diff options
Diffstat (limited to 'nerv/examples/timit_baseline2.lua')
| -rw-r--r-- | nerv/examples/timit_baseline2.lua | 212 |
1 files changed, 212 insertions, 0 deletions
diff --git a/nerv/examples/timit_baseline2.lua b/nerv/examples/timit_baseline2.lua new file mode 100644 index 0000000..d783c3d --- /dev/null +++ b/nerv/examples/timit_baseline2.lua @@ -0,0 +1,212 @@ +require 'kaldi_io' +gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9, frm_ext = 5, + tr_scp = "ark:/speechlab/tools/KALDI/kaldi-master/src/featbin/copy-feats " .. + "scp:/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_prepare/train.scp ark:- |", + cv_scp = "ark:/speechlab/tools/KALDI/kaldi-master/src/featbin/copy-feats " .. + "scp:/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_prepare/cv.scp ark:- |", + initialized_param = {"/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_prepare/nnet_init.nerv", + "/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_prepare/nnet_output.nerv", + "/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_prepare/nnet_trans.nerv"}, + decode_param = {"/speechlab/users/mfy43/timit/nnet_init_20160229015745_iter_13_lr0.013437_tr72.434_cv58.729.nerv", + "/speechlab/users/mfy43/timit/s5/exp/dnn4_nerv_prepare/nnet_trans.nerv"}} + +function make_layer_repo(param_repo) + local layer_repo = nerv.LayerRepo( + { + -- global transf + ["nerv.BiasLayer"] = + { + blayer1 = {dim_in = {440}, dim_out = {440}, params = {bias = "bias0"}} + }, + ["nerv.WindowLayer"] = + { + wlayer1 = {dim_in = {440}, dim_out = {440}, params = {window = "window0"}} + }, + -- biased linearity + ["nerv.AffineLayer"] = + { + affine0 = {dim_in = {440}, dim_out = {1024}, + params = {ltp = "affine0_ltp", bp = "affine0_bp"}}, + affine1 = {dim_in = {1024}, dim_out = {1024}, + params = {ltp = "affine1_ltp", bp = "affine1_bp"}}, + affine2 = {dim_in = {1024}, dim_out = {1024}, + params = {ltp = "affine2_ltp", bp = "affine2_bp"}}, + affine3 = {dim_in = {1024}, dim_out = {1024}, + params = {ltp = "affine3_ltp", bp = "affine3_bp"}}, + affine4 = {dim_in = {1024}, dim_out = {1024}, + params = {ltp = "affine4_ltp", bp = "affine4_bp"}}, + affine5 = {dim_in = {1024}, dim_out = {1024}, + params = {ltp = "affine5_ltp", bp = "affine5_bp"}}, + affine6 = {dim_in = {1024}, dim_out = {1959}, + params = {ltp = "affine6_ltp", bp = "affine6_bp"}} + }, + ["nerv.SigmoidLayer"] = + { + sigmoid0 = {dim_in = {1024}, dim_out = {1024}}, + sigmoid1 = {dim_in = {1024}, dim_out = {1024}}, + sigmoid2 = {dim_in = {1024}, dim_out = {1024}}, + sigmoid3 = {dim_in = {1024}, dim_out = {1024}}, + sigmoid4 = {dim_in = {1024}, dim_out = {1024}}, + sigmoid5 = {dim_in = {1024}, dim_out = {1024}} + }, + ["nerv.SoftmaxCELayer"] = -- softmax + ce criterion layer for finetune output + { + ce_crit = {dim_in = {1959, 1}, dim_out = {1}, compressed = true} + }, + ["nerv.SoftmaxLayer"] = -- softmax for decode output + { + softmax = {dim_in = {1959}, dim_out = {1959}} + } + }, param_repo, gconf) + + layer_repo:add_layers( + { + ["nerv.GraphLayer"] = + { + global_transf = { + dim_in = {440}, dim_out = {440}, + layer_repo = layer_repo, + connections = { + {"<input>[1]", "blayer1[1]", 0}, + {"blayer1[1]", "wlayer1[1]", 0}, + {"wlayer1[1]", "<output>[1]", 0} + } + }, + main = { + dim_in = {440}, dim_out = {1959}, + 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]", "<output>[1]", 0} + } + } + } + }, param_repo, gconf) + + layer_repo:add_layers( + { + ["nerv.GraphLayer"] = + { + ce_output = { + dim_in = {440, 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 = {440}, dim_out = {1959}, + 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.KaldiReader(gconf, + { + id = "main_scp", + feature_rspecifier = scp_file, + conf_file = gconf.htk_conf, + frm_ext = gconf.frm_ext, + mlfs = { + phone_state = { + 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:- |", + format = "map" + } + } + }), + data = {main_scp = 440, phone_state = 1}} + } +end + +function make_decode_readers(scp_file, layer_repo) + return { + {reader = nerv.KaldiReader(gconf, + { + id = "main_scp", + feature_rspecifier = scp_file, + conf_file = gconf.htk_conf, + frm_ext = gconf.frm_ext, + mlfs = {}, + need_key = true + }), + data = {main_scp = 440, phone_state = 1}} + } +end + +function make_buffer(readers) + return nerv.SGDBuffer(gconf, + { + buffer_size = gconf.buffer_size, + batch_size = gconf.batch_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 |