kaldi mpe training support

1 files changed, 185 insertions, 0 deletions

diff --git a/nerv/examples/seq_chime.lua b/nerv/examples/seq_chime.lua
new file mode 100644
index 0000000..be723ca
--- /dev/null
+++ b/nerv/examples/seq_chime.lua
@@ -0,0 +1,185 @@
+require 'kaldi_io'
+gconf = {lrate = 0.00001, wcost = 0, momentum = 0.0,
+        cumat_type = nerv.CuMatrixFloat,
+        mmat_type = nerv.MMatrixFloat,
+        frm_ext = 5,
+        tr_scp = "ark,s,cs:/slfs6/users/ymz09/kaldi/src/featbin/copy-feats scp:/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_smbr/train.scp ark:- |",
+        initialized_param = {"/slfs6/users/ymz09/nerv-project/nerv/nerv-speech/kaldi_seq/test/chime3_init.nerv",
+        "/slfs6/users/ymz09/nerv-project/nerv/nerv-speech/kaldi_seq/test/chime3_global_transf.nerv"},
+        debug = false}
+
+function make_layer_repo(param_repo)
+    local layer_repo = nerv.LayerRepo(
+    {
+        -- global transf
+        ["nerv.BiasLayer"] =
+        {
+            blayer1 = {{bias = "bias1"}, {dim_in = {440}, dim_out = {440}}},
+            blayer2 = {{bias = "bias2"}, {dim_in = {440}, dim_out = {440}}}
+        },
+        ["nerv.WindowLayer"] =
+        {
+            wlayer1 = {{window = "window1"}, {dim_in = {440}, dim_out = {440}}},
+            wlayer2 = {{window = "window2"}, {dim_in = {440}, dim_out = {440}}}
+        },
+        -- biased linearity
+        ["nerv.AffineLayer"] =
+        {
+            affine0 = {{ltp = "affine0_ltp", bp = "affine0_bp"},
+            {dim_in = {440}, 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 = {2011}}}
+        },
+        ["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.MPELayer"] =
+        {
+            mpe_crit = {{}, {dim_in = {2011, -1}, dim_out = {1},
+                        cmd = {
+                            arg = "--class-frame-counts=/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced/ali_train_pdf.counts --acoustic-scale=0.1 --lm-scale=1.0 --learn-rate=0.00001 --do-smbr=true --verbose=1",
+                            mdl = "/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_ali/final.mdl",
+                            lat = "scp:/slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_denlats/lat.scp",
+                            ali = "ark:gunzip -c /slfs5/users/ymz09/chime/baseline/ASR/exp/tri4a_dnn_tr05_multi_enhanced_ali/ali.*.gz |"
+                        }
+                    }
+                }
+        },
+        ["nerv.SoftmaxLayer"] = -- softmax for decode output
+        {
+            softmax = {{}, {dim_in = {2011}, dim_out = {2011}}}
+        }
+    }, param_repo, gconf)
+
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            global_transf = {{}, {
+                dim_in = {440}, dim_out = {440},
+                sub_layers = layer_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 = {440}, dim_out = {2011},
+                sub_layers = layer_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]"] = "<output>[1]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    layer_repo:add_layers(
+    {
+        ["nerv.DAGLayer"] =
+        {
+            mpe_output = {{}, {
+                dim_in = {440, -1}, dim_out = {1},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "mpe_crit[1]",
+                    ["<input>[2]"] = "mpe_crit[2]",
+                    ["mpe_crit[1]"] = "<output>[1]"
+                }
+            }},
+            softmax_output = {{}, {
+                dim_in = {440}, dim_out = {2011},
+                sub_layers = layer_repo,
+                connections = {
+                    ["<input>[1]"] = "main[1]",
+                    ["main[1]"] = "softmax[1]",
+                    ["softmax[1]"] = "<output>[1]"
+                }
+            }}
+        }
+    }, param_repo, gconf)
+
+    return layer_repo
+end
+
+function get_network(layer_repo)
+    return layer_repo:get_layer("mpe_output")
+end
+
+function get_decode_network(layer_repo)
+    return layer_repo:get_layer("softmax_output")
+end
+
+function get_global_transf(layer_repo)
+    return layer_repo:get_layer("global_transf")
+end
+
+function make_readers(feature_rspecifier, layer_repo)
+    return {
+                {reader = nerv.KaldiReader(gconf,
+                    {
+                        id = "main_scp",
+                        feature_rspecifier = feature_rspecifier,
+                        frm_ext = gconf.frm_ext,
+                        global_transf = layer_repo:get_layer("global_transf"),
+                        mlfs = {}
+                    })
+                }
+            }
+end
+
+function get_input_order()
+    return {{id = "main_scp", global_transf = true},
+            {id = "key"}}
+end
+
+function get_accuracy(layer_repo)
+    local mpe_crit = layer_repo:get_layer("mpe_crit")
+    return mpe_crit.total_correct / mpe_crit.total_frames * 100
+end
+
+function print_stat(layer_repo)
+    local mpe_crit = layer_repo:get_layer("mpe_crit")
+    nerv.info("*** training stat begin ***")
+    nerv.printf("correct:\t\t%d\n", mpe_crit.total_correct)
+    nerv.printf("frames:\t\t\t%d\n", mpe_crit.total_frames)
+    nerv.printf("accuracy:\t\t%.3f%%\n", get_accuracy(layer_repo))
+    nerv.info("*** training stat end ***")
+end