merge in recent changes about param updates

Merge branch 'master' into txh18/rnnlm

1 files changed, 26 insertions, 46 deletions

diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index 3e84ec0..ed58d38 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -17,49 +17,46 @@ function MatrixParam:train_init()
     self.correction:fill(0)
 end
 
-function MatrixParam:update_by_gradient(gradient)
+function MatrixParam:_update_by_gradient(gradient, alpha, beta)
     local gconf = self.gconf
+    -- momentum gain
+    local mmt_gain = 1.0 / (1.0 - gconf.momentum)
+    local n = self.gconf.batch_size * mmt_gain
+    -- perform update
     if gconf.momentum > 0 then
         self.correction:add(self.correction, gradient, gconf.momentum, 1.0)
-        -- momentum gain
-        local mmt_gain = 1.0 / (1.0 - gconf.momentum)
-        local n = self.gconf.batch_size * mmt_gain
-        -- perform update
-        self.trans:add(self.trans, self.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n)
+        self.trans:add(self.trans, self.correction, alpha, -gconf.lrate / n * beta)
     else
-        self.trans:add(self.trans, gradient, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / gconf.batch_size)
+        self.trans:add(self.trans, gradient, alpha, -gconf.lrate / n * beta)
     end
 end
 
-function MatrixParam:update_by_err_input(err, input)
+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 = self.gconf.batch_size * mmt_gain
+    -- perform update
     if gconf.momentum > 0 then
         self.correction:mul(input, err, 1.0, gconf.momentum, 'T', 'N')
-        -- momentum gain
-        local mmt_gain = 1.0 / (1.0 - gconf.momentum)
-        local n = self.gconf.batch_size * mmt_gain
-        -- perform update
-        self.trans:add(self.trans, self.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n)
+        self.trans:add(self.trans, self.correction, alpha, -gconf.lrate / n * beta)
     else
-        self.trans:mul(input, err, - gconf.lrate / gconf.batch_size, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, 'T', 'N')
+        self.trans:mul(input, err, -gconf.lrate / n * beta, alpha, 'T', 'N')
     end
 end
 
---[[ --these updates are the same
-function LinearTransParam:update(gradient)
-    MatrixParam.update(self, gradient)
-    -- local gconf = self.gconf
-    -- weight decay(put into MatrixParam:update)
-    -- self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost / gconf.batch_size)
+function MatrixParam:update_by_gradient(gradient)
+    self:_update_by_gradient(gradient, 1.0, 1.0)
+end
+
+function MatrixParam:update_by_err_input(err, input)
+    self:_update_by_err_input(err, input, 1.0, 1.0)
 end
 
-function BiasParam:update(gradient)
-    MatrixParam.update(self, gradient)
-    --local gconf = self.gconf
-    -- weight decay
-    -- self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost / gconf.batch_size)
+function LinearTransParam:update_by_err_input(err, input)
+    local l2 = 1 - self.gconf.lrate * self.gconf.wcost
+    self:_update_by_err_input(err, input, l2, l2)
 end
-]]--
 
 function AffineLayer:__init(id, global_conf, layer_conf)
     self.id = id
@@ -69,7 +66,7 @@ function AffineLayer:__init(id, global_conf, layer_conf)
     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.direct_update = layer_conf.direct_update or global_conf.direct_update
 end
 
 function AffineLayer:init(batch_size)
@@ -92,25 +89,8 @@ function AffineLayer:batch_resize(batch_size)
 end
 
 function AffineLayer:update(bp_err, input, output)
-    if self.direct_update == true then
-        local gconf = self.gconf
-        if gconf.momentum > 0 then
-            self.ltp.correction:mul(input[1], bp_err[1], 1.0, gconf.momentum, 'T', 'N')
-            self.bp.correction:add(self.bp.correction, bp_err[1]:colsum(), gconf.momentum, 1)
-            -- momentum gain
-            local mmt_gain = 1.0 / (1.0 - gconf.momentum)
-            local n = self.gconf.batch_size * mmt_gain
-            -- perform update
-            self.ltp.trans:add(self.ltp.trans, self.ltp.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n)
-            self.bp.trans:add(self.bp.trans, self.bp.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n)
-        else
-            self.ltp.trans:mul(input[1], bp_err[1], - gconf.lrate / gconf.batch_size, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, 'T', 'N')
-            self.bp.trans:add(self.bp.trans, bp_err[1]:colsum(), 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / gconf.batch_size)
-        end
-    else
-        self.ltp:update_by_err_input(bp_err[1], input[1])
-        self.bp:update_by_gradient(bp_err[1]:colsum())
-    end
+    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)