merge seq_buffer and change asr_trainer.lua accordingly

4 files changed, 52 insertions, 34 deletions

diff --git a/nerv/examples/asr_trainer.lua b/nerv/examples/asr_trainer.lua
index 6bdf57c..645f1ef 100644
--- a/nerv/examples/asr_trainer.lua
+++ b/nerv/examples/asr_trainer.lua
@@ -22,9 +22,9 @@ local function build_trainer(ifname)
     local input_order = get_input_order()
 
     network = nerv.Network("nt", gconf, {network = network})
-    network:init(gconf.batch_size, 1)
+    network:init(gconf.batch_size, gconf.chunk_size)
     global_transf = nerv.Network("gt", gconf, {network = global_transf})
-    global_transf:init(gconf.batch_size, 1)
+    global_transf:init(gconf.batch_size, gconf.chunk_size)
 
     local iterative_trainer = function (prefix, scp_file, bp, rebind_param_repo)
         -- rebind the params if necessary
@@ -39,11 +39,17 @@ local function build_trainer(ifname)
         local buffer = make_buffer(make_readers(scp_file, layer_repo))
         -- initialize the network
         gconf.cnt = 0
-        err_input = {mat_type(gconf.batch_size, 1)}
-        err_input[1]:fill(1)
+        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()
         global_transf:epoch_init()
-        for data in buffer.get_data, buffer do
+        for d in buffer.get_data, buffer do
             -- prine stat periodically
             gconf.cnt = gconf.cnt + 1
             if gconf.cnt == 1000 then
@@ -54,35 +60,39 @@ local function build_trainer(ifname)
                 -- break
             end
             local input = {}
+            local err_output = {}
 --            if gconf.cnt == 1000 then break end
             for i, e in ipairs(input_order) do
                 local id = e.id
-                if data[id] == nil then
+                if d.data[id] == nil then
                     nerv.error("input data %s not found", id)
                 end
-                local transformed
+                local transformed = {}
+                local err_output_i = {}
                 if e.global_transf then
-                    transformed = nerv.speech_utils.global_transf(data[id],
-                                        global_transf,
-                                        gconf.frm_ext or 0, 0,
-                                        gconf)
+                    for _, mini_batch in ipairs(d.data[id]) do
+                        table.insert(transformed,
+                                        nerv.speech_utils.global_transf(mini_batch,
+                                            global_transf,
+                                            gconf.frm_ext or 0, 0,
+                                            gconf))
+                    end
                 else
-                    transformed = data[id]
+                    transformed = d.data[id]
+                end
+                for _, mini_batch in ipairs(transformed) do
+                    table.insert(err_output_i, mini_batch:create())
                 end
+                table.insert(err_output, err_output_i)
                 table.insert(input, transformed)
             end
-            local output = {mat_type(gconf.batch_size, 1)}
-            err_output = {}
-            for i = 1, #input do
-                table.insert(err_output, input[i]:create())
-            end
-            network:mini_batch_init({seq_length = table.vector(gconf.batch_size, 1),
-                                    new_seq = {},
+            network:mini_batch_init({seq_length = d.seq_length,
+                                    new_seq = d.new_seq,
                                     do_train = bp,
-                                    input = {input},
-                                    output = {output},
-                                    err_input = {err_input},
-                                    err_output = {err_output}})
+                                    input = input,
+                                    output = output,
+                                    err_input = err_input,
+                                    err_output = err_output})
             network:propagate()
             if bp then
                 network:back_propagate()
@@ -111,19 +121,21 @@ end
 
 local function check_and_add_defaults(spec, opts)
     local function get_opt_val(k)
-        return opts[string.gsub(k, '_', '-')].val
+        local k = string.gsub(k, '_', '-')
+        return opts[k].val, opts[k].specified
     end
     local opt_v = get_opt_val("resume_from")
     if opt_v then
+        nerv.info("resuming from previous training state")
         gconf = dofile(opt_v)
     else
         for k, v in pairs(spec) do
-            local opt_v = get_opt_val(k)
-            if opt_v ~= nil then
+            local opt_v, specified = get_opt_val(k)
+            if (not specified) and gconf[k] ~= nil then
+                nerv.info("using setting in network config file: %s = %s", k, gconf[k])
+            elseif opt_v ~= nil then
+                nerv.info("using setting in options: %s = %s", k, opt_v)
                 gconf[k] = opt_v
-            elseif gconf[k] ~= nil then
-            elseif v ~= nil then
-                gconf[k] = v
             end
         end
     end
@@ -168,6 +180,7 @@ end
 local trainer_defaults = {
     lrate = 0.8,
     batch_size = 256,
+    chunk_size = 1,
     buffer_size = 81920,
     wcost = 1e-6,
     momentum = 0.9,
diff --git a/nerv/examples/swb_baseline.lua b/nerv/examples/swb_baseline.lua
index 0ce8468..ece4d44 100644
--- a/nerv/examples/swb_baseline.lua
+++ b/nerv/examples/swb_baseline.lua
@@ -171,6 +171,7 @@ function make_buffer(readers)
         {
             buffer_size = gconf.buffer_size,
             batch_size = gconf.batch_size,
+            chunk_size = gconf.chunk_size,
             randomize = gconf.randomize,
             readers = readers,
             use_gpu = true
diff --git a/nerv/examples/swb_baseline2.lua b/nerv/examples/swb_baseline2.lua
index 6796f6f..38cfb9a 100644
--- a/nerv/examples/swb_baseline2.lua
+++ b/nerv/examples/swb_baseline2.lua
@@ -6,7 +6,8 @@ gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9, frm_ext = 5,
         cv_scp = "/speechlab/users/mfy43/swb50/train_cv.scp",
         htk_conf = "/speechlab/users/mfy43/swb50/plp_0_d_a.conf",
         initialized_param = {"/speechlab/users/mfy43/swb50/swb_init.nerv",
-                            "/speechlab/users/mfy43/swb50/swb_global_transf.nerv"}}
+                            "/speechlab/users/mfy43/swb50/swb_global_transf.nerv"},
+        chunk_size = 1}
 
 function make_layer_repo(param_repo)
     local layer_repo = nerv.LayerRepo(
@@ -145,7 +146,7 @@ end
 
 function make_readers(scp_file, layer_repo)
     return {
-                {reader = nerv.TNetReader(gconf,
+                {reader = nerv.HTKReader(gconf,
                     {
                         id = "main_scp",
                         scp_file = scp_file,
@@ -166,10 +167,11 @@ function make_readers(scp_file, layer_repo)
 end
 
 function make_buffer(readers)
-    return nerv.SGDBuffer(gconf,
+    return nerv.FrmBuffer(gconf,
         {
             buffer_size = gconf.buffer_size,
             batch_size = gconf.batch_size,
+            chunk_size = gconf.chunk_size,
             randomize = gconf.randomize,
             readers = readers,
             use_gpu = true
diff --git a/nerv/examples/timit_baseline2.lua b/nerv/examples/timit_baseline2.lua
index b1c1e66..658aa2e 100644
--- a/nerv/examples/timit_baseline2.lua
+++ b/nerv/examples/timit_baseline2.lua
@@ -8,7 +8,8 @@ gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9, frm_ext = 5,
                             "/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_dnn/nnet_output.nerv",
                             "/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_dnn/nnet_trans.nerv"},
          -- params in nnet_trans.nerv are included in the trained model
-        decode_param = {"/speechlab/users/mfy43/timit/s5/nerv_20160311205342/nnet_init_20160311211609_iter_13_lr0.013437_tr72.572_cv58.709.nerv"}}
+        decode_param = {"/speechlab/users/mfy43/timit/s5/nerv_20160311205342/nnet_init_20160311211609_iter_13_lr0.013437_tr72.572_cv58.709.nerv"},
+        chunk_size = 1}
 
 function make_layer_repo(param_repo)
     local layer_repo = nerv.LayerRepo(
@@ -176,10 +177,11 @@ function make_decode_readers(scp_file, layer_repo)
 end
 
 function make_buffer(readers)
-    return nerv.SGDBuffer(gconf,
+    return nerv.FrmBuffer(gconf,
         {
             buffer_size = gconf.buffer_size,
             batch_size = gconf.batch_size,
+            chunk_size = gconf.chunk_size,
             randomize = gconf.randomize,
             readers = readers,
             use_gpu = true