diff options
author | TianxingHe <htx_2006@hotmail.com> | 2015-12-04 15:19:44 +0800 |
---|---|---|
committer | TianxingHe <htx_2006@hotmail.com> | 2015-12-04 15:19:44 +0800 |
commit | af684cb95478fc38cc3d9f284b6b518a431c87e2 (patch) | |
tree | faab52eb3f6507331703b656c62a9e2ebf3b3f92 | |
parent | cbcce5ecc2864872b411eebbd307fa0f9a7e9dc0 (diff) | |
parent | 618450eb71817ded45c422f35d8fede2d52a66b2 (diff) |
Merge pull request #15 from cloudygoose/txh18/rnnlm
LSTM&TNN
40 files changed, 20627 insertions, 4060 deletions
@@ -7,7 +7,7 @@ luajit: luarocks: PREFIX=$(PREFIX) ./tools/build_luarocks.sh install: - cd nerv; $(PREFIX)/bin/luarocks make + cd nerv; $(PREFIX)/bin/luarocks make CFLAGS=$(CFLAGS) speech: cd speech/speech_utils; $(PREFIX)/bin/luarocks make cd speech/htk_io; $(PREFIX)/bin/luarocks make @@ -12,6 +12,7 @@ git clone https://github.com/Nerv-SJTU/nerv.git cd nerv git submodule init && git submodule update make +#To include some new CUDA feature(e.x. atomicCAS), use "make CFLAGS=-D__NERV_FUTURE_CUDA_7" #further, if you want the speech modules git clone https://github.com/Nerv-SJTU/nerv-speech.git speech diff --git a/nerv/Makefile b/nerv/Makefile index b449f82..5c329f9 100644 --- a/nerv/Makefile +++ b/nerv/Makefile @@ -7,7 +7,7 @@ INC_PATH := $(LUA_BINDIR)/../include/nerv LUA_DIR = $(INST_LUADIR)/nerv OBJ_DIR := $(BUILD_DIR)/objs ISUBDIR := io matrix luaT -SUBDIR := matrix io layer examples nn lib/io lib/luaT lib/matrix +SUBDIR := matrix io layer examples nn lib/io lib/luaT lib/matrix tnn/layersT INC_SUBDIR := $(addprefix $(INC_PATH)/,$(ISUBDIR)) OBJ_SUBDIR := $(addprefix $(OBJ_DIR)/,$(SUBDIR)) @@ -30,10 +30,13 @@ LUAT_OBJS := $(addprefix $(OBJ_DIR)/,$(LUAT_OBJS)) OBJS := $(CORE_OBJS) $(NERV_OBJS) $(LUAT_OBJS) LIBS := $(INST_LIBDIR)/libnerv.so $(LIB_PATH)/libnervcore.so $(LIB_PATH)/libluaT.so LUA_LIBS := matrix/init.lua io/init.lua init.lua \ - layer/init.lua layer/affine.lua layer/sigmoid.lua layer/softmax_ce.lua layer/softmax.lua \ + layer/init.lua layer/affine.lua layer/sigmoid.lua layer/tanh.lua layer/softmax_ce.lua layer/softmax.lua \ layer/window.lua layer/bias.lua layer/combiner.lua layer/mse.lua layer/affine_recurrent.lua \ + layer/elem_mul.lua layer/gate_fff.lua \ nn/init.lua nn/layer_repo.lua nn/param_repo.lua nn/layer_dag.lua \ - io/sgd_buffer.lua + io/sgd_buffer.lua \ + tnn/init.lua tnn/layer_dag_t.lua tnn/sutil.lua tnn/tnn.lua \ + tnn/layersT/dropout_t.lua tnn/layersT/lstm_t.lua tnn/layersT/softmax_ce_t.lua INCLUDE := -I $(LUA_INCDIR) -DLUA_USE_APICHECK #CUDA_BASE := /usr/local/cuda-7.0 @@ -41,7 +44,7 @@ CUDA_BASE := /usr/local/cuda CUDA_INCLUDE := -I $(CUDA_BASE)/include/ INCLUDE += $(CUDA_INCLUDE) -LDFLAGS := -L$(CUDA_BASE)/lib64/ -Wl,-rpath=$(CUDA_BASE)/lib64/ -lcudart -lcublas +LDFLAGS := -L$(CUDA_BASE)/lib64/ -Wl,-rpath=$(CUDA_BASE)/lib64/ -lcudart -lcublas -lcurand CFLAGS := -Wall -Wextra -O2 NVCC := $(CUDA_BASE)/bin/nvcc NVCC_FLAGS := -Xcompiler -fPIC,-Wall,-Wextra @@ -55,7 +58,7 @@ $(OBJ_DIR) $(LUA_DIR) $(OBJ_SUBDIR) $(LUA_SUBDIR) $(INC_SUBDIR): $(OBJ_DIR)/%.o: %.c $(patsubst /%.o,/%.c,$@) gcc -c -o $@ $< $(INCLUDE) -fPIC $(CFLAGS) $(OBJ_DIR)/lib/matrix/cukernel.o: lib/matrix/cukernel.cu - $(NVCC) -c -o $@ $< $(INCLUDE) $(NVCC_FLAGS) + $(NVCC) -c -o $@ $< $(INCLUDE) $(NVCC_FLAGS) $(CFLAGS) $(LUA_DIR)/%.lua: %.lua cp $< $@ diff --git a/nerv/examples/lmptb/lm_trainer.lua b/nerv/examples/lmptb/lm_trainer.lua index 44862dc..9ef4794 100644 --- a/nerv/examples/lmptb/lm_trainer.lua +++ b/nerv/examples/lmptb/lm_trainer.lua @@ -2,41 +2,55 @@ require 'lmptb.lmvocab' require 'lmptb.lmfeeder' require 'lmptb.lmutil' require 'lmptb.layer.init' -require 'rnn.init' +--require 'tnn.init' require 'lmptb.lmseqreader' local LMTrainer = nerv.class('nerv.LMTrainer') -local printf = nerv.printf +--local printf = nerv.printf + +--The bias param update in nerv don't have wcost added +function nerv.BiasParam:update_by_gradient(gradient) + local gconf = self.gconf + local l2 = 1 - gconf.lrate * gconf.wcost + self:_update_by_gradient(gradient, l2, l2) +end --Returns: LMResult -function LMTrainer.lm_process_file(global_conf, fn, tnn, do_train) +function LMTrainer.lm_process_file_rnn(global_conf, fn, tnn, do_train) local reader = nerv.LMSeqReader(global_conf, global_conf.batch_size, global_conf.chunk_size, global_conf.vocab) reader:open_file(fn) local result = nerv.LMResult(global_conf, global_conf.vocab) result:init("rnn") - + if global_conf.dropout_rate ~= nil then + nerv.info("LMTrainer.lm_process_file_rnn: dropout_rate is %f", global_conf.dropout_rate) + end + global_conf.timer:flush() tnn:flush_all() --caution: will also flush the inputs from the reader! local next_log_wcn = global_conf.log_w_num + local neto_bakm = global_conf.mmat_type(global_conf.batch_size, 1) --space backup matrix for network output while (1) do global_conf.timer:tic('most_out_loop_lmprocessfile') local r, feeds - - r, feeds = tnn:getFeedFromReader(reader) - if (r == false) then break end + global_conf.timer:tic('tnn_beforeprocess') + r, feeds = tnn:getfeed_from_reader(reader) + if r == false then + break + end for t = 1, global_conf.chunk_size do tnn.err_inputs_m[t][1]:fill(1) for i = 1, global_conf.batch_size do - if (bit.band(feeds.flags_now[t][i], nerv.TNN.FC.HAS_LABEL) == 0) then + if bit.band(feeds.flags_now[t][i], nerv.TNN.FC.HAS_LABEL) == 0 then tnn.err_inputs_m[t][1][i - 1][0] = 0 end end end + global_conf.timer:toc('tnn_beforeprocess') --[[ for j = 1, global_conf.chunk_size, 1 do @@ -50,29 +64,33 @@ function LMTrainer.lm_process_file(global_conf, fn, tnn, do_train) tnn:net_propagate() - if (do_train == true) then + if do_train == true then tnn:net_backpropagate(false) tnn:net_backpropagate(true) end - + + global_conf.timer:tic('tnn_afterprocess') for t = 1, global_conf.chunk_size, 1 do + tnn.outputs_m[t][1]:copy_toh(neto_bakm) for i = 1, global_conf.batch_size, 1 do if (feeds.labels_s[t][i] ~= global_conf.vocab.null_token) then - result:add("rnn", feeds.labels_s[t][i], math.exp(tnn.outputs_m[t][1][i - 1][0])) + --result:add("rnn", feeds.labels_s[t][i], math.exp(tnn.outputs_m[t][1][i - 1][0])) + result:add("rnn", feeds.labels_s[t][i], math.exp(neto_bakm[i - 1][0])) end end end + tnn:move_right_to_nextmb({0}) --only copy for time 0 + global_conf.timer:toc('tnn_afterprocess') - tnn:moveRightToNextMB() global_conf.timer:toc('most_out_loop_lmprocessfile') --print log - if (result["rnn"].cn_w > next_log_wcn) then + if result["rnn"].cn_w > next_log_wcn then next_log_wcn = next_log_wcn + global_conf.log_w_num - printf("%s %d words processed %s.\n", global_conf.sche_log_pre, result["rnn"].cn_w, os.date()) - printf("\t%s log prob per sample :%f.\n", global_conf.sche_log_pre, result:logp_sample("rnn")) + nerv.printf("%s %d words processed %s.\n", global_conf.sche_log_pre, result["rnn"].cn_w, os.date()) + nerv.printf("\t%s log prob per sample :%f.\n", global_conf.sche_log_pre, result:logp_sample("rnn")) for key, value in pairs(global_conf.timer.rec) do - printf("\t [global_conf.timer]: time spent on %s:%.5f clock time\n", key, value) + nerv.printf("\t [global_conf.timer]: time spent on %s:%.5f clock time\n", key, value) end global_conf.timer:flush() nerv.LMUtil.wait(0.1) @@ -90,9 +108,9 @@ function LMTrainer.lm_process_file(global_conf, fn, tnn, do_train) --break --debug end - printf("%s Displaying result:\n", global_conf.sche_log_pre) - printf("%s %s\n", global_conf.sche_log_pre, result:status("rnn")) - printf("%s Doing on %s end.\n", global_conf.sche_log_pre, fn) + nerv.printf("%s Displaying result:\n", global_conf.sche_log_pre) + nerv.printf("%s %s\n", global_conf.sche_log_pre, result:status("rnn")) + nerv.printf("%s Doing on %s end.\n", global_conf.sche_log_pre, fn) return result end diff --git a/nerv/examples/lmptb/lmptb/layer/lm_affine_recurrent.lua b/nerv/examples/lmptb/lmptb/layer/lm_affine_recurrent.lua index a5ecce1..c43e567 100644 --- a/nerv/examples/lmptb/lmptb/layer/lm_affine_recurrent.lua +++ b/nerv/examples/lmptb/lmptb/layer/lm_affine_recurrent.lua @@ -14,7 +14,7 @@ function LMRecurrent:propagate(input, output) output[1]:copy_fromd(input[1]) if (self.independent == true) then for i = 1, input[1]:nrow() do - if (self.gconf.input_word_id[self.id][0][i - 1] == self.break_id) then --here is sentence break + if (self.gconf.input_word_id[self.id][i - 1][0] == self.break_id) then --here is sentence break input[2][i - 1]:fill(0) end end diff --git a/nerv/examples/lmptb/lmptb/layer/select_linear.lua b/nerv/examples/lmptb/lmptb/layer/select_linear.lua index 672b7e2..580b9c5 100644 --- a/nerv/examples/lmptb/lmptb/layer/select_linear.lua +++ b/nerv/examples/lmptb/lmptb/layer/select_linear.lua @@ -10,9 +10,9 @@ function SL:__init(id, global_conf, layer_conf) self.dim_out = layer_conf.dim_out self.gconf = global_conf - self.ltp = layer_conf.ltp self.vocab = layer_conf.vocab - + self.ltp = self:find_param("ltp", layer_conf, global_conf, nerv.LinearTransParam, {self.vocab:size(), self.dim_out[1]}) --layer_conf.ltp + self:check_dim_len(1, 1) end @@ -30,12 +30,13 @@ function SL:init(batch_size) end function SL:update(bp_err, input, output) - --for i = 1, input[1]:ncol(), 1 do - -- if (input[1][0][i - 1] ~= 0) then - -- local word_vec = self.ltp.trans[input[1][0][i - 1]] - --word_vec:add(word_vec, bp_err[1][i - 1], 1, - self.gconf.lrate / self.gconf.batch_size) - -- end + --use this to produce reproducable result + --for i = 1, input[1]:nrow(), 1 do + -- local word_vec = self.ltp.trans[input[1][i - 1][0]] + -- word_vec:add(word_vec, bp_err[1][i - 1], 1, - self.gconf.lrate / self.gconf.batch_size) --end + + --I tried the update_select_rows kernel which uses atomicAdd, but it generates unreproducable result self.ltp.trans:update_select_rows(bp_err[1], input[1]:trans(), - self.gconf.lrate / self.gconf.batch_size, 0) self.ltp.trans:add(self.ltp.trans, self.ltp.trans, 1.0, - self.gconf.lrate * self.gconf.wcost / self.gconf.batch_size) end diff --git a/nerv/examples/lmptb/lmptb/lmseqreader.lua b/nerv/examples/lmptb/lmptb/lmseqreader.lua index e0dcd95..40471d5 100644 --- a/nerv/examples/lmptb/lmptb/lmseqreader.lua +++ b/nerv/examples/lmptb/lmptb/lmseqreader.lua @@ -1,4 +1,5 @@ require 'lmptb.lmvocab' +--require 'tnn.init' local LMReader = nerv.class("nerv.LMSeqReader") @@ -7,7 +8,7 @@ local printf = nerv.printf --global_conf: table --batch_size: int --vocab: nerv.LMVocab -function LMReader:__init(global_conf, batch_size, chunk_size, vocab) +function LMReader:__init(global_conf, batch_size, chunk_size, vocab, r_conf) self.gconf = global_conf self.fh = nil --file handle to read, nil means currently no file self.batch_size = batch_size @@ -15,6 +16,13 @@ function LMReader:__init(global_conf, batch_size, chunk_size, vocab) self.log_pre = "[LOG]LMSeqReader:" self.vocab = vocab self.streams = nil + if r_conf == nil then + r_conf = {} + end + self.se_mode = false --sentence end mode, when a sentence end is met, the stream after will be null + if r_conf.se_mode == true then + self.se_mode = true + end end --fn: string @@ -24,12 +32,21 @@ function LMReader:open_file(fn) nerv.error("%s error: in open_file(fn is %s), file handle not nil.", self.log_pre, fn) end printf("%s opening file %s...\n", self.log_pre, fn) - print("batch_size:", self.batch_size, "chunk_size", self.chunk_size) + print(self.log_pre, "batch_size:", self.batch_size, "chunk_size", self.chunk_size) + print(self.log_pre, "se_mode:", self.se_mode) self.fh = io.open(fn, "r") self.streams = {} for i = 1, self.batch_size, 1 do self.streams[i] = {["store"] = {}, ["head"] = 1, ["tail"] = 0} end + self.stat = {} --stat collected during file reading + self.stat.al_sen_start = true --check whether it's always sentence_start at the begining of a minibatch + self.bak_inputs_m = {} --backup MMatrix for temporary storey(then copy to TNN CuMatrix) + for j = 1, self.chunk_size, 1 do + self.bak_inputs_m[j] = {} + self.bak_inputs_m[j][1] = self.gconf.mmat_type(self.batch_size, 1) + --self.bak_inputs_m[j][2] = self.gconf.mmat_type(self.batch_size, self.vocab:size()) --since MMatrix does not yet have fill, this m[j][2] is not used + end end --id: int @@ -78,7 +95,7 @@ function LMReader:get_batch(feeds) local labels_s = feeds.labels_s for i = 1, self.chunk_size, 1 do inputs_s[i] = {} - labels_s[i] = {} + labels_s[i] = {} end local inputs_m = feeds.inputs_m --port 1 : word_id, port 2 : label @@ -86,45 +103,62 @@ function LMReader:get_batch(feeds) local flagsPack = feeds.flagsPack_now local got_new = false + for j = 1, self.chunk_size, 1 do + inputs_m[j][2]:fill(0) + end for i = 1, self.batch_size, 1 do local st = self.streams[i] + local end_stream = false --used for se_mode, indicating that this stream is ended for j = 1, self.chunk_size, 1 do flags[j][i] = 0 - self:refresh_stream(i) - if (st.store[st.head] ~= nil) then - inputs_s[j][i] = st.store[st.head] - inputs_m[j][1][i - 1][0] = self.vocab:get_word_str(st.store[st.head]).id - 1 - else + if end_stream == true then + if self.se_mode == false then + nerv.error("lmseqreader:getbatch: error, end_stream is true while se_mode is false") + end inputs_s[j][i] = self.vocab.null_token - inputs_m[j][1][i - 1][0] = 0 - end - inputs_m[j][2][i - 1]:fill(0) - if (st.store[st.head + 1] ~= nil) then - labels_s[j][i] = st.store[st.head + 1] - inputs_m[j][2][i - 1][self.vocab:get_word_str(st.store[st.head + 1]).id - 1] = 1 + self.bak_inputs_m[j][1][i - 1][0] = 0 + labels_s[j][i] = self.vocab.null_token else - if (inputs_s[j][i] ~= self.vocab.null_token) then - nerv.error("reader error : input not null but label is null_token") + self:refresh_stream(i) + if st.store[st.head] ~= nil then + inputs_s[j][i] = st.store[st.head] + --inputs_m[j][1][i - 1][0] = self.vocab:get_word_str(st.store[st.head]).id - 1 + self.bak_inputs_m[j][1][i - 1][0] = self.vocab:get_word_str(st.store[st.head]).id - 1 + else + inputs_s[j][i] = self.vocab.null_token + --inputs_m[j][1][i - 1][0] = 0 + self.bak_inputs_m[j][1][i - 1][0] = 0 end - labels_s[j][i] = self.vocab.null_token - end - if (inputs_s[j][i] ~= self.vocab.null_token) then - if (labels_s[j][i] == self.vocab.null_token) then - nerv.error("reader error : label is null while input is not null") + if st.store[st.head + 1] ~= nil then + labels_s[j][i] = st.store[st.head + 1] + inputs_m[j][2][i - 1][self.vocab:get_word_str(st.store[st.head + 1]).id - 1] = 1 + else + if (inputs_s[j][i] ~= self.vocab.null_token) then + nerv.error("reader error : input not null but label is null_token") + end + labels_s[j][i] = self.vocab.null_token end - flags[j][i] = bit.bor(flags[j][i], nerv.TNN.FC.SEQ_NORM) - got_new = true - st.store[st.head] = nil - st.head = st.head + 1 - if (labels_s[j][i] == self.vocab.sen_end_token) then - flags[j][i] = bit.bor(flags[j][i], nerv.TNN.FC.SEQ_END) - st.store[st.head] = nil --sentence end is passed + if inputs_s[j][i] ~= self.vocab.null_token then + if labels_s[j][i] == self.vocab.null_token then + nerv.error("reader error : label is null while input is not null") + end + flags[j][i] = bit.bor(flags[j][i], nerv.TNN.FC.SEQ_NORM) --has both input and label + got_new = true + st.store[st.head] = nil st.head = st.head + 1 - end - if (inputs_s[j][i] == self.vocab.sen_end_token) then - flags[j][i] = bit.bor(flags[j][i], nerv.TNN.FC.SEQ_START) - end - end + if labels_s[j][i] == self.vocab.sen_end_token then + flags[j][i] = bit.bor(flags[j][i], nerv.TNN.FC.SEQ_END) + st.store[st.head] = nil --sentence end is passed + st.head = st.head + 1 + if self.se_mode == true then + end_stream = true --meet sentence end, this stream ends now + end + end + if inputs_s[j][i] == self.vocab.sen_end_token then + flags[j][i] = bit.bor(flags[j][i], nerv.TNN.FC.SEQ_START) + end + end + end end end @@ -133,9 +167,20 @@ function LMReader:get_batch(feeds) for i = 1, self.batch_size, 1 do flagsPack[j] = bit.bor(flagsPack[j], flags[j][i]) end + inputs_m[j][1]:copy_fromh(self.bak_inputs_m[j][1]) end - if (got_new == false) then + --check for self.al_sen_start + for i = 1, self.batch_size do + if inputs_s[1][i] ~= self.vocab.sen_end_token and inputs_s[1][i] ~= self.vocab.null_token then + self.stat.al_sen_start = false + end + end + + if got_new == false then + nerv.info("lmseqreader file ends, printing stats...") + print("al_sen_start:", self.stat.al_sen_start) + return false else return true diff --git a/nerv/examples/lmptb/lmptb/lmutil.lua b/nerv/examples/lmptb/lmptb/lmutil.lua index 821aa94..71e8e17 100644 --- a/nerv/examples/lmptb/lmptb/lmutil.lua +++ b/nerv/examples/lmptb/lmptb/lmutil.lua @@ -48,15 +48,15 @@ end --Returns: nerv.MMatrixInt --Set the matrix to be ids of the words, id starting at 1, not 0 function Util.set_id(m, list, vocab) - if (m:ncol() ~= #list or m:nrow() ~= 1) then + if (m:nrow() ~= #list or m:ncol() ~= 1) then nerv.error("nrow of matrix mismatch with list or its col not one") end for i = 1, #list, 1 do --index in matrix starts at 0 if (list[i] ~= vocab.null_token) then - m[0][i - 1] = vocab:get_word_str(list[i]).id + m[i - 1][0] = vocab:get_word_str(list[i]).id else - m[0][i - 1] = 0 + m[i - 1][0] = 0 end end return m diff --git a/nerv/examples/lmptb/logs/LOG-tnn-h300 b/nerv/examples/lmptb/logs/LOG-tnn-h300 new file mode 100644 index 0000000..77fd612 --- /dev/null +++ b/nerv/examples/lmptb/logs/LOG-tnn-h300 @@ -0,0 +1,9372 @@ +Greetings +[SCHEDULER]: applying arg[2](global_conf.hidden_size=300)... +[SCHEDULER]: printing global_conf... +test_fn /home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/PTBdata/ptb.test.txt.adds +train_fn /home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/PTBdata/ptb.train.txt.adds +param_random function: 0x40237598 +train_fn_shuf_bak /home/slhome/txh18/workspace/nerv/play/dagL_test/train_fn_shuf_bak +decay_iter 16 +mmat_type table: 0x409280f8 +vocab_fn /home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/PTBdata/vocab +train_fn_shuf /home/slhome/txh18/workspace/nerv/play/dagL_test/train_fn_shuf +param_fn /home/slhome/txh18/workspace/nerv/play/dagL_test/params +log_w_num 40000 +work_dir /home/slhome/txh18/workspace/nerv/play/dagL_test +batch_size 10 +hidden_size 300 +timer nerv.Timer +sche_log_pre [SCHEDULER]: +nn_act_default 0 +max_iter 35 +valid_fn /home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/PTBdata/ptb.valid.txt.adds +lrate 1 +momentum 0 +wcost 1e-05 +chunk_size 15 +cumat_type table: 0x40935010 +[SCHEDULER]: printing training scheduling options... +lr_half false +start_iter -1 +ppl_last 100000 +[SCHEDULER]: printing training scheduling end. +[SCHEDULER]: creating work_dir... +[SCHEDULER]: building vocab... +[LOG]LMVocab: Vocab building on file /home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/PTBdata/vocab... +[LOG]LMVocab: Building finished, vocab size now is 10000. +[SCHEDULER]: preparing parameters... +[SCHEDULER]: first time, generating parameters... +===INITIAL VALIDATION=== +[SCHEDULER]: preparing parameters... +[SCHEDULER]: loading parameter from file /home/slhome/txh18/workspace/nerv/play/dagL_test/params.0... +reading chunk 0 from 0 +metadata: return {type="nerv.LinearTransParam",id="ltp_ih"} + +reading chunk 1 from 34510155 +metadata: return {type="nerv.LinearTransParam",id="ltp_hh"} + +reading chunk 2 from 35545278 +metadata: return {type="nerv.LinearTransParam",id="ltp_ho"} + +reading chunk 3 from 70045626 +metadata: return {type="nerv.BiasParam",id="bp_h"} + +reading chunk 4 from 70049129 +metadata: return {type="nerv.BiasParam",id="bp_o"} + +reading chunk 5 from 70164107 +[SCHEDULER]: preparing parameters end. +[SCHEDULER]: preparing layers... +(10:08:46 2015-11-18)[nerv] info: create layer: recurrentL1 +(10:08:46 2015-11-18)[nerv] info: create layer: sigmoidL1 +(10:08:46 2015-11-18)[nerv] info: create layer: combinerL1 +(10:08:46 2015-11-18)[nerv] info: create layer: outputL +(10:08:46 2015-11-18)[nerv] info: create layer: softmaxL +(10:08:46 2015-11-18)[nerv] info: create layer: selectL1 +[SCHEDULER]: preparing layers end. +[SCHEDULER]: Generate and initing TNN ... +<input> selectL1 0 +selectL1 recurrentL1 0 +recurrentL1 sigmoidL1 0 +sigmoidL1 combinerL1 0 +combinerL1 recurrentL1 1 +combinerL1 outputL 0 +outputL softmaxL 0 +<input> softmaxL 0 +softmaxL <output> 0 +recurrentL1 #dim_in: 2 #dim_out: 1 #i_conns_p: 2 #o_conns_p 1 +outputL #dim_in: 1 #dim_out: 1 #i_conns_p: 1 #o_conns_p 1 +combinerL1 #dim_in: 1 #dim_out: 2 #i_conns_p: 1 #o_conns_p 2 +sigmoidL1 #dim_in: 1 #dim_out: 1 #i_conns_p: 1 #o_conns_p 1 +selectL1 #dim_in: 1 #dim_out: 1 #i_conns_p: 0 #o_conns_p 1 +softmaxL #dim_in: 2 #dim_out: 1 #i_conns_p: 1 #o_conns_p 0 +TNN initing storage selectL1 -> recurrentL1 +TNN initing storage recurrentL1 -> sigmoidL1 +TNN initing storage sigmoidL1 -> combinerL1 +TNN initing storage combinerL1 -> recurrentL1 +TNN initing storage combinerL1 -> outputL +TNN initing storage outputL -> softmaxL +[SCHEDULER]: Initing TNN end. +[LOG]LMSeqReader: opening file /home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/PTBdata/ptb.valid.txt.adds... +batch_size: 10 chunk_size 15 +[SCHEDULER]: 40095 words processed Wed Nov 18 10:08:51 2015. + [SCHEDULER]: log prob per sample :-4.003230. + [global_conf.timer]: time spent on tnn_beforeprocess:0.47666 clock time + [global_conf.timer]: time spent on most_out_loop_lmprocessfile:4.10684 clock time + [global_conf.timer]: time spent on tnn_actual_layer_propagate:3.41463 clock time + [global_conf.timer]: time spent on tnn_afterprocess:0.12208 clock time +[LOG]LMSeqReader: file expires, closing. +[SCHEDULER]: Displaying result: +[SCHEDULER]: LMResult status of rnn: <SEN_CN 3370> <W_CN 70390> <PPL_NET 10545.667125173> <PPL_OOV 10092.059293004> <LOGP -295333.54956309> +[SCHEDULER]: Doing on /home/slhome/txh18/workspace/nerv/nerv/nerv/examples/lmptb/PTBdata/ptb.valid.txt.adds end. + +[SCHEDULER ITER1 LR1]: preparing parameters... +[SCHEDULER ITER1 LR1]: loading parameter from file /home/slhome/txh18/workspace/nerv/play/dagL_test/params.0... +reading chunk 0 from 0 +metadata: return {type="nerv.LinearTransParam",id="ltp_ih"} + +reading chunk 1 from 34510155 +metadata: return {type="nerv.LinearTransParam",id="ltp_hh"} + +reading chunk 2 from 35545278 +metadata: return {type="nerv.LinearTransParam",id="ltp_ho"} + +reading chunk 3 from 70045626 +metadata: return {type="nerv.BiasParam",id="bp_h"} + +reading chunk 4 from 70049129 +metadata: return {type="nerv.BiasParam",id="bp_o"} + +reading chunk 5 from 70164107 +[SCHEDULER ITER1 LR1]: preparing parameters end. +[SCHEDULER ITER1 LR1]: preparing layers... +(10:08:58 2015-11-18)[nerv] info: create layer: recurrentL1 +(10:08:58 2015-11-18)[nerv] info: create layer: sigmoidL1 +(10:08:58 2015-11-18 |