fixed direct update, did not know the result

4 files changed, 35 insertions, 22 deletions

diff --git a/nerv/examples/lmptb/lm_trainer.lua b/nerv/examples/lmptb/lm_trainer.lua
index 226873b..2be97c8 100644
--- a/nerv/examples/lmptb/lm_trainer.lua
+++ b/nerv/examples/lmptb/lm_trainer.lua
@@ -16,6 +16,7 @@ function LMTrainer.lm_process_file(global_conf, fn, tnn, do_train)
     local result = nerv.LMResult(global_conf, global_conf.vocab)
     result:init("rnn")
     
+    global_conf.timer:flush()
     tnn:flush_all() --caution: will also flush the inputs from the reader!
 
     local next_log_wcn = global_conf.log_w_num
diff --git a/nerv/examples/lmptb/tnn_ptb_main.lua b/nerv/examples/lmptb/tnn_ptb_main.lua
index 891487c..19d0f8a 100644
--- a/nerv/examples/lmptb/tnn_ptb_main.lua
+++ b/nerv/examples/lmptb/tnn_ptb_main.lua
@@ -63,9 +63,11 @@ end
 --Returns: nerv.LayerRepo
 function prepare_layers(global_conf, paramRepo)
     printf("%s preparing layers...\n", global_conf.sche_log_pre)
+    
+    local du = true
 
     --local recurrentLconfig = {{["bp"] = "bp_h", ["ltp_hh"] = "ltp_hh"}, {["dim_in"] = {global_conf.hidden_size, global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}, ["break_id"] = global_conf.vocab:get_sen_entry().id, ["independent"] = global_conf.independent, ["clip"] = 10}}
-    local recurrentLconfig = {{["bp"] = "bp_h", ["ltp_hh"] = "ltp_hh"}, {["dim_in"] = {global_conf.hidden_size, global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}, ["clip"] = 10}}
+    local recurrentLconfig = {{["bp"] = "bp_h", ["ltp_hh"] = "ltp_hh"}, {["dim_in"] = {global_conf.hidden_size, global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}, ["clip"] = 10, ["direct_update"] = du}}
 
     local layers = {
         ["nerv.AffineRecurrentLayer"] = {
@@ -85,7 +87,7 @@ function prepare_layers(global_conf, paramRepo)
         },
 
         ["nerv.AffineLayer"] = {
-            ["outputL"] = {{["ltp"] = "ltp_ho", ["bp"] = "bp_o"}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.vocab:size()}}},
+            ["outputL"] = {{["ltp"] = "ltp_ho", ["bp"] = "bp_o"}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.vocab:size()}, ["direct_update"] = du}},
         },
 
         ["nerv.SoftmaxCELayerT"] = {
@@ -168,7 +170,7 @@ global_conf = {
     mmat_type = nerv.MMatrixFloat,
     nn_act_default = 0, 
 
-    hidden_size = 300, --set to 400 for a stable good test PPL
+    hidden_size = 400, --set to 400 for a stable good test PPL
     chunk_size = 15,
     batch_size = 10, 
     max_iter = 35,
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index 0462383..a2809bf 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -72,18 +72,25 @@ function AffineLayer:batch_resize(batch_size)
 end
 
 function AffineLayer:update(bp_err, input, output)
-    if self.direct_update then
-        self.ltp.correction:mul(input[1], bp_err[1], 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.ltp.trans:add(self.ltp.trans, self.ltp.correction, 1.0, -gconf.lrate / n)
+    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')
+            -- 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 / n)
+            self.bp.trans:add(self.bp.trans, bp_err[1]:colsum(), 1.0-gconf.lrate*gconf.wcost, -gconf.lrate / gconf.batch_size)
+        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_grad:mul(input[1], bp_err[1], 1.0, 0.0, 'T', 'N')
         self.ltp:update(self.ltp_grad)
+        self.bp:update(bp_err[1]:colsum())
     end
-    self.bp:update(bp_err[1]:colsum())
 end
 
 function AffineLayer:propagate(input, output)
diff --git a/nerv/layer/affine_recurrent.lua b/nerv/layer/affine_recurrent.lua
index 92d98e2..3d448d3 100644
--- a/nerv/layer/affine_recurrent.lua
+++ b/nerv/layer/affine_recurrent.lua
@@ -46,17 +46,20 @@ function Recurrent:update(bp_err, input, output)
         local ltp_hh = self.ltp_hh.trans
         local bp = self.bp.trans
         local gconf = self.gconf
-        -- momentum gain
-        local mmt_gain = 1.0 / (1.0 - gconf.momentum);
-        local n = input[1]:nrow() * mmt_gain
-        -- update corrections (accumulated errors)
-        self.ltp_hh.correction:mul(input[2], bp_err[1], 1.0, gconf.momentum, 'T', 'N')
-        self.bp.correction:add(bc, bp_err[1]:colsum(), gconf.momentum, 1.0)
-        -- perform update
-        ltp_hh:add(ltp_hh, self.ltp_hh.correction, 1.0, -gconf.lrate / n)
-        bp:add(bp, self.bp.correction, 1.0, -gconf.lrate / n)
-        -- weight decay
-        ltp_hh:add(ltp_hh, ltp_hh, 1.0, -gconf.lrate * gconf.wcost)
+        if (gconf.momentum > 0) then
+            -- momentum gain
+            local mmt_gain = 1.0 / (1.0 - gconf.momentum);
+            local n = input[1]:nrow() * mmt_gain
+            -- update corrections (accumulated errors)
+            self.ltp_hh.correction:mul(input[2], bp_err[1], 1.0, gconf.momentum, 'T', 'N')
+            self.bp.correction:add(self.bp.correction, bp_err[1]:colsum(), gconf.momentum, 1.0)
+            -- perform update and weight decay
+            ltp_hh:add(ltp_hh, self.ltp_hh.correction, 1.0 - gconf.lrate.gconf.wcost/n, -gconf.lrate / n)
+            bp:add(bp, self.bp.correction, 1.0 - gconf.lrate*gconf.wcost/n, -gconf.lrate / n)
+        else
+            ltp_hh:mul(input[2], bp_err[1], -gconf.lrate/gconf.batch_size, 1.0-gconf.wcost*gconf.lrate/gconf.batch_size, 'T', 'N')
+            bp:add(bp, bp_err[1]:colsum(), 1.0-gconf.lrate*gconf.wcost/gconf.batch_size, -gconf.lrate/gconf.batch_size)
+        end
     else
         self.ltp_hh_grad:mul(input[2], bp_err[1], 1.0, 0.0, 'T', 'N') 
         self.ltp_hh:update(self.ltp_hh_grad)