From 6e720b961f7edac9c3a41affe0ca40dd0ec9fc85 Mon Sep 17 00:00:00 2001 From: Determinant Date: Sun, 7 Jun 2015 11:55:09 +0800 Subject: fix memory leak in profiling; other minor changes --- examples/asr_trainer.lua | 20 +++++++++++++------- examples/swb_baseline.lua | 14 +++++++------- matrix/cuda_helper.h | 13 +++++-------- matrix/cumatrix.c | 1 + matrix/generic/cumatrix.c | 2 ++ matrix/generic/mmatrix.c | 6 ------ nerv.lua | 4 +++- 7 files changed, 31 insertions(+), 29 deletions(-) diff --git a/examples/asr_trainer.lua b/examples/asr_trainer.lua index d72b763..2993192 100644 --- a/examples/asr_trainer.lua +++ b/examples/asr_trainer.lua @@ -35,8 +35,9 @@ function build_trainer(ifname) print_stat(crit) if (not bp) and prefix ~= nil then nerv.info("writing back...") - local accu_cv = get_accuracy(crit) - cf = nerv.ChunkFile(prefix .. "_cv" .. accu_cv .. ".nerv", "w") + local fname = string.format("%s_cv%.3f.nerv", + prefix, get_accuracy(crit)) + cf = nerv.ChunkFile(fname, "w") for i, p in ipairs(network:get_params()) do cf:write_chunk(p) end @@ -53,7 +54,7 @@ halving_factor = 0.6 end_halving_inc = 0.1 min_iter = 1 max_iter = 20 -min_halving = 6 +min_halving = 5 gconf.batch_size = 256 gconf.buffer_size = 81920 @@ -65,11 +66,16 @@ local do_halving = false nerv.info("initial cross validation: %.3f", accu_best) for i = 1, max_iter do - nerv.info("iteration %d with lrate = %.6f", i, gconf.lrate) + nerv.info("[NN] begin iteration %d with lrate = %.6f", i, gconf.lrate) local accu_tr = trainer(nil, gconf.tr_scp, true) nerv.info("[TR] training set %d: %.3f", i, accu_tr) - local accu_new = trainer(pf0 .. "_iter" .. i .. "_tr" .. accu_tr, - gconf.cv_scp, false) + local accu_new = trainer( + string.format("%s_%s_iter_%d_lr%f_tr%.3f", + string.gsub(pf0, "(.*/)(.*)%..*", "%2"), + os.date("%Y%m%d%H%M%S"), + i, gconf.lrate, + accu_tr), + gconf.cv_scp, false) nerv.info("[CV] cross validation %d: %.3f", i, accu_new) -- TODO: revert the weights local accu_diff = accu_new - accu_best @@ -85,5 +91,5 @@ for i = 1, max_iter do if accu_new > accu_best then accu_best = accu_new end + nerv.Matrix.print_profile() end -nerv.Matrix.print_profile() diff --git a/examples/swb_baseline.lua b/examples/swb_baseline.lua index f536777..28cc6d5 100644 --- a/examples/swb_baseline.lua +++ b/examples/swb_baseline.lua @@ -6,8 +6,8 @@ gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9, 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", - global_transf = "global_transf.nerv", - initialized_param = "converted.nerv", + global_transf = "/slfs1/users/mfy43/swb_global_transf.nerv", + initialized_param = "/slfs1/users/mfy43/swb_init.nerv", debug = false} function make_param_repo(param_file) @@ -154,10 +154,10 @@ end function print_stat(crit) nerv.info("*** training stat begin ***") - nerv.utils.printf("cross entropy:\t%.8f\n", crit.total_ce) - nerv.utils.printf("correct:\t%d\n", crit.total_correct) - nerv.utils.printf("frames:\t%d\n", crit.total_frames) - nerv.utils.printf("err/frm:\t%.8f\n", crit.total_ce / crit.total_frames) - nerv.utils.printf("accuracy:\t%.3f%%\n", get_accuracy(crit)) + nerv.utils.printf("cross entropy:\t\t%.8f\n", crit.total_ce) + nerv.utils.printf("correct:\t\t%d\n", crit.total_correct) + nerv.utils.printf("frames:\t\t\t%d\n", crit.total_frames) + nerv.utils.printf("err/frm:\t\t%.8f\n", crit.total_ce / crit.total_frames) + nerv.utils.printf("accuracy:\t\t%.3f%%\n", get_accuracy(crit)) nerv.info("*** training stat end ***") end diff --git a/matrix/cuda_helper.h b/matrix/cuda_helper.h index d6effdb..fde6f18 100644 --- a/matrix/cuda_helper.h +++ b/matrix/cuda_helper.h @@ -62,17 +62,14 @@ static const char *cublasGetErrorString(cublasStatus_t err) { #define PROFILE_START \ do { \ - cudaEvent_t start, stop; \ - cudaEventCreate(&start); \ - cudaEventCreate(&stop); \ - cudaEventRecord(start, 0); + cudaEventRecord(profile_start, 0); #define PROFILE_STOP \ - cudaEventRecord(stop, 0); \ - cudaEventSynchronize(stop); \ + cudaEventRecord(profile_stop, 0); \ + cudaEventSynchronize(profile_stop); \ float milliseconds = 0; \ - cudaEventElapsedTime(&milliseconds, start, stop); \ + cudaEventElapsedTime(&milliseconds, profile_start, profile_stop); \ accu_profile(__func__, milliseconds / 1000); \ } while (0); -#define PROFILE_END +#define PROFILE_END #endif diff --git a/matrix/cumatrix.c b/matrix/cumatrix.c index 4ebc5ff..ee5ecaa 100644 --- a/matrix/cumatrix.c +++ b/matrix/cumatrix.c @@ -2,6 +2,7 @@ #include "../common.h" #include "cuda_helper.h" static cublasHandle_t cublas_handle; +static cudaEvent_t profile_start, profile_stop; static HashMap *profile; int print_profile(lua_State *L) { diff --git a/matrix/generic/cumatrix.c b/matrix/generic/cumatrix.c index 956e1e6..a340aef 100644 --- a/matrix/generic/cumatrix.c +++ b/matrix/generic/cumatrix.c @@ -443,6 +443,8 @@ static const luaL_Reg nerv_matrix_(extra_methods)[] = { static void cuda_matrix_(init)(lua_State *L) { luaN_append_methods(L, nerv_matrix_(extra_methods)); cublasCreate(&cublas_handle); + cudaEventCreate(&profile_start); + cudaEventCreate(&profile_stop); profile = hashmap_create(PROFILE_HASHMAP_SIZE, bkdr_hash, strcmp); } diff --git a/matrix/generic/mmatrix.c b/matrix/generic/mmatrix.c index 2045d65..b0f0791 100644 --- a/matrix/generic/mmatrix.c +++ b/matrix/generic/mmatrix.c @@ -87,15 +87,9 @@ int nerv_matrix_(save)(lua_State *L) { MATRIX_ELEM *row = MATRIX_ROW_PTR(self, i); for (j = 0; j < ncol; j++) if (fprintf(fp, MATRIX_ELEM_WRITE_FMT " ", row[j]) < 0) - { - free(self); return 0; - } if (fprintf(fp, "\n") < 0) - { - free(self); return 0; - } } return 0; } diff --git a/nerv.lua b/nerv.lua index ce6bc44..467d926 100644 --- a/nerv.lua +++ b/nerv.lua @@ -10,7 +10,9 @@ function nerv.error_method_not_implemented() end function nerv.info(fmt, ...) - nerv.utils.printf("[nerv] info: " .. fmt .. "\n", ...) + nerv.utils.printf( + string.format("(%s)[nerv] info: %s\n", + os.date("%H:%M:%S %F"), fmt), ...) end -- Torch C API wrapper -- cgit v1.2.3-70-g09d2