From 12771b3f0e18274db5cc634f06b94fb1597464df Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Thu, 25 Apr 2013 10:16:39 +0200 Subject: parallellized example --- training/dtrain/examples/parallelized/cdec.ini | 2 +- training/dtrain/examples/parallelized/work/out.0.0 | 9 +++++---- training/dtrain/examples/parallelized/work/out.0.1 | 9 +++++---- training/dtrain/examples/parallelized/work/out.1.0 | 9 +++++---- training/dtrain/examples/parallelized/work/out.1.1 | 9 +++++---- 5 files changed, 21 insertions(+), 17 deletions(-) (limited to 'training/dtrain/examples/parallelized') diff --git a/training/dtrain/examples/parallelized/cdec.ini b/training/dtrain/examples/parallelized/cdec.ini index e43ba1c4..5773029a 100644 --- a/training/dtrain/examples/parallelized/cdec.ini +++ b/training/dtrain/examples/parallelized/cdec.ini @@ -4,7 +4,7 @@ intersection_strategy=cube_pruning cubepruning_pop_limit=200 scfg_max_span_limit=15 feature_function=WordPenalty -feature_function=KLanguageModel ../example/nc-wmt11.en.srilm.gz +feature_function=KLanguageModel ../standard//nc-wmt11.en.srilm.gz #feature_function=ArityPenalty #feature_function=CMR2008ReorderingFeatures #feature_function=Dwarf diff --git a/training/dtrain/examples/parallelized/work/out.0.0 b/training/dtrain/examples/parallelized/work/out.0.0 index 7a00ed0f..c559dd4d 100644 --- a/training/dtrain/examples/parallelized/work/out.0.0 +++ b/training/dtrain/examples/parallelized/work/out.0.0 @@ -1,9 +1,9 @@ cdec cfg 'cdec.ini' Loading the LM will be faster if you build a binary file. -Reading ../example/nc-wmt11.en.srilm.gz +Reading ../standard//nc-wmt11.en.srilm.gz ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100 **************************************************************************************************** -Seeding random number sequence to 3121929377 +Seeding random number sequence to 405292278 dtrain Parameters: @@ -16,6 +16,7 @@ Parameters: learning rate 0.0001 gamma 0 loss margin 1 + faster perceptron 0 pairs 'XYX' hi lo 0.1 pair threshold 0 @@ -51,11 +52,11 @@ WEIGHTS non0 feature count: 12 avg list sz: 100 avg f count: 11.32 -(time 0.37 min, 4.4 s/S) +(time 0.35 min, 4.2 s/S) Writing weights file to 'work/weights.0.0' ... done --- Best iteration: 1 [SCORE 'stupid_bleu'=0.17521]. -This took 0.36667 min. +This took 0.35 min. diff --git a/training/dtrain/examples/parallelized/work/out.0.1 b/training/dtrain/examples/parallelized/work/out.0.1 index e2bd6649..8bc7ea9c 100644 --- a/training/dtrain/examples/parallelized/work/out.0.1 +++ b/training/dtrain/examples/parallelized/work/out.0.1 @@ -1,9 +1,9 @@ cdec cfg 'cdec.ini' Loading the LM will be faster if you build a binary file. -Reading ../example/nc-wmt11.en.srilm.gz +Reading ../standard//nc-wmt11.en.srilm.gz ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100 **************************************************************************************************** -Seeding random number sequence to 2767202922 +Seeding random number sequence to 43859692 dtrain Parameters: @@ -16,6 +16,7 @@ Parameters: learning rate 0.0001 gamma 0 loss margin 1 + faster perceptron 0 pairs 'XYX' hi lo 0.1 pair threshold 0 @@ -52,11 +53,11 @@ WEIGHTS non0 feature count: 12 avg list sz: 100 avg f count: 10.496 -(time 0.32 min, 3.8 s/S) +(time 0.35 min, 4.2 s/S) Writing weights file to 'work/weights.0.1' ... done --- Best iteration: 1 [SCORE 'stupid_bleu'=0.26638]. -This took 0.31667 min. +This took 0.35 min. diff --git a/training/dtrain/examples/parallelized/work/out.1.0 b/training/dtrain/examples/parallelized/work/out.1.0 index 6e790e38..65d1e7dc 100644 --- a/training/dtrain/examples/parallelized/work/out.1.0 +++ b/training/dtrain/examples/parallelized/work/out.1.0 @@ -1,9 +1,9 @@ cdec cfg 'cdec.ini' Loading the LM will be faster if you build a binary file. -Reading ../example/nc-wmt11.en.srilm.gz +Reading ../standard//nc-wmt11.en.srilm.gz ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100 **************************************************************************************************** -Seeding random number sequence to 1432415010 +Seeding random number sequence to 4126799437 dtrain Parameters: @@ -16,6 +16,7 @@ Parameters: learning rate 0.0001 gamma 0 loss margin 1 + faster perceptron 0 pairs 'XYX' hi lo 0.1 pair threshold 0 @@ -51,11 +52,11 @@ WEIGHTS non0 feature count: 11 avg list sz: 100 avg f count: 11.814 -(time 0.45 min, 5.4 s/S) +(time 0.43 min, 5.2 s/S) Writing weights file to 'work/weights.1.0' ... done --- Best iteration: 1 [SCORE 'stupid_bleu'=0.10863]. -This took 0.45 min. +This took 0.43333 min. diff --git a/training/dtrain/examples/parallelized/work/out.1.1 b/training/dtrain/examples/parallelized/work/out.1.1 index 0b984761..f479fbbc 100644 --- a/training/dtrain/examples/parallelized/work/out.1.1 +++ b/training/dtrain/examples/parallelized/work/out.1.1 @@ -1,9 +1,9 @@ cdec cfg 'cdec.ini' Loading the LM will be faster if you build a binary file. -Reading ../example/nc-wmt11.en.srilm.gz +Reading ../standard//nc-wmt11.en.srilm.gz ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100 **************************************************************************************************** -Seeding random number sequence to 1771918374 +Seeding random number sequence to 2112412848 dtrain Parameters: @@ -16,6 +16,7 @@ Parameters: learning rate 0.0001 gamma 0 loss margin 1 + faster perceptron 0 pairs 'XYX' hi lo 0.1 pair threshold 0 @@ -52,11 +53,11 @@ WEIGHTS non0 feature count: 12 avg list sz: 100 avg f count: 11.224 -(time 0.42 min, 5 s/S) +(time 0.45 min, 5.4 s/S) Writing weights file to 'work/weights.1.1' ... done --- Best iteration: 1 [SCORE 'stupid_bleu'=0.13169]. -This took 0.41667 min. +This took 0.45 min. -- cgit v1.2.3 From 5c4ef3f9206fd7e1ddfe252132582d854d62f0a4 Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Mon, 24 Jun 2013 17:45:32 +0200 Subject: documentation --- training/dtrain/README.md | 11 +++++++++++ training/dtrain/examples/parallelized/dtrain.ini | 2 -- 2 files changed, 11 insertions(+), 2 deletions(-) (limited to 'training/dtrain/examples/parallelized') diff --git a/training/dtrain/README.md b/training/dtrain/README.md index 2ab2f232..2bae6b48 100644 --- a/training/dtrain/README.md +++ b/training/dtrain/README.md @@ -17,6 +17,17 @@ To build only parts needed for dtrain do cd training/dtrain/; make ``` +Ideas +----- + * get approx_bleu to work? + * implement minibatches (Minibatch and Parallelization for Online Large Margin Structured Learning) + * learning rate 1/T? + * use an oracle? mira-like (model vs. BLEU), feature repr. of reference!? + * implement lc_bleu properly + * merge kbest lists of previous epochs (as MERT does) + * ``walk entire regularization path'' + * rerank after each update? + Running ------- See directories under test/ . diff --git a/training/dtrain/examples/parallelized/dtrain.ini b/training/dtrain/examples/parallelized/dtrain.ini index f19ef891..0b0932d6 100644 --- a/training/dtrain/examples/parallelized/dtrain.ini +++ b/training/dtrain/examples/parallelized/dtrain.ini @@ -11,6 +11,4 @@ pair_sampling=XYX hi_lo=0.1 select_weights=last print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough -# newer version of the grammar extractor use different feature names: -#print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough decoder_config=cdec.ini -- cgit v1.2.3