Merge pull request #4 from uphantom/master

fastnn first version, include follow submodular

3 files changed, 1006 insertions, 0 deletions

diff --git a/fastnn/example/asgd_data_trainer.lua b/fastnn/example/asgd_data_trainer.lua
new file mode 100644
index 0000000..33d579a
--- /dev/null
+++ b/fastnn/example/asgd_data_trainer.lua
@@ -0,0 +1,405 @@
+require 'fastnn'
+require 'libhtkio'
+require 'threads'
+
+dofile("fastnn/fastnn_baseline.lua")
+
+env = string.format([[
+package.path="/home/slhome/wd007/.luarocks/share/lua/5.1/?.lua;/home/slhome/wd007/.luarocks/share/lua/5.1/?/init.lua;/sgfs/users/wd007/src/nerv/install/share/lua/5.1/?.lua;/sgfs/users/wd007/src/nerv/install/share/lua/5.1/?/init.lua;"..package.path; 
+package.cpath="/home/slhome/wd007/.luarocks/lib/lua/5.1/?.so;/sgfs/users/wd007/src/nerv/install/lib/lua/5.1/?.so;"..package.cpath
+local k,l,_=pcall(require,"luarocks.loader") _=k and l.add_context("nerv","scm-1")
+]])
+
+
+
+local data_thread_code = [[
+%s
+
+require 'nerv'
+require 'fastnn'
+dofile("fastnn/fastnn_baseline.lua")
+os.execute("export MALLOC_CHECK_=0")
+     
+local thread_idx = %d
+local example_repo_shareid = %d
+local data_mutex_shareid = %d
+local feat_repo_shareid = %d
+local gpu_shareid = %d
+local batch_size = %d
+local bp = %d
+local scp_file = '%s'
+
+local share_mutex = threads.Mutex(data_mutex_shareid)
+local share_example_repo = fastnn.CExamplesRepo(example_repo_shareid, true)
+local share_gpu = fastnn.CDevice(gpu_shareid)
+
+--print(thread_idx)
+--print(share_mutex)
+--print(share_gpu)
+--print(share_example_repo)
+
+if bp == 0 then
+	bp = false 
+else
+   	bp = true
+end
+gconf.randomize = bp
+--print(gconf.randomize)
+
+share_mutex:lock()
+local gpuid = share_example_repo:get_gpuid()
+if gpuid < 0 then 
+        gpuid = share_gpu:select_gpu()
+        share_example_repo:set_gpuid(gpuid)
+else
+        share_gpu:select_gpu(gpuid)     
+end
+
+nerv.info_stderr("thread %%d loading transf ...", thread_idx)
+local param_transf_repo = nerv.ParamRepo()
+param_transf_repo:import(gconf.transf, nil, gconf)
+local transf_node_repo = make_transf_node_repo(param_transf_repo)
+local transf_layer_repo = make_transf_link_repo(transf_node_repo, param_transf_repo)
+local transf = transf_layer_repo:get_layer("global_transf")
+share_mutex:unlock()
+
+local feat_id = get_feat_id()
+
+local buffer = make_buffer(make_readers(scp_file, transf_layer_repo, feat_repo_shareid, data_mutex_shareid))
+
+        local t = 1;
+        for data in buffer.get_data, buffer do
+                local example = fastnn.Example:PrepareData(data, nil, feat_id)
+                --print(string.format("Accept NO.%%d %%s", t, example)); t = t+1;
+                share_example_repo:accept(example)
+		--print("share_example_repo:accept")
+
+            -- collect garbage in-time to save GPU memory
+            collectgarbage("collect")
+        end 
+        share_example_repo:done()
+--	print("share_example_repo:done")
+
+]]
+
+
+train_thread_code = [[
+%s
+
+require 'nerv'
+require 'fastnn'
+dofile("fastnn/fastnn_baseline.lua")
+os.execute("export MALLOC_CHECK_=0")
+
+local thread_idx = %d
+local example_repo_shareid = %d
+local data_mutex_shareid = %d 
+local master_shareid = %d
+local gpu_shareid = %d
+local xent_shareid = %d
+local batch_size = %d
+local lrate = %f
+local bp = %d
+local nnet_in = '%s'
+local nnet_out = '%s'
+
+local share_example_repo = fastnn.CExamplesRepo(example_repo_shareid, true)
+local share_mutex = threads.Mutex(data_mutex_shareid)
+local share_master = fastnn.ModelSync(master_shareid)
+local share_gpu = fastnn.CDevice(gpu_shareid)
+local share_xent = fastnn.CXent(xent_shareid)
+
+if bp == 0 then
+	bp = false
+else
+	bp = true
+end
+
+gconf.randomize = bp
+gconf.lrate = lrate
+gconf.batch_size = batch_size
+gconf.network[1] = nnet_in
+nerv.info_stderr("input network: %%s", gconf.network[1])
+nerv.info_stderr(gconf.randomize)
+nerv.info_stderr("input batch_size: %%d", gconf.batch_size)
+nerv.info_stderr("input lrate: %%f", gconf.lrate)
+
+share_mutex:lock()
+local gpuid = share_example_repo:get_gpuid()
+if gpuid < 0 then 
+        gpuid = share_gpu:select_gpu()
+        share_example_repo:set_gpuid(gpuid)
+else
+        share_gpu:select_gpu(gpuid)     
+end
+
+nerv.context = nerv.CCuContext()
+--print(nerv.context)
+
+
+nerv.info_stderr("thread %%d loading network ...", thread_idx)
+local param_network_repo = nerv.ParamRepo()
+param_network_repo:import(gconf.network, nil, gconf)
+local network_node_repo = make_network_node_repo(param_network_repo)
+local network_layer_repo = make_network_link_repo(network_node_repo, param_network_repo)
+local network = get_network(network_layer_repo)
+share_mutex:unlock()
+
+
+	local input_order = get_input_order()
+
+	-- initialize the network
+        network:init(gconf.batch_size)
+	gconf.cnt = 0 
+        err_input = {nerv.CuMatrixFloat(gconf.batch_size, 1)} 
+        err_input[1]:fill(1)
+
+	share_master:Initialize(network)
+	share_master:SyncInc()
+
+	for example in share_example_repo.provide, share_example_repo do
+
+		gconf.cnt = gconf.cnt + 1
+		if gconf.cnt == 2000 then
+			print_stat(network_node_repo)
+			gconf.cnt = 0
+		end
+
+		local input = {}
+                local n = example:size()
+                for i = 0, n-1 do
+                        table.insert(input, example:at(i))
+                end 
+
+		local output = {nerv.CuMatrixFloat(gconf.batch_size, 1)} 
+		err_output = {input[1]:create()}
+            	network:propagate(input, output)
+
+		if bp then
+                	network:back_propagate(err_input, err_output, input, output)
+                	network:gradient(err_input, input, output)
+
+			share_master:LockModel()
+			share_master:WeightToD(network)			
+                	network:update_gradient()
+			share_master:WeightFromD(network)			
+			share_master:UnLockModel()
+            	end 
+
+            -- collect garbage in-time to save GPU memory
+            collectgarbage("collect")
+        end 
+
+	--print_stat(network_node_repo)
+	local ce_crit = network_node_repo:get_layer("ce_crit")
+	local xent = fastnn.CXent(ce_crit.total_frames, ce_crit.total_correct, ce_crit.total_ce, ce_crit.total_ce)
+
+	share_master:LockModel()
+	share_xent:add(xent)
+	share_master:SyncDec()
+	--print(string.format("ThreadCount: %%d", share_master:ThreadCount()))
+	if share_master:ThreadCount() == 0 and bp then
+		 share_master:WeightToD(network)
+		 local fname = string.format("%%s_tr%%.3f",
+                            nnet_out, frame_acc(share_xent))
+		 nerv.info_stderr("writing back %%s ...", fname)
+		 network:get_params():export(fname, nil)
+	end
+	share_master:UnLockModel()
+]]
+
+function get_data_thread(data_thread_code, env, thread_idx, example_repo_shareid,  
+                                                data_mutex_shareid, feat_repo_shareid, gpu_shareid,  
+                                                batch_size, bp, scp_file)
+        return string.format(data_thread_code, env, thread_idx, example_repo_shareid,  
+                                                data_mutex_shareid, feat_repo_shareid, gpu_shareid, 
+                                                batch_size, bp, scp_file)
+end
+
+function get_train_thread(train_thread_code, env, thread_idx, example_repo_shareid,  
+                                                data_mutex_shareid, master_shareid, gpu_shareid, xent_shareid, 
+						batch_size, lrate, bp, nnet_in, nnet_out)
+        return string.format(train_thread_code, env, thread_idx, example_repo_shareid,  
+                                                data_mutex_shareid, master_shareid, gpu_shareid, xent_shareid, 
+						batch_size, lrate, bp, nnet_in, nnet_out)
+end
+
+function trainer(batch_size, lrate, bp, scp_file, nnet_in, nnet_out, num_threads)
+	local train_threads={}
+	local trainer = {}
+	local data_threads = {}
+	local data = {}
+	local num_threads=num_threads
+
+	local data_mutex = threads.Mutex()
+	local data_mutex_shareid = data_mutex:id()
+
+	local master = fastnn.CModelSync()
+	local master_shareid = master:id()
+	--print(master)
+
+	local xent = fastnn.CXent()
+	local xent_shareid = xent:id()
+	--print(xent)
+
+	local gpu = fastnn.CDevice()
+	local gpu_shareid = gpu:id()
+	--print(gpu_shareid)
+	gpu:init()
+
+	local example_repo = {}  
+	local example_repo_shareid = {}  
+
+	local feat_repo = nerv.TNetFeatureRepo(scp_file, gconf.htk_conf, gconf.frm_ext)
+	local feat_repo_shareid = feat_repo:id()
+
+	for i=1,num_threads,1 do
+		example_repo[i] = fastnn.CExamplesRepo(128, false)
+        	example_repo_shareid[i] = example_repo[i]:id()  
+		
+		data_threads[i] = get_data_thread(data_thread_code, env, i, example_repo_shareid[i],
+                                                data_mutex_shareid, feat_repo_shareid, gpu_shareid,
+                                                batch_size, bp, scp_file)
+
+        	train_threads[i] = get_train_thread(train_thread_code, env, i, example_repo_shareid[i],
+                                                data_mutex_shareid, master_shareid, gpu_shareid, xent_shareid, 
+						batch_size, lrate, bp, nnet_in, nnet_out)
+		--print(train_threads[i])
+        	data[i] = threads.Thread(data_threads[i])
+        	trainer[i] = threads.Thread(train_threads[i])
+	end
+
+	nerv.info_stderr('| waiting for thread...')
+
+	for i=1,num_threads,1 do
+		data[i]:free()
+        	trainer[i]:free()
+	end
+
+	print_xent(xent)
+
+	nerv.info_stderr('| all thread finished!')
+
+	return frame_acc(xent)
+end
+
+function get_filename(fname)
+	return string.gsub((string.gsub(fname, "(.*/)(.*)", "%2")),"(.*)%..*", "%1")
+end
+
+function do_sds(tr_scp, sds_scp, sds_rate)
+        math.randomseed(os.time())
+        local scp_file = io.open(tr_scp, "r")
+        local sds_file = io.open(sds_scp, "w")
+        for line in scp_file:lines() do
+		rate = math.random()
+		if (rate < sds_rate) then
+			sds_file:write(line.."\n")
+		end 
+	end 
+	scp_file:close()
+	sds_file:close()
+end
+
+function print_tag(iter)
+	 io.stderr:write(string.format("########################################################\n"))
+	 io.stderr:write(string.format("# NN TRAINING ITERATION:%d, %s\n", iter, os.date()))
+	 io.stderr:write(string.format("########################################################\n"))
+end
+	
+
+start_halving_inc = 0.5
+halving_factor = 0.8
+end_halving_inc = 0.1
+min_iter = 1
+max_iter = 20
+min_halving = 0
+gconf.batch_size = 256
+pf0 = get_filename(gconf.network[1])
+nnet_in = gconf.network[1]
+nnet_out = ""
+sds_scp =  "tr_sds_"..string.format("%.4d", math.random()*10000)..".scp"  --"tr_sds.scp"
+sds_factor = 0.4
+num_threads = 1
+global_option = nil
+
+print_gconf()
+os.execute("export MALLOC_CHECK_=0")
+
+-- training begin
+nerv.info_stderr("begin initial cross validation")  
+local accu_best = trainer(gconf.batch_size, gconf.lrate, 0, 
+				gconf.cv_scp, nnet_in, nil, num_threads) 
+local do_halving = false
+local accu_new = accu_best
+
+nerv.info_stderr("initial cross validation: %.3f\n", accu_best)  
+
+for i = 1, max_iter do
+
+	if accu_new >= accu_best then
+		local sds_rate = math.cos((i-1)*11.0/180*math.pi)
+		if (sds_rate <= sds_factor) then
+			sds_rate = sds_factor
+		end
+		nerv.info_stderr("iteration %d sds_rate: %.6f", i, sds_rate)
+		do_sds(gconf.tr_scp, sds_scp, sds_rate)
+	end
+		
+        nnet_out=pf0.."_iter"..i
+	--print(nnet_out)
+	print_tag(i)
+	nerv.info_stderr("[NN] begin iteration %d learning_rate: %.3f batch_size: %d.", i, gconf.lrate, gconf.batch_size)
+	local accu_tr = trainer(gconf.batch_size, gconf.lrate, 1,
+					sds_scp, nnet_in, nnet_out, num_threads)
+	nerv.info_stderr("[TR] end iteration %d frame_accuracy: %.3f.\n", i, accu_tr)	
+	os.execute("sleep " .. 3)
+
+	nnet_out = nnet_out.."_tr"..accu_tr
+	accu_new = trainer(gconf.batch_size, gconf.lrate, 0,
+					gconf.cv_scp, nnet_out, nil, num_threads)
+	nerv.info_stderr("[CV] end iteration %d frame_accuracy: %.3f.\n\n", i, accu_new)	
+	os.execute("sleep " .. 3)
+
+	local nnet_tmp = string.format("%s_%s_iter_%d_lr%f_tr%.3f_cv%.3f", 
+					pf0,
+					os.date("%Y%m%d%H%M%S"),
+   					i, gconf.lrate, accu_tr, accu_new)
+	
+	-- TODO: revert the weights
+	local accu_diff = accu_new - accu_best
+	local cmd
+	if accu_new > accu_best then
+		accu_best = accu_new
+		nnet_in = nnet_tmp
+		gconf.batch_size = gconf.batch_size + 128
+		if gconf.batch_size > 1024 then
+			gconf.batch_size = 1024
+		end
+	else  
+	-- reject
+		nnet_tmp = nnet_tmp.."_rejected"
+		do_halving = true
+	end
+	cmd = "mv "..nnet_out.." "..nnet_tmp
+	os.execute(cmd)	
+
+	if do_halving and accu_diff < end_halving_inc and i > min_iter then
+		break;
+	end
+	
+	if accu_diff < start_halving_inc and i >= min_halving then
+		do_halving = true
+	end
+
+	if do_halving then
+		gconf.lrate = gconf.lrate * halving_factor
+		halving_factor = halving_factor - 0.025
+		if halving_factor < 0.6 then
+			halving_factor = 0.6
+		end
+	end
+	nerv.info_stderr("iteration %d done!", i)
+end
+
+
diff --git a/fastnn/example/asgd_sds_trainer.lua b/fastnn/example/asgd_sds_trainer.lua
new file mode 100644
index 0000000..cf1c7a6
--- /dev/null
+++ b/fastnn/example/asgd_sds_trainer.lua
@@ -0,0 +1,343 @@
+
+NERV_ROOT = "/sgfs/users/wd007/src/nerv-2"
+
+env = string.format([[
+package.path="/home/slhome/wd007/.luarocks/share/lua/5.1/?.lua;/home/slhome/wd007/.luarocks/share/lua/5.1/?/init.lua;%s/install/share/lua/5.1/?.lua;%s/install/share/lua/5.1/?/init.lua;"..package.path; 
+package.cpath="/home/slhome/wd007/.luarocks/lib/lua/5.1/?.so;%s/install/lib/lua/5.1/?.so;"..package.cpath
+local k,l,_=pcall(require,"luarocks.loader") _=k and l.add_context("nerv","scm-1")
+]], NERV_ROOT, NERV_ROOT, NERV_ROOT)
+
+loadstring(env)()
+
+require 'nerv'
+
+require 'fastnn'
+require 'libhtkio'
+require 'threads'
+
+dofile("fastnn/example/fastnn_baseline.lua")
+
+
+
+train_thread_code = [[
+%s
+
+require 'nerv'
+require 'fastnn'
+require 'libhtkio'
+
+dofile("fastnn/example/fastnn_baseline.lua")
+os.execute("export MALLOC_CHECK_=0")
+
+local thread_idx = %d
+local feat_repo_shareid = %d
+local data_mutex_shareid = %d 
+local master_shareid = %d
+local gpu_shareid = %d
+local xent_shareid = %d
+local batch_size = %d
+local lrate = %f
+local bp = %d
+local scp_file = '%s'
+local nnet_in = '%s'
+local nnet_out = '%s'
+
+local share_mutex = threads.Mutex(data_mutex_shareid)
+local share_master = fastnn.ModelSync(master_shareid)
+local share_gpu = fastnn.CDevice(gpu_shareid)
+local share_xent = fastnn.CXent(xent_shareid)
+
+if bp == 0 then
+	bp = false
+else
+	bp = true
+	gconf.tr_scp = scp_file
+end
+
+share_mutex:lock()
+
+gconf.randomize = bp
+gconf.lrate = lrate
+gconf.batch_size = batch_size
+gconf.initialized_param[2] = nnet_in
+nerv.info_stderr("input network: %%s", gconf.initialized_param[2])
+--nerv.info_stderr(gconf.randomize)
+nerv.info_stderr("input batch_size: %%d", gconf.batch_size)
+nerv.info_stderr("input scp_file: %%s", scp_file)
+nerv.info_stderr("input lrate: %%f", gconf.lrate)
+
+
+
+share_gpu:select_gpu()
+
+nerv.context = nerv.CCuContext()
+--print(nerv.context)
+
+nerv.info_stderr("thread %%d loading parameters ...", thread_idx)
+local param_repo = nerv.ParamRepo()
+param_repo:import(gconf.initialized_param, nil, gconf)
+local layer_repo = make_layer_repo(param_repo)
+local network = get_network(layer_repo)
+local global_transf = get_global_transf(layer_repo)
+
+share_mutex:unlock()
+
+local buffer = make_buffer(make_readers(nil, layer_repo, feat_repo_shareid, data_mutex_shareid))
+
+local input_order = get_input_order()
+
+	-- initialize the network
+        network:init(gconf.batch_size)
+	gconf.cnt = 0 
+        err_input = {nerv.CuMatrixFloat(gconf.batch_size, 1)} 
+        err_input[1]:fill(1)
+
+	share_master:Initialize(network)
+	share_master:SyncInc()
+
+        for data in buffer.get_data, buffer do
+
+		gconf.cnt = gconf.cnt + 1
+		if gconf.cnt == 2000 then
+			print_stat(layer_repo)
+			gconf.cnt = 0
+		end
+
+		local input = {}
+		
+		for i, e in ipairs(input_order) do
+                	local id = e.id
+                	if data[id] == nil then
+                    		nerv.error("input data %%s not found", id)
+                	end
+                	local transformed
+                	if e.global_transf then
+                    	transformed = nerv.speech_utils.global_transf(data[id],
+                                        global_transf,
+                                        gconf.frm_ext or 0, 0,
+                                        gconf)
+                	else
+                    		transformed = data[id]
+                	end
+	                table.insert(input, transformed)
+            	end
+
+            	local output = {nerv.CuMatrixFloat(gconf.batch_size, 1)}
+            	err_output = {}
+            	for i = 1, #input do
+                	table.insert(err_output, input[i]:create())
+		end
+
+            	network:propagate(input, output)
+
+		if bp then
+                	network:back_propagate(err_input, err_output, input, output)
+                	network:gradient(err_input, input, output)
+
+			share_master:LockModel()
+			share_master:WeightToD(network)			
+                	network:update_gradient()
+              --  	network:update(err_input, input, output)
+			share_master:WeightFromD(network)			
+			share_master:UnLockModel()
+            	end 
+
+            -- collect garbage in-time to save GPU memory
+            collectgarbage("collect")
+        end 
+
+	--print_stat(network_node_repo)
+	local ce_crit = layer_repo:get_layer("ce_crit")
+	local xent = fastnn.CXent(ce_crit.total_frames, ce_crit.total_correct, ce_crit.total_ce, ce_crit.total_ce)
+
+	share_master:LockModel()
+	share_xent:add(xent)
+	share_master:SyncDec()
+	--print(string.format("ThreadCount: %%d", share_master:ThreadCount()))
+	if share_master:ThreadCount() == 0 and bp then
+		 share_master:WeightToD(network)
+		 local fname = string.format("%%s_tr%%.3f",
+                            nnet_out, frame_acc(share_xent))
+		 nerv.info_stderr("writing back %%s ...", fname)
+		 network:get_params():export(fname, nil)
+	end
+	share_master:UnLockModel()
+]]
+
+
+function get_train_thread(train_thread_code, env, thread_idx, feat_repo_shareid,  
+                                                data_mutex_shareid, master_shareid, gpu_shareid, xent_shareid, 
+						batch_size, lrate, bp, scp_file, nnet_in, nnet_out)
+        return string.format(train_thread_code, env, thread_idx, feat_repo_shareid,  
+                                                data_mutex_shareid, master_shareid, gpu_shareid, xent_shareid, 
+						batch_size, lrate, bp, scp_file, nnet_in, nnet_out)
+end
+
+function trainer(batch_size, lrate, bp, scp_file, nnet_in, nnet_out, num_threads)
+	local train_threads={}
+	local trainer = {}
+	local num_threads=num_threads
+
+	local data_mutex = threads.Mutex()
+	local data_mutex_shareid = data_mutex:id()
+
+	local master = fastnn.CModelSync()
+	local master_shareid = master:id()
+	--print(master)
+
+	local xent = fastnn.CXent()
+	local xent_shareid = xent:id()
+	--print(xent)
+
+	local gpu = fastnn.CDevice()
+	local gpu_shareid = gpu:id()
+	--print(gpu_shareid)
+	gpu:init()
+
+	local feat_repo = nerv.TNetFeatureRepo(scp_file, gconf.htk_conf, gconf.frm_ext)
+	local feat_repo_shareid = feat_repo:id()
+
+	for i=1,num_threads,1 do
+
+        	train_threads[i] = get_train_thread(train_thread_code, env, i, feat_repo_shareid,
+                                                data_mutex_shareid, master_shareid, gpu_shareid, xent_shareid, 
+						batch_size, lrate, bp, scp_file, nnet_in, nnet_out)
+		--print(train_threads[i])
+        	trainer[i] = threads.Thread(train_threads[i])
+	end
+
+	nerv.info_stderr('| waiting for thread...')
+
+	for i=1,num_threads,1 do
+        	trainer[i]:free()
+	end
+
+	print_xent(xent)
+
+	nerv.info_stderr('| all thread finished!')
+
+	return frame_acc(xent)
+end
+
+function get_filename(fname)
+	return string.gsub((string.gsub(fname, "(.*/)(.*)", "%2")),"(.*)%..*", "%1")
+end
+
+function do_sds(tr_scp, sds_scp, sds_rate)
+        math.randomseed(os.time())
+        local scp_file = io.open(tr_scp, "r")
+        local sds_file = io.open(sds_scp, "w")
+        for line in scp_file:lines() do
+		rate = math.random()
+		if (rate < sds_rate) then
+			sds_file:write(line.."\n")
+		end 
+	end 
+	scp_file:close()
+	sds_file:close()
+end
+
+function print_tag(iter)
+	 io.stderr:write(string.format("########################################################\n"))
+	 io.stderr:write(string.format("# NN TRAINING ITERATION:%d, %s\n", iter, os.date()))
+	 io.stderr:write(string.format("########################################################\n"))
+end
+	
+
+start_halving_inc = 0.5
+halving_factor = 0.8
+end_halving_inc = 0.1
+min_iter = 1
+max_iter = 20
+min_halving = 0
+gconf.batch_size = 256
+pf0 = get_filename(gconf.initialized_param[2])
+nnet_in = gconf.initialized_param[2]
+nnet_out = ""
+sds_scp = "tr_sds_"..string.format("%.4d", math.random()*10000)..".scp"  --"tr_sds.scp"
+sds_factor = 0.4
+num_threads = 2
+global_option = nil
+
+os.execute("export MALLOC_CHECK_=0")
+print_gconf()
+
+-- training begin
+nerv.info_stderr("begin initial cross validation")  
+accu_best = trainer(gconf.batch_size, gconf.lrate, 0, 
+				gconf.cv_scp, nnet_in, "", num_threads) 
+local do_halving = false
+local accu_new = accu_best
+
+nerv.info_stderr("initial cross validation: %.3f\n", accu_best)  
+
+for i = 1, max_iter do
+
+	if accu_new >= accu_best then
+		local sds_rate = math.cos((i-1)*11.0/180*math.pi)
+		if (sds_rate <= sds_factor) then
+			sds_rate = sds_factor
+		end
+		nerv.info_stderr("iteration %d sds_rate: %.6f", i, sds_rate)
+		do_sds(gconf.tr_scp, sds_scp, sds_rate)
+	end
+		
+        nnet_out=pf0.."_iter"..i
+	--print(nnet_out)
+	print_tag(i)
+	nerv.info_stderr("[NN] begin iteration %d learning_rate: %.3f batch_size: %d.", i, gconf.lrate, gconf.batch_size)
+	accu_tr = trainer(gconf.batch_size, gconf.lrate, 1,
+					sds_scp, nnet_in, nnet_out, num_threads)
+	collectgarbage("collect")
+	nerv.info_stderr("[TR] end iteration %d frame_accuracy: %.3f.\n", i, accu_tr)	
+	os.execute("sleep " .. 3)
+
+	nnet_out = nnet_out.."_tr"..accu_tr
+	accu_new = trainer(gconf.batch_size, gconf.lrate, 0,
+					gconf.cv_scp, nnet_out, "", num_threads)
+	collectgarbage("collect")
+	nerv.info_stderr("[CV] end iteration %d frame_accuracy: %.3f.\n\n", i, accu_new)	
+	os.execute("sleep " .. 3)
+
+	local nnet_tmp = string.format("%s_%s_iter_%d_lr%f_tr%.3f_cv%.3f", 
+					pf0,
+					os.date("%Y%m%d%H%M%S"),
+   					i, gconf.lrate, accu_tr, accu_new)
+	
+	-- TODO: revert the weights
+	local accu_diff = accu_new - accu_best
+	local cmd
+	if accu_new > accu_best then
+		accu_best = accu_new
+		nnet_in = nnet_tmp
+		gconf.batch_size = gconf.batch_size + 128
+		if gconf.batch_size > 1024 then
+			gconf.batch_size = 1024
+		end
+	else  
+	-- reject
+		nnet_tmp = nnet_tmp.."_rejected"
+		do_halving = true
+	end
+	cmd = "mv "..nnet_out.." "..nnet_tmp
+	os.execute(cmd)	
+
+	if do_halving and accu_diff < end_halving_inc and i > min_iter then
+		break;
+	end
+	
+	if accu_diff < start_halving_inc and i >= min_halving then
+		do_halving = true
+	end
+
+	if do_halving then
+		gconf.lrate = gconf.lrate * halving_factor
+		halving_factor = halving_factor - 0.025
+		if halving_factor < 0.6 then
+			halving_factor = 0.6
+		end
+	end
+	nerv.info_stderr("iteration %d done!", i)
+end
+
+
diff --git a/fastnn/example/fastnn_baseline.lua b/fastnn/example/fastnn_baseline.lua
new file mode 100644
index 0000000..6e774de
--- /dev/null
+++ b/fastnn/example/fastnn_baseline.lua
@@ -0,0 +1,258 @@
+require 'htk_io'
+
+gconf = {lrate = 0.2, wcost = 1e-6, momentum = 0.9,
+        cumat_type = nerv.CuMatrixFloat,
+        mmat_type = nerv.MMatrixFloat,
+        frm_ext = 5,
+        frm_trim = 5,
+	batch_size = 256,
+	buffer_size = 81920,
+	rearrange = true,
+        tr_scp = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/train.scp",
+        cv_scp = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/train_cv.scp",
+        htk_conf = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/fbank_d_a_z.conf",
+	initialized_param = {"/sgfs/users/wd007/src/nerv/tools/nerv.global.transf",
+                             "/sgfs/users/wd007/src/nerv/tools/nerv.svd0.55_3000h_iter1.init"},
+        debug = false}
+
+function make_layer_repo(param_repo) 
+    local layer_repo = nerv.LayerRepo( 
+    {
+        -- global transf
+        ["nerv.BiasLayer"] =
+        {   
+            blayer1 = {{bias = "bias1"}, {dim_in = {1320}, dim_out = {1320}}},
+        },  
+        ["nerv.WindowLayer"] =
+        {   
+            wlayer1 = {{window = "window1"}, {dim_in = {1320}, dim_out = {1320}}},
+	},
+        -- biased linearity
+        ["nerv.AffineLayer"] =
+        {
+            affine0 = {{ltp = "affine0_ltp", bp = "affine0_bp"},
+            {dim_in = {1320}, dim_out = {2048}}},
+            affine1 = {{ltp = "affine1_ltp", bp = "affine1_bp"},
+            {dim_in = {2048}, dim_out = {367}}},
+            affine2 = {{ltp = "affine2_ltp", bp = "affine2_bp"},
+            {dim_in = {367}, dim_out = {2048}}},
+            affine3 = {{ltp = "affine3_ltp", bp = "affine3_bp"},
+            {dim_in = {2048}, dim_out = {408}}},
+            affine4 = {{ltp = "affine4_ltp", bp = "affine4_bp"},
+            {dim_in = {408}, dim_out = {2048}}},
+            affine5 = {{ltp = "affine5_ltp", bp = "affine5_bp"},
+            {dim_in = {2048}, dim_out = {368}}},
+            affine6 = {{ltp = "affine6_ltp", bp = "affine6_bp"},
+            {dim_in = {368}, dim_out = {2048}}},
+            affine7 = {{ltp = "affine7_ltp", bp = "affine7_bp"},
+            {dim_in = {2048}, dim_out = {303}}},
+            affine8 = {{ltp = "affine8_ltp", bp = "affine8_bp"},
+            {dim_in = {303}, dim_out = {2048}}},
+            affine9 = {{ltp = "affine9_ltp", bp = "affine9_bp"},
+            {dim_in = {2048}, dim_out = {277}}},
+            affine10 = {{ltp = "affine10_ltp", bp = "affine10_bp"},
+            {dim_in = {277}, dim_out = {2048}}},
+            affine11 = {{ltp = "affine11_ltp", bp = "affine11_bp"},
+            {dim_in = {2048}, dim_out = {361}}},
+            affine12 = {{ltp = "affine12_ltp", bp = "affine12_bp"},
+            {dim_in = {361}, dim_out = {2048}}},
+            affine13 = {{ltp = "affine13_ltp", bp = "affine13_bp"},
+            {dim_in = {2048}, dim_out = {441}}},
+            affine14 = {{ltp = "affine14_ltp", bp = "affine14_bp"},
+            {dim_in = {441}, dim_out = {10092}}},
+	},
+        ["nerv.SigmoidLayer"] =
+        {
+            sigmoid0 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid1 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid2 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid3 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid4 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid5 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+            sigmoid6 = {{}, {dim_in = {2048}, dim_out = {2048}}},
+        },
+        ["nerv.SoftmaxCELayer"] = -- softmax + ce criterion layer for finetune output
+        {
+            ce_crit = {{}, {dim_in = {10092, 1}, dim_out = {1}, compressed = true}}
+        },
+	["nerv.SoftmaxLayer"] = -- softmax for decode output
+        {
+            softmax = {{}, {dim_in = {10092}, dim_out = {10092}}}
+        }
+    }, param_repo, gconf)
+    
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            global_transf = {{}, {
+                dim_in = {1320}, dim_out = {1320},
+                sub_layers = layer_repo,
+                connections = 
+		{
+                    ["<input>[1]"] = "blayer1[1]",
+                    ["blayer1[1]"] = "wlayer1[1]",
+                    ["wlayer1[1]"] = "<output>[1]"
+                }
+            }},
+	    main = {{}, {
+                dim_in = {1320}, dim_out = {10092},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "affine0[1]",
+                    ["affine0[1]"] = "sigmoid0[1]",
+                    ["sigmoid0[1]"] = "affine1[1]",
+                    ["affine1[1]"] = "affine2[1]",
+                    ["affine2[1]"] = "sigmoid1[1]",
+                    ["sigmoid1[1]"] = "affine3[1]",
+                    ["affine3[1]"] = "affine4[1]",
+                    ["affine4[1]"] = "sigmoid2[1]",
+                    ["sigmoid2[1]"] = "affine5[1]",
+                    ["affine5[1]"] = "affine6[1]",
+                    ["affine6[1]"] = "sigmoid3[1]",
+                    ["sigmoid3[1]"] = "affine7[1]",
+                    ["affine7[1]"] = "affine8[1]",
+                    ["affine8[1]"] = "sigmoid4[1]",
+                    ["sigmoid4[1]"] = "affine9[1]",
+                    ["affine9[1]"] = "affine10[1]",
+                    ["affine10[1]"] = "sigmoid5[1]",
+                    ["sigmoid5[1]"] = "affine11[1]",
+                    ["affine11[1]"] = "affine12[1]",
+                    ["affine12[1]"] = "sigmoid6[1]",
+                    ["sigmoid6[1]"] = "affine13[1]",
+                    ["affine13[1]"] = "affine14[1]",
+                    ["affine14[1]"] = "<output>[1]",
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            ce_output = {{}, {
+                dim_in = {1320, 1}, dim_out = {1},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "ce_crit[1]",
+                    ["<input>[2]"] = "ce_crit[2]",
+                    ["ce_crit[1]"] = "<output>[1]"
+                }
+            }},
+            softmax_output = {{}, {
+                dim_in = {1320}, dim_out = {10092},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "softmax[1]",
+                    ["softmax[1]"] = "<output>[1]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    return layer_repo
+end
+
+
+function get_network(layer_repo)
+    return layer_repo:get_layer("ce_output")
+end
+
+function get_decode_network(layer_repo)
+    return layer_repo:get_layer("softmax_output")
+end
+
+function get_global_transf(layer_repo)
+    return layer_repo:get_layer("global_transf")
+end
+
+
+
+function make_readers(scp_file, layer_repo, feat_repo_shareid, data_mutex_shareid)
+    return {
+                {reader = nerv.TNetReader(gconf,
+                    {   
+                        id = "main_scp",
+                        scp_file = scp_file,
+                        conf_file = gconf.htk_conf,
+                        frm_ext = gconf.frm_ext,
+                        mlfs = { 
+                            phone_state = { 
+                                file = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/ref.mlf",
+                                format = "map",
+                                format_arg = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/dict",
+                                dir = "*/",
+                                ext = "lab"
+                            }   
+                        },  
+                        global_transf = layer_repo:get_layer("global_transf")
+                    }, feat_repo_shareid, data_mutex_shareid), 
+                data = {main_scp = 1320, phone_state = 1}} 
+            }   
+end
+
+function get_feat_id()
+        return {main_scp = true}
+end
+
+
+function make_buffer(readers)
+    return nerv.SGDBuffer(gconf,
+        {
+            buffer_size = gconf.buffer_size,
+            randomize = gconf.randomize,
+            readers = readers,
+            use_gpu = true
+        })
+end
+
+function get_input_order()
+    return {{id = "main_scp", global_transf = true},
+            {id = "phone_state"}}
+end
+
+function get_accuracy(layer_repo)
+    local ce_crit = layer_repo:get_layer("ce_crit")
+    return ce_crit.total_correct / ce_crit.total_frames * 100
+end
+
+function print_stat(layer_repo)
+    local ce_crit = layer_repo:get_layer("ce_crit")
+    nerv.info("*** training stat begin ***")
+    nerv.printf("cross entropy:\t\t%.8f\n", ce_crit.total_ce)
+    nerv.printf("correct:\t\t%d\n", ce_crit.total_correct)
+    nerv.printf("frames:\t\t\t%d\n", ce_crit.total_frames)
+    nerv.printf("err/frm:\t\t%.8f\n", ce_crit.total_ce / ce_crit.total_frames)
+    nerv.printf("accuracy:\t\t%.3f%%\n", get_accuracy(layer_repo))
+    nerv.info("*** training stat end ***")
+end
+
+function print_xent(xent)
+	local totalframes = xent:totalframes()
+	local loss = xent:loss()
+	local correct = xent:correct()
+	nerv.info_stderr("*** training statistics info begin ***")
+	nerv.info_stderr("total frames:\t\t%d", totalframes)
+	nerv.info_stderr("cross entropy:\t%.8f", loss/totalframes)
+	nerv.info_stderr("frame accuracy:\t%.3f%%", 100*correct/totalframes)
+	nerv.info_stderr("*** training statistics info end ***")
+end
+
+function frame_acc(xent)
+	local correct = xent:correct()
+	local totalframes = xent:totalframes()
+	return string.format("%.3f", 100*correct/totalframes)
+end
+
+function print_gconf()
+	nerv.info_stderr("%s \t:= %s", "network", gconf.initialized_param[1])
+	nerv.info_stderr("%s \t:= %s", "transf", gconf.initialized_param[2])
+	nerv.info_stderr("%s \t:= %s", "batch_size", gconf.batch_size)
+	nerv.info_stderr("%s \t:= %s", "buffer_size", gconf.buffer_size)
+	nerv.info_stderr("%s \t:= %s", "lrate", gconf.lrate)
+	nerv.info_stderr("%s \t:= %s", "tr_scp", gconf.tr_scp)
+	nerv.info_stderr("%s \t:= %s", "cv_scp", gconf.cv_scp)
+end