require 'htk_io'
gconf = {lrate = 0.2, wcost = 1e-6, momentum = 0.9,
cumat_type = nerv.CuMatrixFloat,
mmat_type = nerv.MMatrixFloat,
frm_ext = 5,
frm_trim = 5,
batch_size = 256,
buffer_size = 81920,
rearrange = true,
tr_scp = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/train.scp",
cv_scp = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/train_cv.scp",
htk_conf = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/fbank_d_a_z.conf",
initialized_param = {"/sgfs/users/wd007/src/nerv/tools/nerv.global.transf",
"/sgfs/users/wd007/src/nerv/tools/nerv.svd0.55_3000h_iter1.init"},
debug = false}
function make_layer_repo(param_repo)
local layer_repo = nerv.LayerRepo(
{
-- global transf
["nerv.BiasLayer"] =
{
blayer1 = {{bias = "bias1"}, {dim_in = {1320}, dim_out = {1320}}},
},
["nerv.WindowLayer"] =
{
wlayer1 = {{window = "window1"}, {dim_in = {1320}, dim_out = {1320}}},
},
-- biased linearity
["nerv.AffineLayer"] =
{
affine0 = {{ltp = "affine0_ltp", bp = "affine0_bp"},
{dim_in = {1320}, dim_out = {2048}}},
affine1 = {{ltp = "affine1_ltp", bp = "affine1_bp"},
{dim_in = {2048}, dim_out = {367}}},
affine2 = {{ltp = "affine2_ltp", bp = "affine2_bp"},
{dim_in = {367}, dim_out = {2048}}},
affine3 = {{ltp = "affine3_ltp", bp = "affine3_bp"},
{dim_in = {2048}, dim_out = {408}}},
affine4 = {{ltp = "affine4_ltp", bp = "affine4_bp"},
{dim_in = {408}, dim_out = {2048}}},
affine5 = {{ltp = "affine5_ltp", bp = "affine5_bp"},
{dim_in = {2048}, dim_out = {368}}},
affine6 = {{ltp = "affine6_ltp", bp = "affine6_bp"},
{dim_in = {368}, dim_out = {2048}}},
affine7 = {{ltp = "affine7_ltp", bp = "affine7_bp"},
{dim_in = {2048}, dim_out = {303}}},
affine8 = {{ltp = "affine8_ltp", bp = "affine8_bp"},
{dim_in = {303}, dim_out = {2048}}},
affine9 = {{ltp = "affine9_ltp", bp = "affine9_bp"},
{dim_in = {2048}, dim_out = {277}}},
affine10 = {{ltp = "affine10_ltp", bp = "affine10_bp"},
{dim_in = {277}, dim_out = {2048}}},
affine11 = {{ltp = "affine11_ltp", bp = "affine11_bp"},
{dim_in = {2048}, dim_out = {361}}},
affine12 = {{ltp = "affine12_ltp", bp = "affine12_bp"},
{dim_in = {361}, dim_out = {2048}}},
affine13 = {{ltp = "affine13_ltp", bp = "affine13_bp"},
{dim_in = {2048}, dim_out = {441}}},
affine14 = {{ltp = "affine14_ltp", bp = "affine14_bp"},
{dim_in = {441}, dim_out = {10092}}},
},
["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 = {10092, 1}, dim_out = {1}, compressed = true}}
},
["nerv.SoftmaxLayer"] = -- softmax for decode output
{
softmax = {{}, {dim_in = {10092}, dim_out = {10092}}}
}
}, param_repo, gconf)
layer_repo:add_layers(
{
["nerv.DAGLayer"] =
{
global_transf = {{}, {
dim_in = {1320}, dim_out = {1320},
sub_layers = layer_repo,
connections =
{
["<input>[1]"] = "blayer1[1]",
["blayer1[1]"] = "wlayer1[1]",
["wlayer1[1]"] = "<output>[1]"
}
}},
main = {{}, {
dim_in = {1320}, dim_out = {10092},
sub_layers = layer_repo,
connections = {
["<input>[1]"] = "affine0[1]",
["affine0[1]"] = "sigmoid0[1]",
["sigmoid0[1]"] = "affine1[1]",
["affine1[1]"] = "affine2[1]",
["affine2[1]"] = "sigmoid1[1]",
["sigmoid1[1]"] = "affine3[1]",
["affine3[1]"] = "affine4[1]",
["affine4[1]"] = "sigmoid2[1]",
["sigmoid2[1]"] = "affine5[1]",
["affine5[1]"] = "affine6[1]",
["affine6[1]"] = "sigmoid3[1]",
["sigmoid3[1]"] = "affine7[1]",
["affine7[1]"] = "affine8[1]",
["affine8[1]"] = "sigmoid4[1]",
["sigmoid4[1]"] = "affine9[1]",
["affine9[1]"] = "affine10[1]",
["affine10[1]"] = "sigmoid5[1]",
["sigmoid5[1]"] = "affine11[1]",
["affine11[1]"] = "affine12[1]",
["affine12[1]"] = "sigmoid6[1]",
["sigmoid6[1]"] = "affine13[1]",
["affine13[1]"] = "affine14[1]",
["affine14[1]"] = "<output>[1]",
}
}}
}
}, param_repo, gconf)
layer_repo:add_layers(
{
["nerv.DAGLayer"] =
{
ce_output = {{}, {
dim_in = {1320, 1}, dim_out = {1},
sub_layers = layer_repo,
connections = {
["<input>[1]"] = "main[1]",
["main[1]"] = "ce_crit[1]",
["<input>[2]"] = "ce_crit[2]",
["ce_crit[1]"] = "<output>[1]"
}
}},
softmax_output = {{}, {
dim_in = {1320}, dim_out = {10092},
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("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, feat_repo_shareid, data_mutex_shareid)
return {
{reader = nerv.TNetReader(gconf,
{
id = "main_scp",
scp_file = scp_file,
conf_file = gconf.htk_conf,
frm_ext = gconf.frm_ext,
mlfs = {
phone_state = {
file = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/ref.mlf",
format = "map",
format_arg = "/sgfs/users/wd007/asr/baseline_chn_50h/finetune/finetune_baseline/dict",
dir = "*/",
ext = "lab"
}
},
global_transf = layer_repo:get_layer("global_transf")
}, feat_repo_shareid, data_mutex_shareid),
data = {main_scp = 1320, phone_state = 1}}
}
end
function get_feat_id()
return {main_scp = true}
end
function make_buffer(readers)
return nerv.SGDBuffer(gconf,
{
buffer_size = gconf.buffer_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_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
function print_xent(xent)
local totalframes = xent:totalframes()
local loss = xent:loss()
local correct = xent:correct()
nerv.info_stderr("*** training statistics info begin ***")
nerv.info_stderr("total frames:\t\t%d", totalframes)
nerv.info_stderr("cross entropy:\t%.8f", loss/totalframes)
nerv.info_stderr("frame accuracy:\t%.3f%%", 100*correct/totalframes)
nerv.info_stderr("*** training statistics info end ***")
end
function frame_acc(xent)
local correct = xent:correct()
local totalframes = xent:totalframes()
return string.format("%.3f", 100*correct/totalframes)
end
function print_gconf()
nerv.info_stderr("%s \t:= %s", "network", gconf.initialized_param[1])
nerv.info_stderr("%s \t:= %s", "transf", gconf.initialized_param[2])
nerv.info_stderr("%s \t:= %s", "batch_size", gconf.batch_size)
nerv.info_stderr("%s \t:= %s", "buffer_size", gconf.buffer_size)
nerv.info_stderr("%s \t:= %s", "lrate", gconf.lrate)
nerv.info_stderr("%s \t:= %s", "tr_scp", gconf.tr_scp)
nerv.info_stderr("%s \t:= %s", "cv_scp", gconf.cv_scp)
end