correction_acc = self.correction:create()
     self.correction:fill(0)
+    self.correction_acc:fill(0)
 end
 
 function MatrixParam:copy(copier)
@@ -34,46 +36,37 @@ function MatrixParam:copy(copier)
     return target
 end
 
-function MatrixParam:_update_by_gradient(gradient, alpha, beta)
+function MatrixParam:_update(alpha, beta)
     local gconf = self.gconf
     -- momentum gain
     local mmt_gain = 1.0 / (1.0 - gconf.momentum)
     local n = gconf.batch_size * mmt_gain
     -- perform update
     if gconf.momentum > 0 then
-        self.correction:add(self.correction, gradient, gconf.momentum, 1.0)
+        self.correction:add(self.correction, self.correction_acc, gconf.momentum, 1.0)
         self.trans:add(self.trans, self.correction, alpha, -gconf.lrate / n * beta)
     else
-        self.trans:add(self.trans, gradient, alpha, -gconf.lrate / n * beta)
+        self.trans:add(self.trans, self.correction_acc, alpha, -gconf.lrate / n * beta)
     end
+    self.correction_acc:fill(0)
 end
 
-function MatrixParam:_update_by_err_input(err, input, alpha, beta)
-    local gconf = self.gconf
-    -- momentum gain
-    local mmt_gain = 1.0 / (1.0 - gconf.momentum)
-    local n = gconf.batch_size * mmt_gain
-    -- perform update
-    if gconf.momentum > 0 then
-        self.correction:mul(input, err, 1.0, gconf.momentum, 'T', 'N')
-        self.trans:add(self.trans, self.correction, alpha, -gconf.lrate / n * beta)
-    else
-        self.trans:mul(input, err, -gconf.lrate / n * beta, alpha, 'T', 'N')
-    end
+function MatrixParam:back_propagate_by_gradient(gradient)
+    self.correction_acc:add(self.correction_acc, gradient, 1.0, 1.0)
 end
 
-function MatrixParam:update_by_gradient(gradient)
-    self:_update_by_gradient(gradient, 1.0, 1.0)
+function MatrixParam:back_propagate_by_err_input(err, input)
+    self.correction_acc:mul(input, err, 1.0, 1.0, 'T', 'N')
 end
 
-function MatrixParam:update_by_err_input(err, input)
-    self:_update_by_err_input(err, input, 1.0, 1.0)
+function MatrixParam:update_by_gradient()
+    self:_update(1.0, 1.0)
 end
 
-function LinearTransParam:update_by_err_input(err, input)
+function MatrixParam:update_by_err_input()
     local gconf = self.gconf
     local l2 = 1 - gconf.lrate * gconf.wcost
-    self:_update_by_err_input(err, input, l2, l2)
+    self:_update(l2, l2)
 end
 
 --- A fully-connected linear transform layer.
@@ -121,11 +114,11 @@ function AffineLayer:batch_resize(batch_size)
     -- do nothing
 end
 
-function AffineLayer:update(bp_err, input, output)
+function AffineLayer:update()
     for i = 1, #self.dim_in do
-        self["ltp" .. i]:update_by_err_input(bp_err[1], input[i])
+        self["ltp" .. i]:update_by_err_input()
     end
-    self.bp:update_by_gradient(bp_err[1]:colsum())
+    self.bp:update_by_gradient()
 end
 
 function AffineLayer:propagate(input, output)
@@ -141,7 +134,9 @@ end
 function AffineLayer: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
+    self.bp:back_propagate_by_gradient(bp_err[1]:colsum())
 end
 
 function AffineLayer:get_params()