diff options
| author | Patrick Simianer <simianer@cl.uni-heidelberg.de> | 2012-04-26 21:39:11 +0200 |
|---|---|---|
| committer | Patrick Simianer <simianer@cl.uni-heidelberg.de> | 2012-04-26 21:39:11 +0200 |
| commit | 01110e92e7429df7882879e026b28aa9c89c724d (patch) | |
| tree | f5e03f63c8ae907696582aaa66953cc5cd911610 /dtrain/test | |
| parent | 28806638345e60bd442bf5fa8e7471f9115b0296 (diff) | |
made pair sampling configurable
Diffstat (limited to 'dtrain/test')
| -rw-r--r-- | dtrain/test/example/README | 6 | ||||
| -rw-r--r-- | dtrain/test/example/dtrain.ini | 9 |
2 files changed, 11 insertions, 4 deletions
diff --git a/dtrain/test/example/README b/dtrain/test/example/README new file mode 100644 index 00000000..e5a5de59 --- /dev/null +++ b/dtrain/test/example/README @@ -0,0 +1,6 @@ +Small example of input format for distributed training. +Call dtrain from cdec/dtrain/ with ./dtrain -c test/example/dtrain.ini . + +For this to work, disable '#define DTRAIN_LOCAL' from dtrain.h +and recompile. + diff --git a/dtrain/test/example/dtrain.ini b/dtrain/test/example/dtrain.ini index b59250f3..cd2c75e7 100644 --- a/dtrain/test/example/dtrain.ini +++ b/dtrain/test/example/dtrain.ini @@ -1,5 +1,5 @@ input=test/example/nc-wmt11.1k.gz # use '-' for STDIN -output=weights.gz # a weights file (add .gz for gzip compression) or STDOUT '-' +output=- # a weights file (add .gz for gzip compression) or STDOUT '-' decoder_config=test/example/cdec.ini # config for cdec # weights for these features will be printed on each iteration print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PhraseModel_5 PhraseModel_6 PassThrough @@ -10,11 +10,12 @@ stop_after=100 # stop epoch after 100 inputs epochs=3 # run over input 3 times k=100 # use 100best lists N=4 # optimize (approx) BLEU4 -scorer=stupid_bleu # use 'stupid' BLEU+1 +scorer=approx_bleu # use 'stupid' BLEU+1 learning_rate=0.0001 # learning rate gamma=0 # use SVM reg sample_from=kbest # use kbest lists (as opposed to forest) filter=uniq # only unique entries in kbest (surface form) -pair_sampling=108010 # 10 vs 80 vs 10 and 80 vs 10 +pair_sampling=XYX +hi_lo=0.1 # 10 vs 80 vs 10 and 80 vs 10 pair_threshold=0 # minimum distance in BLEU (this will still only use pairs with diff > 0) -select_weights=last # just output last weights +select_weights=VOID # don't output weights |