require 'kaldi_io'
gconf = {lrate = 0.8,
wcost = 1e-6,
momentum = 0.9,
frm_ext = 5,
chunk_size = 1,
tr_scp = "ark:/speechlab/tools/KALDI/kaldi-master/src/featbin/copy-feats " ..
"scp:/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_dnn/train.scp ark:- |",
cv_scp = "ark:/speechlab/tools/KALDI/kaldi-master/src/featbin/copy-feats " ..
"scp:/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_dnn/cv.scp ark:- |",
ali = {targets_rspecifier = "ark:/speechlab/tools/KALDI/kaldi-master/src/bin/ali-to-pdf " ..
"/speechlab/users/mfy43/timit/s5/exp/tri3_ali/final.mdl " ..
"\"ark:gunzip -c /speechlab/users/mfy43/timit/s5/exp/tri3_ali/ali.*.gz |\" " ..
"ark:- | " ..
"/speechlab/tools/KALDI/kaldi-master/src/bin/ali-to-post " ..
"ark:- ark:- |"},
initialized_param = {"/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_dnn/nnet_init.nerv",
"/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_2016-05-06_17:40:54/2016-05-06_19:44:43_iter_20_lr0.012500_tr0.867_cv1.464.nerv"}
}
local input_size = 440
local output_size = 1959
local hidden_size = 1024
local trainer = nerv.Trainer
function trainer:make_layer_repo(param_repo)
local layer_repo = nerv.LayerRepo(
{
-- global transf
["nerv.BiasLayer"] =
{
blayer1 = {dim_in = {input_size}, dim_out = {input_size}, params = {bias = "bias0"}, no_update_all = true}
},
["nerv.WindowLayer"] =
{
wlayer1 = {dim_in = {input_size}, dim_out = {input_size}, params = {window = "window0"}, no_update_all = true}
},
-- biased linearity
["nerv.AffineLayer"] =
{
affine0 = {dim_in = {input_size}, dim_out = {hidden_size},
params = {ltp = "affine0_ltp", bp = "affine0_bp"}},
affine1 = {dim_in = {hidden_size}, dim_out = {hidden_size},
params = {ltp = "affine1_ltp", bp = "affine1_bp"}},
affine2 = {dim_in = {hidden_size}, dim_out = {hidden_size},
params = {ltp = "affine2_ltp", bp = "affine2_bp"}},
affine3 = {dim_in = {hidden_size}, dim_out = {hidden_size},
params = {ltp = "affine3_ltp", bp = "affine3_bp"}},
affine4 = {dim_in = {hidden_size}, dim_out = {hidden_size},
params = {ltp = "affine4_ltp", bp = "affine4_bp"}},
affine5 = {dim_in = {hidden_size}, dim_out = {hidden_size},
params = {ltp = "affine5_ltp", bp = "affine5_bp"}},
affine6 = {dim_in = {hidden_size}, dim_out = {output_size},
params = {ltp = "affine6_ltp", bp = "affine6_bp"}}
},
["nerv.SigmoidLayer"] =
{
sigmoid0 = {dim_in = {hidden_size}, dim_out = {hidden_size}},
sigmoid1 = {dim_in = {hidden_size}, dim_out = {hidden_size}},
sigmoid2 = {dim_in = {hidden_size}, dim_out = {hidden_size}},
sigmoid3 = {dim_in = {hidden_size}, dim_out = {hidden_size}},
sigmoid4 = {dim_in = {hidden_size}, dim_out = {hidden_size}},
sigmoid5 = {dim_in = {hidden_size}, dim_out = {hidden_size}}
},
["nerv.SoftmaxCELayer"] = -- softmax + ce criterion layer for finetune output
{
ce_crit = {dim_in = {output_size, 1}, dim_out = {1}, compressed = true}
},
["nerv.SoftmaxLayer"] = -- softmax for decode output
{
softmax = {dim_in = {output_size}, dim_out = {output_size}}
}
}, param_repo, gconf)
layer_repo:add_layers(
{
["nerv.GraphLayer"] =
{
global_transf = {
dim_in = {input_size}, dim_out = {input_size},
layer_repo = layer_repo,
connections = {
{"[1]", "blayer1[1]", 0},
{"blayer1[1]", "wlayer1[1]", 0},
{"wlayer1[1]", "