merge seq_buffer and change asr_trainer.lua accordingly

12 files changed, 80 insertions, 54 deletions

diff --git a/Makefile b/Makefile
index e98302b..728d14b 100644
--- a/Makefile
+++ b/Makefile
@@ -1,4 +1,4 @@
-.PHONY: all clean install luajit luarocks speech submodule
+.PHONY: all clean install luajit luarocks speech
 ##############  EDIT THESE LINES  #####################
 SHELL := /bin/bash
 PREFIX := $(CURDIR)/install/
@@ -26,8 +26,7 @@ export BLAS_LDFLAGS
 		nerv-clean speech-clean speech/speech_utils-clean speech/htk_io-clean speech/kaldi_io-clean speech/kaldi_decode-clean \
 		Penlight
 
-all: nerv
-submodule: luajit luarocks Penlight
+all: luajit luarocks Penlight nerv
 luajit:
 	PREFIX=$(PREFIX) ./tools/build_luajit.sh
 luarocks:
diff --git a/nerv/Makefile b/nerv/Makefile
index 68465a1..a5e4f66 100644
--- a/nerv/Makefile
+++ b/nerv/Makefile
@@ -44,7 +44,7 @@ LUA_LIBS := matrix/init.lua io/init.lua init.lua \
 			layer/elem_mul.lua layer/lstm.lua layer/lstm_gate.lua layer/dropout.lua layer/gru.lua \
 			layer/graph.lua layer/rnn.lua layer/duplicate.lua layer/identity.lua \
 			nn/init.lua nn/layer_repo.lua nn/param_repo.lua nn/network.lua \
-			io/sgd_buffer.lua io/seq_buffer.lua
+			io/frm_buffer.lua io/seq_buffer.lua
 
 INCLUDE := -I $(LUA_INCDIR) -DLUA_USE_APICHECK
 CUDA_INCLUDE := -I $(CUDA_BASE)/include/
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
diff --git a/nerv/init.lua b/nerv/init.lua
index ff944b8..439a83e 100644
--- a/nerv/init.lua
+++ b/nerv/init.lua
@@ -109,7 +109,7 @@ function table.val_to_str(v)
                     (("number" == type(v) or
                     "string" == type(v) or
                     "boolean" == type(v)) and tostring(v)) or
-                    nil -- failed to serialize
+                    "" -- failed to serialize
     end
 end
 
@@ -226,7 +226,8 @@ function nerv.parse_args(argv, options, unordered)
         local opt_type = v[3]
         local opt_meta = {type = opt_type,
                             desc = v.desc or "",
-                            val = v.default}
+                            val = v.default,
+                            specified = false}
         if opt_short ~= nil then
             if type(opt_short) ~= "string" or #opt_short ~= 1 then err() end
             if opt_type ~= "boolean" then
@@ -246,6 +247,7 @@ function nerv.parse_args(argv, options, unordered)
                 for c in k:gmatch"." do
                     if sopts[c] then
                         sopts[c].val = true
+                        sopts[c].specified = true
                     else
                         nerv.error("invalid option -%s", c)
                     end
@@ -262,6 +264,7 @@ function nerv.parse_args(argv, options, unordered)
                                     k, opts[k].type)
                     else
                         opts[k].val = true
+                        opts[k].specified = true
                     end
                 else
                     local k, v = token:match(opt_with_val_exp)
@@ -269,6 +272,7 @@ function nerv.parse_args(argv, options, unordered)
                         if opts[k] == nil then
                             nerv.error("invalid option %s", token)
                         end
+                        opts[k].specified = true
                         if opts[k].type == "boolean" then
                             if v == "yes" then
                                 opts[k].val = true
diff --git a/nerv/io/sgd_buffer.lua b/nerv/io/frm_buffer.lua
index d78f6d1..9761f16 100644
--- a/nerv/io/sgd_buffer.lua
+++ b/nerv/io/frm_buffer.lua
@@ -1,6 +1,6 @@
-local SGDBuffer = nerv.class("nerv.SGDBuffer", "nerv.DataBuffer")
+local FrmBuffer = nerv.class("nerv.FrmBuffer", "nerv.DataBuffer")
 
-function SGDBuffer:__init(global_conf, buffer_conf)
+function FrmBuffer:__init(global_conf, buffer_conf)
     self.gconf = global_conf
     self.batch_size = buffer_conf.batch_size
     self.buffer_size = math.floor(buffer_conf.buffer_size /
@@ -57,7 +57,7 @@ function SGDBuffer:__init(global_conf, buffer_conf)
     end
 end
 
-function SGDBuffer:saturate()
+function FrmBuffer:saturate()
     local buffer_size = self.buffer_size
     self.head = 0
     self.tail = buffer_size
@@ -116,7 +116,7 @@ function SGDBuffer:saturate()
     return self.tail >= self.batch_size
 end
 
-function SGDBuffer:get_data()
+function FrmBuffer:get_data()
     local batch_size = self.batch_size
     if self.head >= self.tail then -- buffer is empty
         local t = os.clock()
@@ -132,7 +132,9 @@ function SGDBuffer:get_data()
         return nil -- the remaining data cannot build a batch
     end
     actual_batch_size = math.min(batch_size, self.tail - self.head)
-    local res = {}
+    local res = {seq_length = table.vector(gconf.batch_size, 1),
+                new_seq = {},
+                data = {}}
     for i, reader in ipairs(self.readers) do
         for id, buff in pairs(reader.buffs) do
             local batch = self.output_mat_type(actual_batch_size, buff.width)
@@ -141,7 +143,7 @@ function SGDBuffer:get_data()
             else
                 self.copy_from(batch, buff.data, self.head, self.head + actual_batch_size)
             end
-            res[id] = batch
+            res.data[id] = {batch}
         end
     end
     self.head = self.head + actual_batch_size
diff --git a/nerv/io/init.lua b/nerv/io/init.lua
index c36d850..d3ba27c 100644
--- a/nerv/io/init.lua
+++ b/nerv/io/init.lua
@@ -56,5 +56,5 @@ function DataBuffer:get_data()
     nerv.error_method_not_implemented()
 end
 
-nerv.include('sgd_buffer.lua')
+nerv.include('frm_buffer.lua')
 nerv.include('seq_buffer.lua')
diff --git a/nerv/io/seq_buffer.lua b/nerv/io/seq_buffer.lua
index ad1b3f7..029e7b8 100644
--- a/nerv/io/seq_buffer.lua
+++ b/nerv/io/seq_buffer.lua
@@ -5,7 +5,10 @@ function SeqBuffer:__init(global_conf, buffer_conf)
 
     self.batch_size = buffer_conf.batch_size
     self.chunk_size = buffer_conf.chunk_size
-    self.readers = buffer_conf.readers
+    self.readers = {}
+    for _, v in ipairs(buffer_conf.readers) do
+        table.insert(self.readers, v.reader)
+    end
     self.nn_act_default = buffer_conf.nn_act_default
     if self.nn_act_default == nil then
         self.nn_act_default = 0
@@ -29,7 +32,7 @@ function SeqBuffer:new_mini_batch()
 end
 
 function SeqBuffer:saturate(batch)
-    if self.queue[self.head] ~= nil and self.queue[self.head].seq_length[batch] ~= 0 then    
+    if self.queue[self.head] ~= nil and self.queue[self.head].seq_length[batch] ~= 0 then
         return true
     end
     local data = {}
diff --git a/nerv/layer/rnn.lua b/nerv/layer/rnn.lua
index 0b5ccaa..333be9e 100644
--- a/nerv/layer/rnn.lua
+++ b/nerv/layer/rnn.lua
@@ -20,7 +20,7 @@ function RNNLayer:__init(id, global_conf, layer_conf)
         ['nerv.AffineLayer'] = {
             main = {dim_in = {din, dout}, dim_out = {dout}, pr = pr},
         },
-        [layers.activation] = {
+        [layer_conf.activation] = {
             activation = {dim_in = {dout}, dim_out = {dout}},
         },
         ['nerv.DuplicateLayer'] = {
diff --git a/nerv/nn/network.lua b/nerv/nn/network.lua
index 6f7fe10..7e2af4e 100644
--- a/nerv/nn/network.lua
+++ b/nerv/nn/network.lua
@@ -413,7 +413,7 @@ function network:set_input(input)
             local edge = self.socket.inputs[i]
             local id, port, time = edge[1], edge[2], edge[3]
             if t + time >= 1 and t + time <= self.chunk_size then
-                self.input[t + time][id][port] = input[t][i]
+                self.input[t + time][id][port] = input[i][t]
             end
         end
     end
@@ -425,7 +425,7 @@ function network:set_output(output)
             local edge = self.socket.outputs[i]
             local id, port, time = edge[1], edge[2], edge[3]
             if t - time >= 1 and t - time <= self.chunk_size then
-                self.output[t - time][id][port] = output[t][i]
+                self.output[t - time][id][port] = output[i][t]
             end
         end
     end
@@ -437,7 +437,7 @@ function network:set_err_input(err_input)
             local edge = self.socket.outputs[i]
             local id, port, time = edge[1], edge[2], edge[3]
             if t - time >= 1 and t - time <= self.chunk_size then
-                self.err_input[t - time][id][port] = err_input[t][i]
+                self.err_input[t - time][id][port] = err_input[i][t]
             end
         end
     end
@@ -449,7 +449,7 @@ function network:set_err_output(err_output)
             local edge = self.socket.inputs[i]
             local id, port, time = edge[1], edge[2], edge[3]
             if t + time >= 1 and t + time <= self.chunk_size then
-                self.err_output[t + time][id][port] = err_output[t][i]
+                self.err_output[t + time][id][port] = err_output[i][t]
             end
         end
     end