From d88a57f4852c50a2678de950ee650ed9b6a895f0 Mon Sep 17 00:00:00 2001 From: Determinant Date: Sun, 8 May 2016 12:05:53 +0800 Subject: use `trainer.lua` in acoustic model examples --- nerv/examples/swb_baseline2.lua | 150 +++++++++++++++++----------------- nerv/examples/timit_baseline2.lua | 166 +++++++++++++++++++------------------- nerv/examples/trainer.lua | 2 +- 3 files changed, 160 insertions(+), 158 deletions(-) diff --git a/nerv/examples/swb_baseline2.lua b/nerv/examples/swb_baseline2.lua index 38cfb9a..87b01fa 100644 --- a/nerv/examples/swb_baseline2.lua +++ b/nerv/examples/swb_baseline2.lua @@ -1,65 +1,79 @@ 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 +gconf = {lrate = 0.8, + wcost = 1e-6, + momentum = 0.9, + frm_ext = 5, + rearrange = true, -- just to make the context order consistent with old TNet results, deprecated frm_trim = 5, -- trim the first and last 5 frames, TNet just does this, deprecated + chunk_size = 1, tr_scp = "/speechlab/users/mfy43/swb50/train_bp.scp", cv_scp = "/speechlab/users/mfy43/swb50/train_cv.scp", + ali = {file = "/speechlab/users/mfy43/swb50/ref.mlf", + format = "map", + format_arg = "/speechlab/users/mfy43/swb50/dict", + dir = "*/", + ext = "lab"}, 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 input_size = 429 +local output_size = 3001 +local hidden_size = 2048 +local trainer = nerv.Trainer + +function trainer: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"}} + blayer1 = {dim_in = {input_size}, dim_out = {input_size}, params = {bias = "bias0"}}, + blayer2 = {dim_in = {input_size}, dim_out = {input_size}, 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"}} + wlayer1 = {dim_in = {input_size}, dim_out = {input_size}, params = {window = "window0"}}, + wlayer2 = {dim_in = {input_size}, dim_out = {input_size}, params = {window = "window1"}} }, -- biased linearity ["nerv.AffineLayer"] = { - affine0 = {dim_in = {429}, dim_out = {2048}, + affine0 = {dim_in = {input_size}, dim_out = {hidden_size}, params = {ltp = "affine0_ltp", bp = "affine0_bp"}}, - affine1 = {dim_in = {2048}, dim_out = {2048}, + affine1 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine1_ltp", bp = "affine1_bp"}}, - affine2 = {dim_in = {2048}, dim_out = {2048}, + affine2 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine2_ltp", bp = "affine2_bp"}}, - affine3 = {dim_in = {2048}, dim_out = {2048}, + affine3 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine3_ltp", bp = "affine3_bp"}}, - affine4 = {dim_in = {2048}, dim_out = {2048}, + affine4 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine4_ltp", bp = "affine4_bp"}}, - affine5 = {dim_in = {2048}, dim_out = {2048}, + affine5 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine5_ltp", bp = "affine5_bp"}}, - affine6 = {dim_in = {2048}, dim_out = {2048}, + affine6 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine6_ltp", bp = "affine6_bp"}}, - affine7 = {dim_in = {2048}, dim_out = {3001}, + affine7 = {dim_in = {hidden_size}, dim_out = {output_size}, 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}} + 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}}, + sigmoid6 = {dim_in = {hidden_size}, dim_out = {hidden_size}} }, ["nerv.SoftmaxCELayer"] = -- softmax + ce criterion layer for finetune output { - ce_crit = {dim_in = {3001, 1}, dim_out = {1}, compressed = true} + ce_crit = {dim_in = {output_size, 1}, dim_out = {1}, compressed = true} }, ["nerv.SoftmaxLayer"] = -- softmax for decode output { - softmax = {dim_in = {3001}, dim_out = {3001}} + softmax = {dim_in = {output_size}, dim_out = {output_size}} } }, param_repo, gconf) @@ -68,7 +82,7 @@ function make_layer_repo(param_repo) ["nerv.GraphLayer"] = { global_transf = { - dim_in = {429}, dim_out = {429}, + dim_in = {input_size}, dim_out = {input_size}, layer_repo = layer_repo, connections = { {"[1]", "blayer1[1]", 0}, @@ -79,7 +93,7 @@ function make_layer_repo(param_repo) } }, main = { - dim_in = {429}, dim_out = {3001}, + dim_in = {input_size}, dim_out = {output_size}, layer_repo = layer_repo, connections = { {"[1]", "affine0[1]", 0}, @@ -108,20 +122,22 @@ function make_layer_repo(param_repo) ["nerv.GraphLayer"] = { ce_output = { - dim_in = {429, 1}, dim_out = {1}, + dim_in = {input_size, 1}, dim_out = {1}, layer_repo = layer_repo, connections = { - {"[1]", "main[1]", 0}, + {"[1]", "global_transf[1]", 0}, + {"global_transf[1]", "main[1]", 0}, {"main[1]", "ce_crit[1]", 0}, {"[2]", "ce_crit[2]", 0}, {"ce_crit[1]", "[1]", 0} } }, softmax_output = { - dim_in = {429}, dim_out = {3001}, + dim_in = {input_size}, dim_out = {output_size}, layer_repo = layer_repo, connections = { - {"[1]", "main[1]", 0}, + {"[1]", "global_transf[1]", 0}, + {"global_transf[1]", "main[1]", 0}, {"main[1]", "softmax[1]", 0}, {"softmax[1]", "[1]", 0} } @@ -132,73 +148,59 @@ function make_layer_repo(param_repo) return layer_repo end -function get_network(layer_repo) +function trainer: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, +function trainer:get_readers(dataset) + local function reader_gen(scp, ali) + return {{reader = nerv.HTKReader(gconf, { id = "main_scp", - scp_file = scp_file, + scp_file = scp, 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" - } + phone_state = ali } }), - data = {main_scp = 429, phone_state = 1}} - } + data = {main_scp = input_size, phone_state = 1}}} + end + if dataset == 'train' then + return reader_gen(gconf.tr_scp, gconf.tr_ali or gconf.ali) + elseif dataset == 'validate' then + return reader_gen(gconf.cv_scp, gconf.cv_ali or gconf.ali) + else + nerv.error('no such dataset') + end 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 - }) +function trainer:get_input_order() + return {"main_scp", "phone_state"} end -function get_input_order() - return {{id = "main_scp", global_transf = true}, - {id = "phone_state"}} +function trainer:get_decode_input_order() + return {"main_scp"} end -function get_decode_input_order() - return {{id = "main_scp", global_transf = true}} +function trainer:get_error() + local ce_crit = self.layer_repo:get_layer("ce_crit") + return ce_crit.total_ce / ce_crit.total_frames 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 +function trainer:mini_batch_afterprocess(cnt, info) + if cnt % 1000 == 0 then + self:epoch_afterprocess() + end end -function print_stat(layer_repo) - local ce_crit = layer_repo:get_layer("ce_crit") +function trainer:epoch_afterprocess() + local ce_crit = self.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.printf("accuracy:\t\t%.3f%%\n", ce_crit.total_correct / ce_crit.total_frames * 100) nerv.info("*** training stat end ***") end diff --git a/nerv/examples/timit_baseline2.lua b/nerv/examples/timit_baseline2.lua index 658aa2e..313156f 100644 --- a/nerv/examples/timit_baseline2.lua +++ b/nerv/examples/timit_baseline2.lua @@ -1,62 +1,77 @@ require 'kaldi_io' -gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9, frm_ext = 5, +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_20160311205342/nnet_init_20160311211609_iter_13_lr0.013437_tr72.572_cv58.709.nerv"}, - chunk_size = 1} + 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"} +} -function make_layer_repo(param_repo) +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 = {440}, dim_out = {440}, params = {bias = "bias0"}} + blayer1 = {dim_in = {input_size}, dim_out = {input_size}, params = {bias = "bias0"}, no_update_all = true} }, ["nerv.WindowLayer"] = { - wlayer1 = {dim_in = {440}, dim_out = {440}, params = {window = "window0"}} + wlayer1 = {dim_in = {input_size}, dim_out = {input_size}, params = {window = "window0"}, no_update_all = true} }, -- biased linearity ["nerv.AffineLayer"] = { - affine0 = {dim_in = {440}, dim_out = {1024}, + affine0 = {dim_in = {input_size}, dim_out = {hidden_size}, params = {ltp = "affine0_ltp", bp = "affine0_bp"}}, - affine1 = {dim_in = {1024}, dim_out = {1024}, + affine1 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine1_ltp", bp = "affine1_bp"}}, - affine2 = {dim_in = {1024}, dim_out = {1024}, + affine2 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine2_ltp", bp = "affine2_bp"}}, - affine3 = {dim_in = {1024}, dim_out = {1024}, + affine3 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine3_ltp", bp = "affine3_bp"}}, - affine4 = {dim_in = {1024}, dim_out = {1024}, + affine4 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine4_ltp", bp = "affine4_bp"}}, - affine5 = {dim_in = {1024}, dim_out = {1024}, + affine5 = {dim_in = {hidden_size}, dim_out = {hidden_size}, params = {ltp = "affine5_ltp", bp = "affine5_bp"}}, - affine6 = {dim_in = {1024}, dim_out = {1959}, + affine6 = {dim_in = {hidden_size}, dim_out = {output_size}, 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}} + 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 = {1959, 1}, dim_out = {1}, compressed = true} + ce_crit = {dim_in = {output_size, 1}, dim_out = {1}, compressed = true} }, ["nerv.SoftmaxLayer"] = -- softmax for decode output { - softmax = {dim_in = {1959}, dim_out = {1959}} + softmax = {dim_in = {output_size}, dim_out = {output_size}} } }, param_repo, gconf) @@ -65,7 +80,7 @@ function make_layer_repo(param_repo) ["nerv.GraphLayer"] = { global_transf = { - dim_in = {440}, dim_out = {440}, + dim_in = {input_size}, dim_out = {input_size}, layer_repo = layer_repo, connections = { {"[1]", "blayer1[1]", 0}, @@ -74,7 +89,7 @@ function make_layer_repo(param_repo) } }, main = { - dim_in = {440}, dim_out = {1959}, + dim_in = {input_size}, dim_out = {output_size}, layer_repo = layer_repo, connections = { {"[1]", "affine0[1]", 0}, @@ -101,20 +116,22 @@ function make_layer_repo(param_repo) ["nerv.GraphLayer"] = { ce_output = { - dim_in = {440, 1}, dim_out = {1}, + dim_in = {input_size, 1}, dim_out = {1}, layer_repo = layer_repo, connections = { - {"[1]", "main[1]", 0}, + {"[1]", "global_transf[1]", 0}, + {"global_transf[1]", "main[1]", 0}, {"main[1]", "ce_crit[1]", 0}, {"[2]", "ce_crit[2]", 0}, {"ce_crit[1]", "[1]", 0} } }, softmax_output = { - dim_in = {440}, dim_out = {1959}, + dim_in = {input_size}, dim_out = {output_size}, layer_repo = layer_repo, connections = { - {"[1]", "main[1]", 0}, + {"[1]", "global_transf[1]", 0}, + {"global_transf[1]", "main[1]", 0}, {"main[1]", "softmax[1]", 0}, {"softmax[1]", "[1]", 0} } @@ -125,90 +142,73 @@ function make_layer_repo(param_repo) return layer_repo end -function get_network(layer_repo) +function trainer: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, +function trainer:get_readers(dataset) + local function reader_gen(scp, ali) + return {{reader = nerv.KaldiReader(gconf, { id = "main_scp", - feature_rspecifier = scp_file, - conf_file = gconf.htk_conf, + feature_rspecifier = scp, 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" - } + phone_state = ali } }), - data = {main_scp = 440, phone_state = 1}} - } + data = {main_scp = input_size, phone_state = 1}}} + end + if dataset == 'train' then + return reader_gen(gconf.tr_scp, gconf.tr_ali or gconf.ali) + elseif dataset == 'validate' then + return reader_gen(gconf.cv_scp, gconf.cv_ali or gconf.ali) + else + nerv.error('no such dataset') + end +end + +function trainer:get_input_order() + return {"main_scp", "phone_state"} +end + +function trainer:get_decode_network(layer_repo) + return layer_repo:get_layer("softmax_output") end -function make_decode_readers(scp_file, layer_repo) - return { - {reader = nerv.KaldiReader(gconf, +function trainer:make_decode_readers(scp_file) + 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 + mlfs = {} }), - data = {main_scp = 440, 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 - }) + data = {main_scp = input_size, phone_state = 1}}} end -function get_input_order() - return {{id = "main_scp", global_transf = true}, - {id = "phone_state"}} +function trainer:get_decode_input_order() + return {"main_scp"} end -function get_decode_input_order() - return {{id = "main_scp", global_transf = true}} +function trainer:get_error() + local ce_crit = self.layer_repo:get_layer("ce_crit") + return ce_crit.total_ce / ce_crit.total_frames 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 +function trainer:mini_batch_afterprocess(cnt, info) + if cnt % 1000 == 0 then + self:epoch_afterprocess() + end end -function print_stat(layer_repo) - local ce_crit = layer_repo:get_layer("ce_crit") +function trainer:epoch_afterprocess() + local ce_crit = self.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.printf("accuracy:\t\t%.3f%%\n", ce_crit.total_correct / ce_crit.total_frames * 100) nerv.info("*** training stat end ***") end diff --git a/nerv/examples/trainer.lua b/nerv/examples/trainer.lua index 8e3efcb..f6c7a5a 100644 --- a/nerv/examples/trainer.lua +++ b/nerv/examples/trainer.lua @@ -36,7 +36,7 @@ local function make_options(spec) end local function print_help(options) - nerv.printf("Usage: [options] network_config.lua\n") + nerv.printf("Usage: [options] \n") nerv.print_usage(options) end -- cgit v1.2.3