fix bug in passing err_input to network; gen zero vectors for biasalpha-3.3

6 files changed, 27 insertions, 14 deletions

diff --git a/nerv/examples/asr_trainer.lua b/nerv/examples/asr_trainer.lua
index 38ba6e9..9a764fc 100644
--- a/nerv/examples/asr_trainer.lua
+++ b/nerv/examples/asr_trainer.lua
@@ -39,12 +39,8 @@ local function build_trainer(ifname)
         local buffer = make_buffer(make_readers(scp_file, layer_repo))
         -- initialize the network
         gconf.cnt = 0
-        local err_input = {{}}
         local output = {{}}
         for i = 1, gconf.chunk_size do
-            local mini_batch = mat_type(gconf.batch_size, 1)
-            mini_batch:fill(1)
-            table.insert(err_input[1], mini_batch)
             table.insert(output[1], mat_type(gconf.batch_size, 1))
         end
         network:epoch_init()
@@ -91,7 +87,7 @@ local function build_trainer(ifname)
                                     do_train = bp,
                                     input = input,
                                     output = output,
-                                    err_input = err_input,
+                                    err_input = {gconf.mask},
                                     err_output = err_output})
             network:propagate()
             if bp then
@@ -254,8 +250,15 @@ nerv.set_logfile(path.join(working_dir, logfile_name))
 -- start the training
 local trainer = build_trainer(pf0)
 local pr_prev
-gconf.accu_best, pr_prev = trainer(nil, gconf.cv_scp, false)
+-- initial cross-validation
+local param_prefix = string.format("%s_%s",
+                                string.gsub(
+                                    (string.gsub(pf0[1], "(.*/)(.*)", "%2")),
+                                    "(.*)%..*", "%1"),
+                                os.date(date_pattern))
+gconf.accu_best, pr_prev = trainer(path.join(working_dir, param_prefix), gconf.cv_scp, false)
 nerv.info("initial cross validation: %.3f", gconf.accu_best)
+-- main loop
 for i = gconf.cur_iter, gconf.max_iter do
     local stop = false
     gconf.cur_iter = i
@@ -264,7 +267,7 @@ for i = gconf.cur_iter, gconf.max_iter do
         nerv.info("[NN] begin iteration %d with lrate = %.6f", i, gconf.lrate)
         local accu_tr = trainer(nil, gconf.tr_scp, true, rebind_param_repo)
         nerv.info("[TR] training set %d: %.3f", i, accu_tr)
-        local param_prefix = string.format("%s_%s_iter_%d_lr%f_tr%.3f",
+        param_prefix = string.format("%s_%s_iter_%d_lr%f_tr%.3f",
                                 string.gsub(
                                     (string.gsub(pf0[1], "(.*/)(.*)", "%2")),
                                     "(.*)%..*", "%1"),
diff --git a/nerv/layer/affine.lua b/nerv/layer/affine.lua
index a05ae17..a1c92b1 100644
--- a/nerv/layer/affine.lua
+++ b/nerv/layer/affine.lua
@@ -90,7 +90,8 @@ function AffineLayer:bind_params()
     self.ltp = self.ltp1 -- alias of ltp1
     self.bp = self:find_param("bp", self.lconf, self.gconf,
                                 nerv.BiasParam,
-                                {1, self.dim_out[1]})
+                                {1, self.dim_out[1]},
+                                nerv.Param.gen_zero)
 
 end
 
diff --git a/nerv/layer/bias.lua b/nerv/layer/bias.lua
index 191be78..03e310d 100644
--- a/nerv/layer/bias.lua
+++ b/nerv/layer/bias.lua
@@ -9,7 +9,8 @@ end
 function BiasLayer:bind_params()
     self.bias = self:find_param("bias", self.lconf, self.gconf,
                                 nerv.BiasParam,
-                                {1, self.dim_out[1]})
+                                {1, self.dim_out[1]},
+                                nerv.Param.gen_zero)
 end
 
 function BiasLayer:init()
diff --git a/nerv/layer/init.lua b/nerv/layer/init.lua
index d266773..c5b7657 100644
--- a/nerv/layer/init.lua
+++ b/nerv/layer/init.lua
@@ -27,6 +27,10 @@ function Param:update(gradient)
     nerv.error_method_not_implemented()
 end
 
+function Param:gen_zero()
+    return 0
+end
+
 local Layer = nerv.class('nerv.Layer')
 
 function Layer:__init(id, global_conf, layer_conf)
@@ -93,7 +97,7 @@ function Layer:get_sublayer(id)
     nerv.error('primitive layer does not have sublayers')
 end
 
-function Layer:find_param(plist, lconf, gconf, p_type, p_dim)
+function Layer:find_param(plist, lconf, gconf, p_type, p_dim, p_gen)
     if type(plist) == "string" then
         plist = {plist}
     end
@@ -120,10 +124,12 @@ function Layer:find_param(plist, lconf, gconf, p_type, p_dim)
                 "switch to auto-generate", plist_str, self.id)
     local p = p_type(pid, gconf)
     p.trans = self.mat_type(unpack(p_dim))
-    if type(gconf.param_random) ~= "function" then
+    p_gen = p_gen or gconf.param_gen
+                    or gconf.param_random -- obsolete name
+    if type(p_gen) ~= "function" then
         nerv.error("a param generate function is needed")
     end
-    p.trans:generate(gconf.param_random)
+    p.trans:generate(p_gen)
     return p
 end
 
diff --git a/nerv/layer/lstm_gate.lua b/nerv/layer/lstm_gate.lua
index 9d79b04..e3b11b4 100644
--- a/nerv/layer/lstm_gate.lua
+++ b/nerv/layer/lstm_gate.lua
@@ -18,7 +18,8 @@ function LSTMGateLayer:bind_params()
         end
     end
     self.bp = self:find_param("bp", self.lconf, self.gconf,
-                                nerv.BiasParam, {1, self.dim_out[1]})
+                                nerv.BiasParam, {1, self.dim_out[1]},
+                                nerv.Param.gen_zero)
 end
 
 function LSTMGateLayer:init(batch_size)
diff --git a/nerv/layer/window.lua b/nerv/layer/window.lua
index 364929f..729ab58 100644
--- a/nerv/layer/window.lua
+++ b/nerv/layer/window.lua
@@ -9,7 +9,8 @@ end
 function WindowLayer:bind_params()
     self.window = self:find_param("window", self.lconf, self.gconf,
                                 nerv.BiasParam,
-                                {1, self.dim_out[1]})
+                                {1, self.dim_out[1]},
+                                nerv.Param.gen_zero)
 end
 
 function WindowLayer:init()