From 0bb9cd4271f127c311fd9839855def8f9ea91dab Mon Sep 17 00:00:00 2001
From: Determinant <ted.sybil@gmail.com>
Date: Sat, 6 Jun 2015 11:02:48 +0800
Subject: add ASR DNN trainer

---
 examples/asr_trainer.lua | 87 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 87 insertions(+)
 create mode 100644 examples/asr_trainer.lua

(limited to 'examples/asr_trainer.lua')

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()
-- 
cgit v1.2.3