got good result when batch_size=1, strange!

3 files changed, 20 insertions, 6 deletions

diff --git a/nerv/examples/lmptb/m-tests/tnn_test.lua b/nerv/examples/lmptb/m-tests/tnn_test.lua
index 40e332c..a2c38f0 100644
--- a/nerv/examples/lmptb/m-tests/tnn_test.lua
+++ b/nerv/examples/lmptb/m-tests/tnn_test.lua
@@ -155,6 +155,9 @@ function lm_process_file(global_conf, fn, tnn, do_train)
 
     local next_log_wcn = global_conf.log_w_num
 
+    global_conf.fz = 0
+    global_conf.fz2 = 0
+
     while (1) do
         local r, feeds
 
@@ -198,7 +201,7 @@ function lm_process_file(global_conf, fn, tnn, do_train)
             next_log_wcn = next_log_wcn + global_conf.log_w_num
             printf("%s %d words processed %s.\n", global_conf.sche_log_pre, result["rnn"].cn_w, os.date()) 
             printf("\t%s log prob per sample :%f.\n", global_conf.sche_log_pre, result:logp_sample("rnn"))
-            nerv.LMUtil.wait(1)
+            nerv.LMUtil.wait(0.1)
         end
          
         --[[
@@ -213,6 +216,9 @@ function lm_process_file(global_conf, fn, tnn, do_train)
 
         --break --debug
     end
+    
+    print("gconf.fz", global_conf.fz)
+    print("gconf.fz2", global_conf.fz2)
 
     printf("%s Displaying result:\n", global_conf.sche_log_pre)
     printf("%s %s\n", global_conf.sche_log_pre, result:status("rnn"))
@@ -232,14 +238,14 @@ valid_fn = data_dir .. '/ptb.valid.txt.adds'
 test_fn = data_dir .. '/ptb.test.txt.adds'
 
 global_conf = {
-    lrate = 1, wcost = 1e-6, momentum = 0,
+    lrate = 0.1, wcost = 1e-6, momentum = 0,
     cumat_type = nerv.CuMatrixFloat,
     mmat_type = nerv.MMatrixFloat,
     nn_act_default = 0, 
 
     hidden_size = 200,
     chunk_size = 15,
-    batch_size = 10, 
+    batch_size = 1, 
     max_iter = 25,
     param_random = function() return (math.random() / 5 - 0.1) end,
 
diff --git a/nerv/examples/lmptb/rnn/tnn.lua b/nerv/examples/lmptb/rnn/tnn.lua
index 10b501e..fc5321d 100644
--- a/nerv/examples/lmptb/rnn/tnn.lua
+++ b/nerv/examples/lmptb/rnn/tnn.lua
@@ -386,6 +386,7 @@ function TNN:propagate_dfs(ref, t)
     end
     ]]--
     ref.layer:propagate(ref.inputs_m[t], ref.outputs_m[t], t) --propagate!
+
     if (bit.band(self.feeds_now.flagsPack_now[t], bit.bor(nerv.TNN.FC.SEQ_START, nerv.TNN.FC.SEQ_END)) > 0) then --restore cross-border history
         for i = 1, self.batch_size do
             local seq_start = bit.band(self.feeds_now.flags_now[t][i], nerv.TNN.FC.SEQ_START) 
@@ -393,13 +394,13 @@ function TNN:propagate_dfs(ref, t)
             if (seq_start > 0 or seq_end > 0) then
                 for p, conn in pairs(ref.o_conns_p) do
                     if ((ref.o_conns_p[p].time > 0 and seq_end > 0) or (ref.o_conns_p[p].time < 0 and seq_start > 0)) then 
+                        self.gconf.fz2 = self.gconf.fz2 + 1
                         ref.outputs_m[t][p][i - 1]:fill(self.gconf.nn_act_default)                       
                     end
                 end
             end
         end
     end
- 
     --set input flag for future layers
     for i = 1, #ref.dim_out do
         if (ref.outputs_b[t][i] == true) then
@@ -501,13 +502,13 @@ function TNN:backpropagate_dfs(ref, t, do_update)
             if (seq_start > 0 or seq_end > 0) then
                 for p, conn in pairs(ref.i_conns_p) do
                     if ((ref.i_conns_p[p].time > 0 and seq_start > 0) or (ref.i_conns_p[p].time < 0 and seq_end > 0)) then --cross-border, set to zero
+                        self.gconf.fz = self.gconf.fz + 1
                         ref.err_outputs_m[t][p][i - 1]:fill(0)
                     end
                 end
             end
         end
     end
- 
 
     for i = 1, #ref.dim_in do
         if (ref.err_outputs_b[t][i] == true) then
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index 015ec3f..0462383 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -31,7 +31,14 @@ function LinearTransParam: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)
+    self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost / gconf.batch_size)
+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)
 end
 
 function AffineLayer:__init(id, global_conf, layer_conf)