diff options
| -rw-r--r-- | nerv/examples/asr_trainer.lua | 7 | ||||
| -rw-r--r-- | nerv/examples/gen_global_transf.lua | 62 | ||||
| -rw-r--r-- | nerv/examples/swb_baseline.lua | 204 |
3 files changed, 66 insertions, 207 deletions
diff --git a/nerv/examples/asr_trainer.lua b/nerv/examples/asr_trainer.lua index 52cb754..aa1019d 100644 --- a/nerv/examples/asr_trainer.lua +++ b/nerv/examples/asr_trainer.lua @@ -248,7 +248,7 @@ end dir.copyfile(arg[1], working_dir) -- set logfile path nerv.set_logfile(path.join(working_dir, logfile_name)) -path.chdir(working_dir) +--path.chdir(working_dir) -- start the training local trainer = build_trainer(pf0) @@ -258,7 +258,7 @@ nerv.info("initial cross validation: %.3f", gconf.accu_best) for i = gconf.cur_iter, gconf.max_iter do local stop = false gconf.cur_iter = i - dump_gconf(string.format("iter_%d.meta", i)) + dump_gconf(path.join(working_dir, string.format("iter_%d.meta", i))) repeat -- trick to implement `continue` statement nerv.info("[NN] begin iteration %d with lrate = %.6f", i, gconf.lrate) local accu_tr = trainer(nil, gconf.tr_scp, true, rebind_param_repo) @@ -270,7 +270,8 @@ for i = gconf.cur_iter, gconf.max_iter do os.date(date_pattern), i, gconf.lrate, accu_tr) - local accu_new, pr_new, param_fname = trainer(param_prefix, gconf.cv_scp, false) + local accu_new, pr_new, param_fname = + trainer(path.join(working_dir, param_prefix), gconf.cv_scp, false) nerv.info("[CV] cross validation %d: %.3f", i, accu_new) local accu_prev = gconf.accu_best if accu_new < gconf.accu_best then diff --git a/nerv/examples/gen_global_transf.lua b/nerv/examples/gen_global_transf.lua new file mode 100644 index 0000000..c4a3b42 --- /dev/null +++ b/nerv/examples/gen_global_transf.lua @@ -0,0 +1,62 @@ +if #arg < 1 then + return +end + +dofile(arg[1]) + +gconf.mmat_type = nerv.MMatrixFloat +gconf.cumat_type = nerv.CuMatrixFloat +local scp_file = gconf.tr_scp +local loc_type = nerv.ParamRepo.LOC_TYPES.ON_HOST +local reader_spec = make_readers(scp_file)[1] +local reader = reader_spec.reader +local width = reader_spec.data['main_scp'] +local mean = gconf.mmat_type(1, width) +local std = gconf.mmat_type(1, width) +local colsum = gconf.mmat_type(1, width) +local total = 0.0 +local EPS = 1e-7 + +mean:fill(0) +std:fill(0) + +local cnt = 0 +while (true) do + ret = reader:get_data() + if ret == nil then + break + end + + local utt = ret['main_scp'] + colsum = utt:colsum() + mean:add(mean, colsum, 1, 1) + + utt:mul_elem(utt, utt) + colsum = utt:colsum() + std:add(std, colsum, 1, 1) + + total = total + utt:nrow() + cnt = cnt + 1 + if cnt == 1000 then + nerv.info("accumulated %d utterances", cnt) + cnt = 0 + end +end + +local bparam = nerv.BiasParam("bias0", gconf) +bparam.trans = gconf.mmat_type(1, width) +mean:add(mean,mean, -1.0 / total, 0) -- -E(X) +bparam.trans:copy_fromh(mean) + +mean:mul_elem(mean, mean) -- E^2(X) +std:add(std, mean, 1 / total, -1) -- sigma ^ 2 + +for i = 0, width - 1 do + std[0][i] = math.sqrt(std[0][i] + EPS) + std[0][i] = 1 / (std[0][i] + EPS) +end + +local wparam = nerv.BiasParam("window0", gconf) +wparam.trans = std +local pr = nerv.ParamRepo({bparam, wparam}, loc_type) +pr:export("global_transf.nerv", nil) diff --git a/nerv/examples/swb_baseline.lua b/nerv/examples/swb_baseline.lua deleted file mode 100644 index ece4d44..0000000 --- a/nerv/examples/swb_baseline.lua +++ /dev/null @@ -1,204 +0,0 @@ -require 'htk_io' -gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9, - rearrange = true, -- just to make the context order consistent with old results, deprecated - frm_ext = 5, - frm_trim = 5, -- trim the first and last 5 frames, TNet just does this, deprecated - tr_scp = "/slfs1/users/mfy43/swb_ivec/train_bp.scp", - cv_scp = "/slfs1/users/mfy43/swb_ivec/train_cv.scp", - htk_conf = "/slfs1/users/mfy43/swb_ivec/plp_0_d_a.conf", - initialized_param = {"/slfs1/users/mfy43/swb_init.nerv", - "/slfs1/users/mfy43/swb_global_transf.nerv"}} - -function make_layer_repo(param_repo) - local layer_repo = nerv.LayerRepo( - { - -- global transf - ["nerv.BiasLayer"] = - { - blayer1 = {dim_in = {429}, dim_out = {429}, params = {bias = "bias1"}}, - blayer2 = {dim_in = {429}, dim_out = {429}, params = {bias = "bias2"}} - }, - ["nerv.WindowLayer"] = - { - wlayer1 = {dim_in = {429}, dim_out = {429}, params = {window = "window1"}}, - wlayer2 = {dim_in = {429}, dim_out = {429}, params = {window = "window2"}} - }, - -- 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.DAGLayer"] = - { - global_transf = { - dim_in = {429}, dim_out = {429}, - 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 = {429}, dim_out = {3001}, - 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"] = - { - ce_output = { - dim_in = {429, 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 = {429}, dim_out = {3001}, - 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) - 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 = "/slfs1/users/mfy43/swb_ivec/ref.mlf", - format = "map", - format_arg = "/slfs1/users/mfy43/swb_ivec/dict", - dir = "*/", - ext = "lab" - } - } - }), - data = {main_scp = 429, phone_state = 1}} - } -end - -function make_buffer(readers) - return nerv.SGDBuffer(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 |