From 0dbc40a998e2879f27110bd449f4cae9550c4e41 Mon Sep 17 00:00:00 2001 From: txh18 Date: Fri, 4 Dec 2015 21:44:06 +0800 Subject: trying to use lstm_t_v2 for ptb --- nerv/examples/lmptb/lmptb/lstm_t_v2.lua | 123 ++++++++ nerv/examples/lmptb/lstmlm_ptb_main.lua | 10 +- nerv/examples/lmptb/lstmlm_v2_ptb_main.lua | 470 +++++++++++++++++++++++++++++ nerv/layer/gate_fff.lua | 56 ++-- nerv/tnn/layersT/lstm_t.lua | 2 +- 5 files changed, 628 insertions(+), 33 deletions(-) create mode 100644 nerv/examples/lmptb/lmptb/lstm_t_v2.lua create mode 100644 nerv/examples/lmptb/lstmlm_v2_ptb_main.lua diff --git a/nerv/examples/lmptb/lmptb/lstm_t_v2.lua b/nerv/examples/lmptb/lmptb/lstm_t_v2.lua new file mode 100644 index 0000000..d3fc920 --- /dev/null +++ b/nerv/examples/lmptb/lmptb/lstm_t_v2.lua @@ -0,0 +1,123 @@ +local LSTMLayerT = nerv.class('nerv.LSTMLayerTv2', 'nerv.LayerT') +--a version of LSTM that only feed h into the gates + +function LSTMLayerT:__init(id, global_conf, layer_conf) + --input1:x input2:h input3:c + self.id = id + self.dim_in = layer_conf.dim_in + self.dim_out = layer_conf.dim_out + self.gconf = global_conf + + --prepare a DAGLayerT to hold the lstm structure + local pr = layer_conf.pr + if pr == nil then + pr = nerv.ParamRepo() + end + + local function ap(str) + return self.id .. '.' .. str + end + + local layers = { + ["nerv.CombinerLayer"] = { + [ap("inputXDup")] = {{}, {["dim_in"] = {self.dim_in[1]}, + ["dim_out"] = {self.dim_in[1], self.dim_in[1], self.dim_in[1], self.dim_in[1]}, ["lambda"] = {1}}}, + [ap("inputHDup")] = {{}, {["dim_in"] = {self.dim_in[2]}, + ["dim_out"] = {self.dim_in[2], self.dim_in[2], self.dim_in[2], self.dim_in[2]}, ["lambda"] = {1}}}, + [ap("inputCDup")] = {{}, {["dim_in"] = {self.dim_in[3]}, + ["dim_out"] = {self.dim_in[3]}, ["lambda"] = {1}}}, + [ap("mainCDup")] = {{}, {["dim_in"] = {self.dim_in[3], self.dim_in[3]}, + ["dim_out"] = {self.dim_in[3], self.dim_in[3]}, ["lambda"] = {1, 1}}}, + }, + ["nerv.AffineLayer"] = { + [ap("mainAffineL")] = {{}, {["dim_in"] = {self.dim_in[1], self.dim_in[2], self.dim_in[3]}, + ["dim_out"] = {self.dim_out[1]}, ["pr"] = pr}}, + }, + ["nerv.TanhLayer"] = { + [ap("mainTanhL")] = {{}, {["dim_in"] = {self.dim_out[1]}, ["dim_out"] = {self.dim_out[1]}}}, + [ap("outputTanhL")] = {{}, {["dim_in"] = {self.dim_out[1]}, ["dim_out"] = {self.dim_out[1]}}}, + }, + ["nerv.GateFLayer"] = { + [ap("forgetGateL")] = {{}, {["dim_in"] = {self.dim_in[1], self.dim_in[2]}, + ["dim_out"] = {self.dim_in[3]}, ["pr"] = pr}}, + [ap("inputGateL")] = {{}, {["dim_in"] = {self.dim_in[1], self.dim_in[2]}, + ["dim_out"] = {self.dim_in[3]}, ["pr"] = pr}}, + [ap("outputGateL")] = {{}, {["dim_in"] = {self.dim_in[1], self.dim_in[2]}, + ["dim_out"] = {self.dim_in[3]}, ["pr"] = pr}}, + + }, + ["nerv.ElemMulLayer"] = { + [ap("inputGMulL")] = {{}, {["dim_in"] = {self.dim_in[3], self.dim_in[3]}, ["dim_out"] = {self.dim_in[3]}}}, + [ap("forgetGMulL")] = {{}, {["dim_in"] = {self.dim_in[3], self.dim_in[3]}, ["dim_out"] = {self.dim_in[3]}}}, + [ap("outputGMulL")] = {{}, {["dim_in"] = {self.dim_in[3], self.dim_in[3]}, ["dim_out"] = {self.dim_in[3]}}}, + }, + } + + local layerRepo = nerv.LayerRepo(layers, pr, global_conf) + + local connections_t = { + ["[1]"] = ap("inputXDup[1]"), + ["[2]"] = ap("inputHDup[1]"), + ["[3]"] = ap("inputCDup[1]"), + + [ap("inputXDup[1]")] = ap("mainAffineL[1]"), + [ap("inputHDup[1]")] = ap("mainAffineL[2]"), + + [ap("mainAffineL[1]")] = ap("mainTanhL[1]"), + + [ap("inputXDup[2]")] = ap("inputGateL[1]"), + [ap("inputHDup[2]")] = ap("inputGateL[2]"), + + [ap("inputXDup[3]")] = ap("forgetGateL[1]"), + [ap("inputHDup[3]")] = ap("forgetGateL[2]"), + + [ap("mainTanhL[1]")] = ap("inputGMulL[1]"), + [ap("inputGateL[1]")] = ap("inputGMulL[2]"), + + [ap("inputCDup[1]")] = ap("forgetGMulL[1]"), + [ap("forgetGateL[1]")] = ap("forgetGMulL[2]"), + + [ap("inputGMulL[1]")] = ap("mainCDup[1]"), + [ap("forgetGMulL[1]")] = ap("mainCDup[2]"), + + [ap("inputXDup[4]")] = ap("outputGateL[1]"), + [ap("inputHDup[4]")] = ap("outputGateL[2]"), + + [ap("mainCDup[2]")] = "[2]", + [ap("mainCDup[1]")] = ap("outputTanhL[1]"), + + [ap("outputTanhL[1]")] = ap("outputGMulL[1]"), + [ap("outputGateL[1]")] = ap("outputGMulL[2]"), + + [ap("outputGMulL[1]")] = "[1]", + } + self.dagL = nerv.DAGLayerT(self.id, global_conf, + {["dim_in"] = self.dim_in, ["dim_out"] = self.dim_out, ["sub_layers"] = layerRepo, + ["connections"] = connections_t}) + + self:check_dim_len(3, 2) -- x, h, c and h, c +end + +function LSTMLayerT:init(batch_size, chunk_size) + self.dagL:init(batch_size, chunk_size) +end + +function LSTMLayerT:batch_resize(batch_size, chunk_size) + self.dagL:batch_resize(batch_size, chunk_size) +end + +function LSTMLayerT:update(bp_err, input, output, t) + self.dagL:update(bp_err, input, output, t) +end + +function LSTMLayerT:propagate(input, output, t) + self.dagL:propagate(input, output, t) +end + +function LSTMLayerT:back_propagate(bp_err, next_bp_err, input, output, t) + self.dagL:back_propagate(bp_err, next_bp_err, input, output, t) +end + +function LSTMLayerT:get_params() + return self.dagL:get_params() +end diff --git a/nerv/examples/lmptb/lstmlm_ptb_main.lua b/nerv/examples/lmptb/lstmlm_ptb_main.lua index 08e4849..d5408ba 100644 --- a/nerv/examples/lmptb/lstmlm_ptb_main.lua +++ b/nerv/examples/lmptb/lstmlm_ptb_main.lua @@ -343,11 +343,11 @@ end nerv.LMUtil.wait(2) nerv.printf("%s printing training scheduling options...\n", global_conf.sche_log_pre) -nerv.printf("lr_half:%s\n", tostring(lr_half)) -nerv.printf("start_iter:%s\n", tostring(start_iter)) -nerv.printf("ppl_last:%s\n", tostring(ppl_last)) -nerv.printf("commds_str:%s\n", commands_str) -nerv.printf("test_iter:%s\n", tostring(test_iter)) +nerv.printf("lr_half:\t%s\n", tostring(lr_half)) +nerv.printf("start_iter:\t%s\n", tostring(start_iter)) +nerv.printf("ppl_last:\t%s\n", tostring(ppl_last)) +nerv.printf("commds_str:\t%s\n", commands_str) +nerv.printf("test_iter:\t%s\n", tostring(test_iter)) nerv.printf("%s printing training scheduling end.\n", global_conf.sche_log_pre) nerv.LMUtil.wait(2) ------------------printing options end------------------------------ diff --git a/nerv/examples/lmptb/lstmlm_v2_ptb_main.lua b/nerv/examples/lmptb/lstmlm_v2_ptb_main.lua new file mode 100644 index 0000000..a3d7584 --- /dev/null +++ b/nerv/examples/lmptb/lstmlm_v2_ptb_main.lua @@ -0,0 +1,470 @@ +require 'lmptb.lmvocab' +require 'lmptb.lmfeeder' +require 'lmptb.lmutil' +require 'lmptb.layer.init' +--require 'tnn.init' +require 'lmptb.lmseqreader' +require 'lm_trainer' +require 'lmptb.lstm_t_v2' + +--[[global function rename]]-- +--local printf = nerv.printf +local LMTrainer = nerv.LMTrainer +--[[global function rename ends]]-- + +--global_conf: table +--first_time: bool +--Returns: a ParamRepo +function prepare_parameters(global_conf, iter) + nerv.printf("%s preparing parameters...\n", global_conf.sche_log_pre) + + global_conf.paramRepo = nerv.ParamRepo() + local paramRepo = global_conf.paramRepo + + if iter == -1 then --first time + nerv.printf("%s first time, prepare some pre-set parameters, and leaving other parameters to auto-generation...\n", global_conf.sche_log_pre) + local f = nerv.ChunkFile(global_conf.param_fn .. '.0', 'w') + f:close() + --[[ + ltp_ih = nerv.LinearTransParam("ltp_ih", global_conf) + ltp_ih.trans = global_conf.cumat_type(global_conf.vocab:size(), global_conf.hidden_size) --index 0 is for zero, others correspond to vocab index(starting from 1) + ltp_ih.trans:generate(global_conf.param_random) + + ltp_hh = nerv.LinearTransParam("ltp_hh", global_conf) + ltp_hh.trans = global_conf.cumat_type(global_conf.hidden_size, global_conf.hidden_size) + ltp_hh.trans:generate(global_conf.param_random) + + --ltp_ho = nerv.LinearTransParam("ltp_ho", global_conf) + --ltp_ho.trans = global_conf.cumat_type(global_conf.hidden_size, global_conf.vocab:size()) + --ltp_ho.trans:generate(global_conf.param_random) + + bp_h = nerv.BiasParam("bp_h", global_conf) + bp_h.trans = global_conf.cumat_type(1, global_conf.hidden_size) + bp_h.trans:generate(global_conf.param_random) + + --bp_o = nerv.BiasParam("bp_o", global_conf) + --bp_o.trans = global_conf.cumat_type(1, global_conf.vocab:size()) + --bp_o.trans:generate(global_conf.param_random) + + local f = nerv.ChunkFile(global_conf.param_fn .. '.0', 'w') + f:write_chunk(ltp_ih) + f:write_chunk(ltp_hh) + --f:write_chunk(ltp_ho) + f:write_chunk(bp_h) + --f:write_chunk(bp_o) + f:close() + ]]-- + return nil + end + + nerv.printf("%s loading parameter from file %s...\n", global_conf.sche_log_pre, global_conf.param_fn .. '.' .. tostring(iter)) + paramRepo:import({global_conf.param_fn .. '.' .. tostring(iter)}, nil, global_conf) + + nerv.printf("%s preparing parameters end.\n", global_conf.sche_log_pre) + + return nil +end + +--global_conf: table +--Returns: nerv.LayerRepo +function prepare_layers(global_conf) + nerv.printf("%s preparing layers...\n", global_conf.sche_log_pre) + + local pr = global_conf.paramRepo + + local du = false + + --local recurrentLconfig = {{["bp"] = "bp_h", ["ltp_hh"] = "ltp_hh"}, {["dim_in"] = {global_conf.hidden_size, global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}, ["break_id"] = global_conf.vocab:get_sen_entry().id, ["independent"] = global_conf.independent, ["clip"] = 10}} + --local recurrentLconfig = {{}, {["dim_in"] = {global_conf.hidden_size, global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}, ["clip"] = 10, ["direct_update"] = du, ["pr"] = pr}} + + local layers = { + --["nerv.AffineRecurrentLayer"] = { + -- ["recurrentL1"] = recurrentLconfig, + --}, + + ["nerv.LSTMLayerTv2"] = { + ["lstmL1"] = {{}, {["dim_in"] = {global_conf.hidden_size, global_conf.hidden_size, global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size, global_conf.hidden_size}, ["pr"] = pr}}, + }, + + ["nerv.DropoutLayerT"] = { + ["dropoutL1"] = {{}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}}}, + }, + + ["nerv.SelectLinearLayer"] = { + ["selectL1"] = {{}, {["dim_in"] = {1}, ["dim_out"] = {global_conf.hidden_size}, ["vocab"] = global_conf.vocab, ["pr"] = pr}}, + }, + + --["nerv.SigmoidLayer"] = { + -- ["sigmoidL1"] = {{}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}}} + --}, + + ["nerv.CombinerLayer"] = { + ["combinerL1"] = {{}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size, global_conf.hidden_size}, ["lambda"] = {1}}}, + }, + + ["nerv.AffineLayer"] = { + ["outputL"] = {{}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.vocab:size()}, ["direct_update"] = du, ["pr"] = pr}}, + }, + + ["nerv.SoftmaxCELayerT"] = { + ["softmaxL"] = {{}, {["dim_in"] = {global_conf.vocab:size(), global_conf.vocab:size()}, ["dim_out"] = {1}}}, + }, + } + + for l = 2, global_conf.layer_num do + layers["nerv.DropoutLayerT"]["dropoutL" .. l] = {{}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}}} + layers["nerv.LSTMLayerTv2"]["lstmL" .. l] = {{}, {["dim_in"] = {global_conf.hidden_size, global_conf.hidden_size, global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size, global_conf.hidden_size}, ["pr"] = pr}} + layers["nerv.CombinerLayer"]["combinerL" .. l] = {{}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size, global_conf.hidden_size}, ["lambda"] = {1}}} + end + --[[ --we do not need those in the new tnn framework + printf("%s adding %d bptt layers...\n", global_conf.sche_log_pre, global_conf.bptt) + for i = 1, global_conf.bptt do + layers["nerv.IndRecurrentLayer"]["recurrentL" .. (i + 1)] = recurrentLconfig + layers["nerv.SigmoidLayer"]["sigmoidL" .. (i + 1)] = {{}, {["dim_in"] = {global_conf.hidden_size}, ["dim_out"] = {global_conf.hidden_size}}} + layers["nerv.SelectLinearLayer"]["selectL" .. (i + 1)] = {{["ltp"] = "ltp_ih"}, {["dim_in"] = {1}, ["dim_out"] = {global_conf.hidden_size}}} + end + --]] + + local layerRepo = nerv.LayerRepo(layers, pr, global_conf) + nerv.printf("%s preparing layers end.\n", global_conf.sche_log_pre) + return layerRepo +end + +--global_conf: table +--layerRepo: nerv.LayerRepo +--Returns: a nerv.TNN +function prepare_tnn(global_conf, layerRepo) + nerv.printf("%s Generate and initing TNN ...\n", global_conf.sche_log_pre) + + --input: input_w, input_w, ... input_w_now, last_activation + local connections_t = { + {"[1]", "selectL1[1]", 0}, + + --{"selectL1[1]", "recurrentL1[1]", 0}, + --{"recurrentL1[1]", "sigmoidL1[1]", 0}, + --{"sigmoidL1[1]", "combinerL1[1]", 0}, + --{"combinerL1[1]", "recurrentL1[2]", 1}, + + {"selectL1[1]", "lstmL1[1]", 0}, + {"lstmL1[2]", "lstmL1[3]", 1}, + {"lstmL1[1]", "combinerL1[1]", 0}, + {"combinerL1[1]", "lstmL1[2]", 1}, + {"combinerL1[2]", "dropoutL1[1]", 0}, + + {"dropoutL"..global_conf.layer_num.."[1]", "outputL[1]", 0}, + {"outputL[1]", "softmaxL[1]", 0}, + {"[2]", "softmaxL[2]", 0}, + {"softmaxL[1]", "[1]", 0} + } + + for l = 2, global_conf.layer_num do + table.insert(connections_t, {"dropoutL"..(l-1).."[1]", "lstmL"..l.."[1]", 0}) + table.insert(connections_t, {"lstmL"..l.."[2]", "lstmL"..l.."[3]", 1}) + table.insert(connections_t, {"lstmL"..l.."[1]", "combinerL"..l.."[1]", 0}) + table.insert(connections_t, {"combinerL"..l.."[1]", "lstmL"..l.."[2]", 1}) + table.insert(connections_t, {"combinerL"..l.."[2]", "dropoutL"..l.."[1]", 0}) + end + + --[[ + printf("%s printing DAG connections:\n", global_conf.sche_log_pre) + for key, value in pairs(connections_t) do + printf("\t%s->%s\n", key, value) + end + ]]-- + + local tnn = nerv.TNN("TNN", global_conf, {["dim_in"] = {1, global_conf.vocab:size()}, + ["dim_out"] = {1}, ["sub_layers"] = layerRepo, + ["connections"] = connections_t, ["clip_t"] = global_conf.clip_t, + }) + + tnn:init(global_conf.batch_size, global_conf.chunk_size) + + nerv.printf("%s Initing TNN end.\n", global_conf.sche_log_pre) + return tnn +end + +function load_net(global_conf, next_iter) + prepare_parameters(global_conf, next_iter) + local layerRepo = prepare_layers(global_conf) + local tnn = prepare_tnn(global_conf, layerRepo) + return tnn +end + +local train_fn, valid_fn, test_fn +global_conf = {} +local set = arg[1] --"test" + +if (set == "ptb") then + +root_dir = '/home/slhome/txh18/workspace' +data_dir = root_dir .. '/ptb/DATA' +train_fn = data_dir .. '/ptb.train.txt.adds' +valid_fn = data_dir .. '/ptb.valid.txt.adds' +test_fn = data_dir .. '/ptb.test.txt.adds' +vocab_fn = data_dir .. '/vocab' + +qdata_dir = root_dir .. '/ptb/questionGen/gen' + +global_conf = { + lrate = 0.15, wcost = 1e-5, momentum = 0, clip_t = 5, + cumat_type = nerv.CuMatrixFloat, + mmat_type = nerv.MMatrixFloat, + nn_act_default = 0, + + hidden_size = 300, + layer_num = 1, + chunk_size = 15, + batch_size = 20, + max_iter = 45, + lr_decay = 1.003, + decay_iter = 10, + param_random = function() return (math.random() / 5 - 0.1) end, + dropout_str = "0.5", + + train_fn = train_fn, + valid_fn = valid_fn, + test_fn = test_fn, + vocab_fn = vocab_fn, + max_sen_len = 90, + sche_log_pre = "[SCHEDULER]:", + log_w_num = 40000, --give a message when log_w_num words have been processed + timer = nerv.Timer(), + work_dir_base = '/home/slhome/txh18/workspace/ptb/EXP-nerv/lstmlm_v2.0' +} + +elseif (set == "msr_sc") then + +data_dir = '/home/slhome/txh18/workspace/sentenceCompletion/DATA_PV2' +train_fn = data_dir .. '/normed_all.sf.len60.adds.train' +valid_fn = data_dir .. '/normed_all.sf.len60.adds.dev' +test_fn = data_dir .. '/answer_normed.adds' +vocab_fn = data_dir .. '/normed_all.choose.vocab30000.addqvocab' + +global_conf = { + lrate = 1, wcost = 1e-6, momentum = 0, + cumat_type = nerv.CuMatrixFloat, + mmat_type = nerv.MMatrixFloat, + nn_act_default = 0, + + hidden_size = 300, + layer_num = 1, + chunk_size = 15, + batch_size = 10, + max_iter = 30, + decay_iter = 10, + lr_decay = 1.003, + param_random = function() return (math.random() / 5 - 0.1) end, + dropout_str = "0", + + train_fn = train_fn, + valid_fn = valid_fn, + test_fn = test_fn, + vocab_fn = vocab_fn, + sche_log_pre = "[SCHEDULER]:", + log_w_num = 400000, --give a message when log_w_num words have been processed + timer = nerv.Timer(), + work_dir_base = '/home/slhome/txh18/workspace/sentenceCompletion/EXP-Nerv/rnnlm_test' +} + +else + +valid_fn = '/home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/m-tests/some-text' +train_fn = '/home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/m-tests/some-text' +test_fn = '/home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/m-tests/some-text' +vocab_fn = '/home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/m-tests/some-text' + +global_conf = { + lrate = 0.01, wcost = 1e-5, momentum = 0, + cumat_type = nerv.CuMatrixFloat, + mmat_type = nerv.MMatrixFloat, + nn_act_default = 0, + + hidden_size = 20, + layer_num = 1, + chunk_size = 2, + batch_size = 10, + max_iter = 3, + param_random = function() return (math.random() / 5 - 0.1) end, + dropout_str = "0", + + train_fn = train_fn, + valid_fn = valid_fn, + test_fn = test_fn, + lr_decay = 1.003, + decay_iter = 10, + vocab_fn = vocab_fn, + sche_log_pre = "[SCHEDULER]:", + log_w_num = 10, --give a message when log_w_num words have been processed + timer = nerv.Timer(), + work_dir_base = '/home/slhome/txh18/workspace/nerv/play/testEXP/tnn_lstmlm_test' +} + +end + +local lr_half = false --can not be local, to be set by loadstring +local start_iter = -1 +local ppl_last = 100000 +local commands_str = "train:test" +local commands = {} +local test_iter = -1 + +--for testout(question) +local q_file = "ptb.test.txt.q10rs1_Msss.adds" + +if arg[2] ~= nil then + nerv.printf("%s applying arg[2](%s)...\n", global_conf.sche_log_pre, arg[2]) + loadstring(arg[2])() + nerv.LMUtil.wait(0.5) +else + nerv.printf("%s no user setting, all default...\n", global_conf.sche_log_pre) +end + +global_conf.work_dir = global_conf.work_dir_base .. 'h' .. global_conf.hidden_size .. 'l' .. global_conf.layer_num .. 'ch' .. global_conf.chunk_size .. 'ba' .. global_conf.batch_size .. 'slr' .. global_conf.lrate .. 'wc' .. global_conf.wcost .. 'dr' .. global_conf.dropout_str +global_conf.train_fn_shuf = global_conf.work_dir .. '/train_fn_shuf' +global_conf.train_fn_shuf_bak = global_conf.train_fn_shuf .. '_bak' +global_conf.param_fn = global_conf.work_dir .. "/params" +global_conf.dropout_list = nerv.SUtil.parse_schedule(global_conf.dropout_str) +global_conf.log_fn = global_conf.work_dir .. '/log_lstm_tnn_' .. commands_str ..os.date("_TT%X_%m_%d",os.time()) +commands = nerv.SUtil.parse_commands_set(commands_str) + +nerv.printf("%s creating work_dir(%s)...\n", global_conf.sche_log_pre, global_conf.work_dir) +nerv.LMUtil.wait(2) +os.execute("mkdir -p "..global_conf.work_dir) +os.execute("cp " .. global_conf.train_fn .. " " .. global_conf.train_fn_shuf) + +--redirecting log outputs! +nerv.SUtil.log_redirect(global_conf.log_fn) +nerv.LMUtil.wait(2) + +----------------printing options--------------------------------- +nerv.printf("%s printing global_conf...\n", global_conf.sche_log_pre) +for id, value in pairs(global_conf) do + nerv.printf("%s:\t%s\n", id, tostring(value)) +end +nerv.LMUtil.wait(2) + +nerv.printf("%s printing training scheduling options...\n", global_conf.sche_log_pre) +nerv.printf("lr_half:\t%s\n", tostring(lr_half)) +nerv.printf("start_iter:\t%s\n", tostring(start_iter)) +nerv.printf("ppl_last:\t%s\n", tostring(ppl_last)) +nerv.printf("commds_str:\t%s\n", commands_str) +nerv.printf("test_iter:\t%s\n", tostring(test_iter)) +nerv.printf("%s printing training scheduling end.\n", global_conf.sche_log_pre) +nerv.LMUtil.wait(2) +------------------printing options end------------------------------ + +math.randomseed(1) + +local vocab = nerv.LMVocab() +global_conf["vocab"] = vocab +nerv.printf("%s building vocab...\n", global_conf.sche_log_pre) +global_conf.vocab:build_file(global_conf.vocab_fn, false) +ppl_rec = {} + +local final_iter = -1 +if commands["train"] == 1 then + if start_iter == -1 then + prepare_parameters(global_conf, -1) --write pre_generated params to param.0 file + end + + if start_iter == -1 or start_iter == 0 then + nerv.printf("===INITIAL VALIDATION===\n") + local tnn = load_net(global_conf, 0) + global_conf.paramRepo = tnn:get_params() --get auto-generted params + global_conf.paramRepo:export(global_conf.param_fn .. '.0', nil) --some parameters are auto-generated, saved again to param.0 file + global_conf.dropout_rate = 0 + local result = LMTrainer.lm_process_file_rnn(global_conf, global_conf.valid_fn, tnn, false) --false update! + nerv.LMUtil.wait(1) + ppl_rec[0] = {} + ppl_rec[0].valid = result:ppl_all("rnn") + ppl_last = ppl_rec[0].valid + ppl_rec[0].train = 0 + ppl_rec[0].test = 0 + ppl_rec[0].lr = 0 + + start_iter = 1 + + nerv.printf("\n") + end + + for iter = start_iter, global_conf.max_iter, 1 do + final_iter = iter --for final testing + global_conf.sche_log_pre = "[SCHEDULER ITER"..iter.." LR"..global_conf.lrate.."]:" + tnn = load_net(global_conf, iter - 1) + nerv.printf("===ITERATION %d LR %f===\n", iter, global_conf.lrate) + global_conf.dropout_rate = nerv.SUtil.sche_get(global_conf.dropout_list, iter) + result = LMTrainer.lm_process_file_rnn(global_conf, global_conf.train_fn_shuf, tnn, true) --true update! + global_conf.dropout_rate = 0 + ppl_rec[iter] = {} + ppl_rec[iter].train = result:ppl_all("rnn") + --shuffling training file + nerv.printf("%s shuffling training file\n", global_conf.sche_log_pre) + os.execute('cp ' .. global_conf.train_fn_shuf .. ' ' .. global_conf.train_fn_shuf_bak) + os.execute('cat ' .. global_conf.train_fn_shuf_bak .. ' | sort -R --random-source=/dev/zero > ' .. global_conf.train_fn_shuf) + nerv.printf("===PEEK ON TEST %d===\n", iter) + result = LMTrainer.lm_process_file_rnn(global_conf, global_conf.test_fn, tnn, false) --false update! + ppl_rec[iter].test = result:ppl_all("rnn") + nerv.printf("===VALIDATION %d===\n", iter) + result = LMTrainer.lm_process_file_rnn(global_conf, global_conf.valid_fn, tnn, false) --false update! + ppl_rec[iter].valid = result:ppl_all("rnn") + ppl_rec[iter].lr = global_conf.lrate + if ((ppl_last / ppl_rec[iter].valid < global_conf.lr_decay or lr_half == true) and iter > global_conf.decay_iter) then + global_conf.lrate = (global_conf.lrate * 0.6) + end + if ppl_rec[iter].valid < ppl_last then + nerv.printf("%s PPL improves, saving net to file %s.%d...\n", global_conf.sche_log_pre, global_conf.param_fn, iter) + global_conf.paramRepo:export(global_conf.param_fn .. '.' .. tostring(iter), nil) + else + nerv.printf("%s PPL did not improve, rejected, copying param file of last iter...\n", global_conf.sche_log_pre) + os.execute('cp ' .. global_conf.param_fn..'.'..tostring(iter - 1) .. ' ' .. global_conf.param_fn..'.'..tostring(iter)) + end + if ppl_last / ppl_rec[iter].valid < global_conf.lr_decay or lr_half == true then + lr_half = true + end + if ppl_rec[iter].valid < ppl_last then + ppl_last = ppl_rec[iter].valid + end + nerv.printf("\n") + nerv.LMUtil.wait(2) + end + nerv.info("saving final nn to param.final") + os.execute('cp ' .. global_conf.param_fn .. '.' .. tostring(final_iter) .. ' ' .. global_conf.param_fn .. '.final') + + nerv.printf("===VALIDATION PPL record===\n") + for i, _ in pairs(ppl_rec) do + nerv.printf(" \n", i, ppl_rec[i].lr, ppl_rec[i].train, ppl_rec[i].valid, ppl_rec[i].test) + end + nerv.printf("\n") +end --if commands["train"] + +if commands["test"] == 1 then + nerv.printf("===FINAL TEST===\n") + global_conf.sche_log_pre = "[SCHEDULER FINAL_TEST]:" + if final_iter ~= -1 and test_iter == -1 then + test_iter = final_iter + end + if test_iter == -1 then + test_iter = "final" + end + tnn = load_net(global_conf, test_iter) + global_conf.dropout_rate = 0 + LMTrainer.lm_process_file_rnn(global_conf, global_conf.test_fn, tnn, false) --false update! +end --if commands["test"] + +if commands["testout"] == 1 then + nerv.printf("===TEST OUT===\n") + nerv.printf("q_file:\t%s\n", q_file) + local q_fn = qdata_dir .. q_file + global_conf.sche_log_pre = "[SCHEDULER FINAL_TEST]:" + if final_iter ~= -1 and test_iter == -1 then + test_iter = final_iter + end + if test_iter == -1 then + test_iter = "final" + end + tnn = load_net(global_conf, test_iter) + global_conf.dropout_rate = 0 + LMTrainer.lm_process_file_rnn(global_conf, q_fn, tnn, false) --false update! +end --if commands["testout"] + + diff --git a/nerv/layer/gate_fff.lua b/nerv/layer/gate_fff.lua index 751dde1..6082e27 100644 --- a/nerv/layer/gate_fff.lua +++ b/nerv/layer/gate_fff.lua @@ -1,36 +1,33 @@ -local GateFFFLayer = nerv.class('nerv.GateFFFLayer', 'nerv.Layer') +local GateFFFLayer = nerv.class('nerv.GateFLayer', 'nerv.Layer') --Full matrix gate function GateFFFLayer:__init(id, global_conf, layer_conf) self.id = id self.dim_in = layer_conf.dim_in self.dim_out = layer_conf.dim_out self.gconf = global_conf - - self.ltp1 = self:find_param("ltp1", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[1], self.dim_out[1]}) --layer_conf.ltp - self.ltp2 = self:find_param("ltp2", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[2], self.dim_out[1]}) --layer_conf.ltp - self.ltp3 = self:find_param("ltp3", layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[3], self.dim_out[1]}) --layer_conf.ltp + + for i = 1, #self.dim_in do + self["ltp" .. i] = self:find_param("ltp" .. i, layer_conf, global_conf, nerv.LinearTransParam, {self.dim_in[i], self.dim_out[1]}) --layer_conf.ltp + end self.bp = self:find_param("bp", layer_conf, global_conf, nerv.BiasParam, {1, self.dim_out[1]})--layer_conf.bp - self:check_dim_len(3, 1) -- exactly one input and one output + self:check_dim_len(-1, 1) --accept multiple inputs end function GateFFFLayer:init(batch_size) - if self.ltp1.trans:ncol() ~= self.bp.trans:ncol() or - self.ltp2.trans:ncol() ~= self.bp.trans:ncol() or - self.ltp3.trans:ncol() ~= self.bp.trans:ncol() then - nerv.error("mismatching dimensions of linear transform and bias paramter") - end - if self.dim_in[1] ~= self.ltp1.trans:nrow() or - self.dim_in[2] ~= self.ltp2.trans:nrow() or - self.dim_in[3] ~= self.ltp3.trans:nrow() then - nerv.error("mismatching dimensions of linear transform parameter and input") + for i = 1, #self.dim_in do + if self["ltp" .. i].trans:ncol() ~= self.bp.trans:ncol() then + nerv.error("mismatching dimensions of linear transform and bias paramter") + end + if self.dim_in[i] ~= self["ltp" .. i].trans:nrow() then + nerv.error("mismatching dimensions of linear transform parameter and input") + end + self["ltp"..i]:train_init() end + if self.dim_out[1] ~= self.ltp1.trans:ncol() then nerv.error("mismatching dimensions of linear transform parameter and output") end - self.ltp1:train_init() - self.ltp2:train_init() - self.ltp3:train_init() self.bp:train_init() self.err_bakm = self.gconf.cumat_type(batch_size, self.dim_out[1]) end @@ -44,8 +41,9 @@ end function GateFFFLayer:propagate(input, output) -- apply linear transform output[1]:mul(input[1], self.ltp1.trans, 1.0, 0.0, 'N', 'N') - output[1]:mul(input[2], self.ltp2.trans, 1.0, 1.0, 'N', 'N') - output[1]:mul(input[3], self.ltp3.trans, 1.0, 1.0, 'N', 'N') + for i = 2, #self.dim_in do + output[1]:mul(input[i], self["ltp" .. i].trans, 1.0, 1.0, 'N', 'N') + end -- add bias output[1]:add_row(self.bp.trans, 1.0) output[1]:sigmoid(output[1]) @@ -53,19 +51,23 @@ end function GateFFFLayer:back_propagate(bp_err, next_bp_err, input, output) self.err_bakm:sigmoid_grad(bp_err[1], output[1]) - next_bp_err[1]:mul(self.err_bakm, self.ltp1.trans, 1.0, 0.0, 'N', 'T') - next_bp_err[2]:mul(self.err_bakm, self.ltp2.trans, 1.0, 0.0, 'N', 'T') - next_bp_err[3]:mul(self.err_bakm, self.ltp3.trans, 1.0, 0.0, 'N', 'T') + for i = 1, #self.dim_in do + next_bp_err[i]:mul(self.err_bakm, self["ltp" .. i].trans, 1.0, 0.0, 'N', 'T') + end end function GateFFFLayer:update(bp_err, input, output) self.err_bakm:sigmoid_grad(bp_err[1], output[1]) - self.ltp1:update_by_err_input(self.err_bakm, input[1]) - self.ltp2:update_by_err_input(self.err_bakm, input[2]) - self.ltp3:update_by_err_input(self.err_bakm, input[3]) + for i = 1, #self.dim_in do + self["ltp" .. i]:update_by_err_input(self.err_bakm, input[i]) + end self.bp:update_by_gradient(self.err_bakm:colsum()) end function GateFFFLayer:get_params() - return nerv.ParamRepo({self.ltp1, self.ltp2, self.ltp3, self.bp}) + local pr = nerv.ParamRepo({self.bp}) + for i = 1, #self.dim_in do + pr:add(self["ltp" .. i].id, self["ltp" .. i]) + end + return pr end diff --git a/nerv/tnn/layersT/lstm_t.lua b/nerv/tnn/layersT/lstm_t.lua index ded6058..2a3342d 100644 --- a/nerv/tnn/layersT/lstm_t.lua +++ b/nerv/tnn/layersT/lstm_t.lua @@ -36,7 +36,7 @@ function LSTMLayerT:__init(id, global_conf, layer_conf) [ap("mainTanhL")] = {{}, {["dim_in"] = {self.dim_out[1]}, ["dim_out"] = {self.dim_out[1]}}}, [ap("outputTanhL")] = {{}, {["dim_in"] = {self.dim_out[1]}, ["dim_out"] = {self.dim_out[1]}}}, }, - ["nerv.GateFFFLayer"] = { + ["nerv.GateFLayer"] = { [ap("forgetGateL")] = {{}, {["dim_in"] = {self.dim_in[1], self.dim_in[2], self.dim_in[3]}, ["dim_out"] = {self.dim_in[3]}, ["pr"] = pr}}, [ap("inputGateL")] = {{}, {["dim_in"] = {self.dim_in[1], self.dim_in[2], self.dim_in[3]}, -- cgit v1.2.3-70-g09d2