made lm sampling code cleaner

2 files changed, 72 insertions, 57 deletions

diff --git a/nerv/examples/lmptb/lm_sampler.lua b/nerv/examples/lmptb/lm_sampler.lua
index 2a4f1c3..d194af9 100644
--- a/nerv/examples/lmptb/lm_sampler.lua
+++ b/nerv/examples/lmptb/lm_sampler.lua
@@ -3,18 +3,19 @@ local LMSampler = nerv.class('nerv.LMSampler')
 function LMSampler:__init(global_conf)
     self.log_pre = "LMSampler"
     self.gconf = global_conf
+    self.batch_size = self.gconf.batch_size
+    self.chunk_size = self.gconf.chunk_size --largest sample sentence length
     self.vocab = self.gconf.vocab
     self.sen_end_token = self.vocab.sen_end_token
     self.sen_end_id = self.vocab:get_word_str(self.sen_end_token).id 
 end
 
 function LMSampler:load_dagL(dagL)
-    self.batch_size = self.gconf.batch_size
-    self.chunk_size = self.gconf.chunk_size
-    
+   
     nerv.printf("%s loading dagL\n", self.log_pre)
 
     self.dagL = dagL
+    self.dagL:init(self.batch_size)
 
     self.dagL_inputs = {}
     self.dagL_inputs[1] = global_conf.cumat_type(global_conf.batch_size, 1)
diff --git a/nerv/examples/lmptb/sample_grulm_ptb_main.lua b/nerv/examples/lmptb/sample_grulm_ptb_main.lua
index 30dfe26..42a5787 100644
--- a/nerv/examples/lmptb/sample_grulm_ptb_main.lua
+++ b/nerv/examples/lmptb/sample_grulm_ptb_main.lua
@@ -134,10 +134,39 @@ function prepare_tnn(global_conf, layerRepo)
     return tnn
 end
 
-function prepare_dagL(global_conf, layerRepo)
-    nerv.printf("%s Generate and initing dagL ...\n", global_conf.sche_log_pre)
+function load_net_tnn(global_conf, fn)
+    prepare_parameters(global_conf, fn)
+    local layerRepo = prepare_layers(global_conf)
+    local tnn = prepare_tnn(global_conf, layerRepo)
+    return tnn
+end
+
+function prepare_sampler(sm_conf)
+    sm_conf.pr = nerv.ParamRepo()
+    sm_conf.pr:import({sm_conf.fn_to_sample}, nil, sm_conf)
+
+    local layers = {
+        ["nerv.GRULayerT"] = {
+            ["gruL1"] = {{}, {["dim_in"] = {sm_conf.hidden_size, sm_conf.hidden_size}, ["dim_out"] = {sm_conf.hidden_size}, ["pr"] = sm_conf.pr}}, 
+        },
+        ["nerv.DropoutLayerT"] = {
+            ["dropoutL1"] = {{}, {["dim_in"] = {sm_conf.hidden_size}, ["dim_out"] = {sm_conf.hidden_size}}}, 
+        },
+        ["nerv.SelectLinearLayer"] = {
+            ["selectL1"] = {{}, {["dim_in"] = {1}, ["dim_out"] = {sm_conf.hidden_size}, ["vocab"] = sm_conf.vocab, ["pr"] = sm_conf.pr}},
+        },
+        ["nerv.CombinerLayer"] = {
+           ["combinerL1"] = {{}, {["dim_in"] = {sm_conf.hidden_size}, ["dim_out"] = {sm_conf.hidden_size, sm_conf.hidden_size}, ["lambda"] = {1}}},
+        },
+        ["nerv.AffineLayer"] = {
+            ["outputL"] = {{}, {["dim_in"] = {sm_conf.hidden_size}, ["dim_out"] = {sm_conf.vocab:size()},  ["pr"] = sm_conf.pr}},
+        },
+        ["nerv.SoftmaxCELayerT"] = {
+            ["softmaxL"] = {{}, {["dim_in"] = {sm_conf.vocab:size(), sm_conf.vocab:size()}, ["dim_out"] = {1}}},
+        },
+    }
+    local layerRepo = nerv.LayerRepo(layers, sm_conf.pr, sm_conf)
 
-    --input: input_w, input_w, ... input_w_now, last_activation
     local connections_t = {
         ["<input>[1]"] = "selectL1[1]",
        
@@ -151,48 +180,19 @@ function prepare_dagL(global_conf, layerRepo)
         ["combinerL1[2]"] = "<output>[2]",
     }
     
-    if global_conf.layer_num > 1 then
+    if sm_conf.layer_num > 1 then
         nerv.error("multiple layer is currently not supported(not hard to implement though)")
     end
-    --[[
-    for l = 2, global_conf.layer_num do
-        table.insert(connections_t, {"dropoutL"..(l-1).."[1]", "gruL"..l.."[1]", 0})
-        table.insert(connections_t, {"gruL"..l.."[1]", "combinerL"..l.."[1]", 0})
-        table.insert(connections_t, {"combinerL"..l.."[1]", "gruL"..l.."[2]", 1})
-        table.insert(connections_t, {"combinerL"..l.."[2]", "dropoutL"..l.."[1]", 0})
-    end
-    ]]--
 
-    --[[
-    printf("%s printing DAG connections:\n", global_conf.sche_log_pre)
-    for key, value in pairs(connections_t) do
-        printf("\t%s->%s\n", key, value)
-    end
-    ]]--
-
-    local dagL = nerv.DAGLayerT("dagL", global_conf, {["dim_in"] = {1, global_conf.hidden_size}, 
-            ["dim_out"] = {global_conf.vocab:size(), global_conf.hidden_size}, ["sub_layers"] = layerRepo,
+    local dagL = nerv.DAGLayerT("dagL", sm_conf, {["dim_in"] = {1, sm_conf.hidden_size}, 
+            ["dim_out"] = {sm_conf.vocab:size(), sm_conf.hidden_size}, ["sub_layers"] = layerRepo,
             ["connections"] = connections_t
         })
+    
+    local sampler = nerv.LMSampler(sm_conf)
+    sampler:load_dagL(dagL)
 
-    dagL:init(global_conf.batch_size)
-
-    nerv.printf("%s Initing DAGL end.\n", global_conf.sche_log_pre)
-    return dagL
-end
-
-function load_net_tnn(global_conf, fn)
-    prepare_parameters(global_conf, fn)
-    local layerRepo = prepare_layers(global_conf)
-    local tnn = prepare_tnn(global_conf, layerRepo)
-    return tnn
-end
-
-function load_net_dagL(global_conf, fn)
-    prepare_parameters(global_conf, fn)
-    local layerRepo = prepare_layers(global_conf)
-    local dagL = prepare_dagL(global_conf, layerRepo)
-    return dagL
+    return sampler
 end
 
 local train_fn, valid_fn, test_fn
@@ -240,6 +240,23 @@ global_conf = {
     fn_to_sample = root_dir .. '/ptb/EXP-nerv/grulm_v1.0h300l1ch15ba32slr0.15wc1e-05dr0.5/params.final',
 }
 
+sm_conf = {
+    cumat_type = nerv.CuMatrixFloat,
+    mmat_type = nerv.MMatrixFloat,
+    nn_act_default = 0, 
+
+    hidden_size = 300,
+    layer_num = 1,
+    batch_size = 32, 
+    chunk_size = 85, --largest sample sentence length
+    max_iter = 35,
+    max_sen_len = 90,
+    sche_log_pre = "[SAMPLER_S]:",
+
+    timer = global_conf.timer,
+    fn_to_sample = root_dir .. '/ptb/EXP-nerv/grulm_v1.0h300l1ch15ba32slr0.15wc1e-05dr0.5/params.final',
+}
+
 elseif (set == "msr_sc") then
 
 data_dir = '/home/slhome/txh18/workspace/sentenceCompletion/DATA_PV2'
@@ -276,15 +293,13 @@ global_conf = {
 
 elseif (set == "twitter") then
 
-data_dir = root_dir .. '/twitter_new/DATA'
-train_fn = data_dir .. '/twitter.choose2.adds'
-valid_fn = data_dir .. '/twitter.valid.adds'
-test_fn = data_dir .. '/comm.test.choose-ppl.adds'
-vocab_fn = data_dir .. '/twitter.choose.train.vocab'
-
---qdata_dir = root_dir .. '/ptb/questionGen/gen'
-
-global_conf = {
+    data_dir = root_dir .. '/twitter_new/DATA'
+    train_fn = data_dir .. '/twitter.choose2.adds'
+    valid_fn = data_dir .. '/twitter.valid.adds'
+    test_fn = data_dir .. '/comm.test.choose-ppl.adds'
+    vocab_fn = data_dir .. '/twitter.choose.train.vocab'
+    --qdata_dir = root_dir .. '/ptb/questionGen/gen'
+    global_conf = {
     lrate = 0.15, wcost = 1e-5, momentum = 0, clip_t = 5,
     cumat_type = nerv.CuMatrixFloat,
     mmat_type = nerv.MMatrixFloat,
@@ -309,7 +324,7 @@ global_conf = {
     log_w_num = 40000, --give a message when log_w_num words have been processed
     timer = nerv.Timer(),
     work_dir_base = root_dir .. '/twitter_new/EXP-nerv/grulm_v1.0'
-}
+    }
 
 else
 
@@ -385,8 +400,8 @@ nerv.LMUtil.wait(2)
 ]]--
 
 ----------------printing options---------------------------------
-nerv.printf("%s printing global_conf...\n", global_conf.sche_log_pre)
-for id, value in pairs(global_conf) do
+nerv.printf("%s printing sm_conf...\n", sm_conf.sche_log_pre)
+for id, value in pairs(sm_conf) do
     nerv.printf("%s:\t%s\n", id, tostring(value))
 end
 nerv.LMUtil.wait(2)
@@ -405,6 +420,7 @@ math.randomseed(1)
 
 local vocab = nerv.LMVocab()
 global_conf["vocab"] = vocab
+sm_conf["vocab"] = global_conf.vocab
 nerv.printf("%s building vocab...\n", global_conf.sche_log_pre)
 global_conf.vocab:build_file(global_conf.vocab_fn, false)
 ppl_rec = {} 
@@ -421,9 +437,7 @@ end --if commands["test"]
 if commands["sampling"] == 1 then
     nerv.printf("===SAMPLE===\n") 
     global_conf.sche_log_pre = "[SCHEDULER SAMPLING]:" 
-    local dagL = load_net_dagL(global_conf, global_conf.fn_to_sample) 
-    local sampler = nerv.LMSampler(global_conf)
-    sampler:load_dagL(dagL)
+    local sampler = prepare_sampler(sm_conf)
     for k = 1, 1 do
         local res = sampler:lm_sample_rnn_dagL(10, {})
         for i = 1, #res do