require 'htk_io'
gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9, frm_ext = 5,
rearrange = true, -- just to make the context order consistent with old results, deprecated
frm_trim = 5, -- trim the first and last 5 frames, TNet just does this, deprecated
tr_scp = "/speechlab/users/mfy43/swb50/train_bp.scp",
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"},
chunk_size = 1}
function make_layer_repo(param_repo)
local layer_repo = nerv.LayerRepo(
{
-- global transf
["nerv.BiasLayer"] =
{
blayer1 = {dim_in = {429}, dim_out = {429}, params = {bias = "bias0"}},
blayer2 = {dim_in = {429}, dim_out = {429}, params = {bias = "bias1"}}
},
["nerv.WindowLayer"] =
{
wlayer1 = {dim_in = {429}, dim_out = {429}, params = {window = "window0"}},
wlayer2 = {dim_in = {429}, dim_out = {429}, params = {window = "window1"}}
},
-- biased linearity
["nerv.AffineLayer"] =
{
affine0 = {dim_in = {429}, dim_out = {2048},
params = {ltp = "affine0_ltp", bp = "affine0_bp"}},
affine1 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine1_ltp", bp = "affine1_bp"}},
affine2 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine2_ltp", bp = "affine2_bp"}},
affine3 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine3_ltp", bp = "affine3_bp"}},
affine4 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine4_ltp", bp = "affine4_bp"}},
affine5 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine5_ltp", bp = "affine5_bp"}},
affine6 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine6_ltp", bp = "affine6_bp"}},
affine7 = {dim_in = {2048}, dim_out = {3001},
params = {ltp = "affine7_ltp", bp = "affine7_bp"}}
},
["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"] = -- softmax + ce criterion layer for finetune output
{
ce_crit = {dim_in = {3001, 1}, dim_out = {1}, compressed = true}
},
["nerv.SoftmaxLayer"] = -- softmax for decode output
{
softmax = {dim_in = {3001}, dim_out = {3001}}
}
}, param_repo, gconf)
layer_repo:add_layers(
{
["nerv.GraphLayer"] =
{
global_transf = {
dim_in = {429}, dim_out = {429},
layer_repo = layer_repo,
connections = {
{"<input>[1]", "blayer1[1]", 0},
{"blayer1[1]", "wlayer1[1]", 0},
{"wlayer1[1]", "blayer2[1]", 0},
{"blayer2[1]", "wlayer2[1]", 0},
{"wlayer2[1]", "<output>[1]", 0}
}
},
main = {
dim_in = {429}, dim_out = {3001},
layer_repo = layer_repo,
connections = {
{"<input>[1]", "affine0[1]", 0},
{"affine0[1]", "sigmoid0[1]", 0},
{"sigmoid0[1]", "affine1[1]", 0},
{"affine1[1]", "sigmoid1[1]", 0},
{"sigmoid1[1]", "affine2[1]", 0},
{"affine2[1]", "sigmoid2[1]", 0},
{"sigmoid2[1]", "affine3[1]", 0},
{"affine3[1]", "sigmoid3[1]", 0},
{"sigmoid3[1]", "affine4[1]", 0},
{"affine4[1]", "sigmoid4[1]", 0},
{"sigmoid4[1]", "affine5[1]", 0},
{"affine5[1]", "sigmoid5[1]", 0},
{"sigmoid5[1]", "affine6[1]", 0},
{"affine6[1]", "sigmoid6[1]", 0},
{"sigmoid6[1]", "affine7[1]", 0},
{"affine7[1]", "<output>[1]", 0}
}
}
}
}, param_repo, gconf)
layer_repo:add_layers(
{
["nerv.GraphLayer"] =
{
ce_output = {
dim_in = {429, 1}, dim_out = {1},
layer_repo = layer_repo,
connections = {
{"<input>[1]", "main[1]", 0},
{"main[1]", "ce_crit[1]", 0},
{"<input>[2]", "ce_crit[2]", 0},
{"ce_crit[1]", "<output>[1]", 0}
}
},
softmax_output = {
dim_in = {429}, dim_out = {3001},
layer_repo = layer_repo,
connections = {
{"<input>[1]", "main[1]", 0},
{"main[1]", "softmax[1]", 0},
{"softmax[1]", "<output>[1]", 0}
}
}
}
}, param_repo, gconf)
return layer_repo
end
function get_network(layer_repo)
return layer_repo:get_layer("ce_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(scp_file, layer_repo)
return {
{reader = nerv.HTKReader(gconf,
{
id = "main_scp",
scp_file = scp_file,
conf_file = gconf.htk_conf,
frm_ext = gconf.frm_ext,
mlfs = {
phone_state = {
file = "/speechlab/users/mfy43/swb50/ref.mlf",
format = "map",
format_arg = "/speechlab/users/mfy43/swb50/dict",
dir = "*/",
ext = "lab"
}
}
}),
data = {main_scp = 429, phone_state = 1}}
}
end
function make_buffer(readers)
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
})
end
function get_input_order()
return {{id = "main_scp", global_transf = true},
{id = "phone_state"}}
end
function get_decode_input_order()
return {{id = "main_scp", global_transf = true}}
end
function get_accuracy(layer_repo)
local ce_crit = layer_repo:get_layer("ce_crit")
return ce_crit.total_correct / ce_crit.total_frames * 100
end
function print_stat(layer_repo)
local ce_crit = layer_repo:get_layer("ce_crit")
nerv.info("*** training stat begin ***")
nerv.printf("cross entropy:\t\t%.8f\n", ce_crit.total_ce)
nerv.printf("correct:\t\t%d\n", ce_crit.total_correct)
nerv.printf("frames:\t\t\t%d\n", ce_crit.total_frames)
nerv.printf("err/frm:\t\t%.8f\n", ce_crit.total_ce / ce_crit.total_frames)
nerv.printf("accuracy:\t\t%.3f%%\n", get_accuracy(layer_repo))
nerv.info("*** training stat end ***")
end