summaryrefslogtreecommitdiff
path: root/training/dtrain/examples/standard/dtrain.ini
diff options
context:
space:
mode:
Diffstat (limited to 'training/dtrain/examples/standard/dtrain.ini')
-rw-r--r--training/dtrain/examples/standard/dtrain.ini29
1 files changed, 6 insertions, 23 deletions
diff --git a/training/dtrain/examples/standard/dtrain.ini b/training/dtrain/examples/standard/dtrain.ini
index a515db02..f2698007 100644
--- a/training/dtrain/examples/standard/dtrain.ini
+++ b/training/dtrain/examples/standard/dtrain.ini
@@ -1,27 +1,10 @@
-#input=./nc-wmt11.de.gz
-#refs=./nc-wmt11.en.gz
-bitext=./nc-wmt11.gz
+bitext=./nc-wmt11.gz # input bitext
output=- # a weights file (add .gz for gzip compression) or STDOUT '-'
-select_weights=avg # output average (over epochs) weight vector
decoder_config=./cdec.ini # config for cdec
-# weights for these features will be printed on each iteration
+iterations=3 # run over input 3 times
+k=100 # use 100best lists
+N=4 # optimize (approx.) BLEU4
+learning_rate=0.1 # learning rate
+error_margin=1.0 # margin for margin perceptron
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= EgivenFCoherent SampleCountF CountEF MaxLexFgivenE MaxLexEgivenF IsSingletonF IsSingletonFE Glue WordPenalty PassThrough LanguageModel LanguageModel_OOV
-stop_after=10 # stop epoch after 10 inputs
-# interesting stuff
-epochs=3 # run over input 3 times
-k=100 # use 100best lists
-N=4 # optimize (approx) BLEU4
-scorer=fixed_stupid_bleu # use 'stupid' BLEU+1
-learning_rate=0.1 # learning rate, don't care if gamma=0 (perceptron) and loss_margin=0 (not margin perceptron)
-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=XYX #
-hi_lo=0.1 # 10 vs 80 vs 10 and 80 vs 10 here
-pair_threshold=0 # minimum distance in BLEU (here: > 0)
-loss_margin=0 # update if correctly ranked, but within this margin
-repeat=1 # repeat training on a kbest list 1 times
-#batch=true # batch tuning, update after accumulating over all sentences and all kbest lists