add ASR DNN trainer

2 files changed, 250 insertions, 0 deletions

diff --git a/examples/asr_trainer.lua b/examples/asr_trainer.lua
new file mode 100644
index 0000000..b43a547
--- /dev/null
+++ b/examples/asr_trainer.lua
@@ -0,0 +1,87 @@
+function build_trainer(ifname)
+    local param_repo = make_param_repo(ifname)
+    local sublayer_repo = make_sublayer_repo(param_repo)
+    local layer_repo = make_layer_repo(sublayer_repo, param_repo)
+    local crit = get_criterion_layer(sublayer_repo)
+    local network = get_network(layer_repo)
+    local iterative_trainer = function (ofname, scp_file, bp)
+        gconf.randomize = bp
+        -- build buffer
+        local buffer = make_buffer(make_reader(scp_file, layer_repo))
+        -- initialize the network
+        network:init(gconf.batch_size)
+        gconf.cnt = 0
+        for data in buffer.get_data, buffer do
+            -- prine stat periodically
+            gconf.cnt = gconf.cnt + 1
+            if gconf.cnt == 1000 then
+                print_stat(crit)
+                gconf.cnt = 0
+            end
+            if gconf.cnt == 100 then break end
+
+            input = {data.main_scp, data.phone_state}
+            output = {}
+            err_input = {}
+            err_output = {input[1]:create()}
+            network:propagate(input, output)
+            if bp then
+                network:back_propagate(err_output, err_input, input, output)
+                network:update(err_input, input, output)
+            end
+            -- collect garbage in-time to save GPU memory
+            collectgarbage("collect")
+        end
+        print_stat(crit)
+        if bp then
+            nerv.info("writing back...")
+            cf = nerv.ChunkFile(ofname, "w")
+            for i, p in ipairs(network:get_params()) do
+                cf:write_chunk(p)
+            end
+            cf:close()
+        end
+        return get_accuracy(crit)
+    end
+    return iterative_trainer
+end
+
+dofile(arg[1])
+start_halving_inc = 0.5
+halving_factor = 0.6
+end_halving_inc = 0.1
+min_iter = 1
+max_iter = 20
+min_halving = 6
+gconf.batch_size = 256
+gconf.buffer_size = 81920
+
+local pf0 = gconf.initialized_param
+local trainer = build_trainer(pf0)
+--local trainer = build_trainer("c3.nerv")
+local accu_best = trainer(nil, gconf.cv_scp, false)
+local do_halving = false
+
+nerv.info("initial cross validation: %.3f", accu_best)
+for i = 1, max_iter do
+    nerv.info("iteration %d with lrate = %.6f", i, gconf.lrate)
+    local accu_tr = trainer(pf0 .. "_iter" .. i .. ".nerv", gconf.tr_scp, true)
+    nerv.info("[TR] training set %d: %.3f", i, accu_tr)
+    local accu_new = trainer(nil, gconf.cv_scp, false)
+    nerv.info("[CV] cross validation %d: %.3f", i, accu_new)
+    -- TODO: revert the weights
+    local accu_diff = accu_new - accu_best
+    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
+    end
+    if accu_new > accu_best then
+        accu_best = accu_new
+    end
+end
+nerv.Matrix.print_profile()
diff --git a/examples/swb_baseline.lua b/examples/swb_baseline.lua
new file mode 100644
index 0000000..f536777
--- /dev/null
+++ b/examples/swb_baseline.lua
@@ -0,0 +1,163 @@
+require 'speech.init'
+gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9,
+        cumat_type = nerv.CuMatrixFloat,
+        mmat_type = nerv.MMatrixFloat,
+        frm_ext = 5,
+        tr_scp = "/slfs1/users/mfy43/swb_ivec/train_bp.scp",
+        cv_scp = "/slfs1/users/mfy43/swb_ivec/train_cv.scp",
+        htk_conf = "/slfs1/users/mfy43/swb_ivec/plp_0_d_a.conf",
+        global_transf = "global_transf.nerv",
+        initialized_param = "converted.nerv",
+        debug = false}
+
+function make_param_repo(param_file)
+    return nerv.ParamRepo({param_file, gconf.global_transf})
+end
+
+function make_sublayer_repo(param_repo)
+    return nerv.LayerRepo(
+    {
+        -- global transf
+        ["nerv.BiasLayer"] =
+        {
+            blayer1 = {{bias = "bias1"}, {dim_in = {429}, dim_out = {429}}},
+            blayer2 = {{bias = "bias2"}, {dim_in = {429}, dim_out = {429}}}
+        },
+        ["nerv.WindowLayer"] =
+        {
+            wlayer1 = {{window = "window1"}, {dim_in = {429}, dim_out = {429}}},
+            wlayer2 = {{window = "window2"}, {dim_in = {429}, dim_out = {429}}}
+        },
+        -- biased linearity
+        ["nerv.AffineLayer"] =
+        {
+            affine0 = {{ltp = "affine0_ltp", bp = "affine0_bp"},
+            {dim_in = {429}, dim_out = {2048}}},
+            affine1 = {{ltp = "affine1_ltp", bp = "affine1_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine2 = {{ltp = "affine2_ltp", bp = "affine2_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine3 = {{ltp = "affine3_ltp", bp = "affine3_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine4 = {{ltp = "affine4_ltp", bp = "affine4_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine5 = {{ltp = "affine5_ltp", bp = "affine5_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine6 = {{ltp = "affine6_ltp", bp = "affine6_bp"},
+            {dim_in = {2048}, dim_out = {2048}}},
+            affine7 = {{ltp = "affine7_ltp", bp = "affine7_bp"},
+            {dim_in = {2048}, dim_out = {3001}}}
+        },
+        ["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"] =
+        {
+            criterion = {{}, {dim_in = {3001, 1}, dim_out = {}, compressed = true}}
+        }
+    }, param_repo, gconf)
+end
+
+function make_layer_repo(sublayer_repo, param_repo)
+    return nerv.LayerRepo(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            global_transf = {{}, {
+                dim_in = {429}, dim_out = {429},
+                sub_layers = sublayer_repo,
+                connections = {
+                    ["<input>[1]"] = "blayer1[1]",
+                    ["blayer1[1]"] = "wlayer1[1]",
+                    ["wlayer1[1]"] = "blayer2[1]",
+                    ["blayer2[1]"] = "wlayer2[1]",
+                    ["wlayer2[1]"] = "<output>[1]"
+                }
+            }},
+            main = {{}, {
+                dim_in = {429, 1}, dim_out = {},
+                sub_layers = sublayer_repo,
+                connections = {
+                    ["<input>[1]"] = "affine0[1]",
+                    ["affine0[1]"] = "sigmoid0[1]",
+                    ["sigmoid0[1]"] = "affine1[1]",
+                    ["affine1[1]"] = "sigmoid1[1]",
+                    ["sigmoid1[1]"] = "affine2[1]",
+                    ["affine2[1]"] = "sigmoid2[1]",
+                    ["sigmoid2[1]"] = "affine3[1]",
+                    ["affine3[1]"] = "sigmoid3[1]",
+                    ["sigmoid3[1]"] = "affine4[1]",
+                    ["affine4[1]"] = "sigmoid4[1]",
+                    ["sigmoid4[1]"] = "affine5[1]",
+                    ["affine5[1]"] = "sigmoid5[1]",
+                    ["sigmoid5[1]"] = "affine6[1]",
+                    ["affine6[1]"] = "sigmoid6[1]",
+                    ["sigmoid6[1]"] = "affine7[1]",
+                    ["affine7[1]"] = "criterion[1]",
+                    ["<input>[2]"] = "criterion[2]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+end
+
+function get_criterion_layer(sublayer_repo)
+    return sublayer_repo:get_layer("criterion")
+end
+
+function get_network(layer_repo)
+    return layer_repo:get_layer("main")
+end
+
+function make_reader(scp_file, layer_repo)
+    return nerv.TNetReader(gconf,
+        {
+            id = "main_scp",
+            scp_file = scp_file,
+            conf_file = gconf.htk_conf,
+            frm_ext = gconf.frm_ext,
+            mlfs = {
+                phone_state = {
+                    file = "/slfs1/users/mfy43/swb_ivec/ref.mlf",
+                    format = "map",
+                    format_arg = "/slfs1/users/mfy43/swb_ivec/dict",
+                    dir = "*/",
+                    ext = "lab"
+                }
+            },
+            global_transf = layer_repo:get_layer("global_transf")
+        })
+end
+
+function make_buffer(reader, buffer)
+    return nerv.SGDBuffer(gconf,
+        {
+            buffer_size = gconf.buffer_size,
+            randomize = gconf.randomize,
+            readers = {
+                { reader = reader,
+                  data = {main_scp = 429, phone_state = 1}}
+            }
+        })
+end
+
+function get_accuracy(crit)
+    return crit.total_correct / crit.total_frames * 100
+end
+
+function print_stat(crit)
+    nerv.info("*** training stat begin ***")
+    nerv.utils.printf("cross entropy:\t%.8f\n", crit.total_ce)
+    nerv.utils.printf("correct:\t%d\n", crit.total_correct)
+    nerv.utils.printf("frames:\t%d\n", crit.total_frames)
+    nerv.utils.printf("err/frm:\t%.8f\n", crit.total_ce / crit.total_frames)
+    nerv.utils.printf("accuracy:\t%.3f%%\n", get_accuracy(crit))
+    nerv.info("*** training stat end ***")
+end