1) use the last iteration of coordinate descent, 2) fix variational i…

…nference and dirichlet priors, 3) less logging.

alignment/train-latentCrfAligner.cc

@@ -395,10 +395,8 @@ void endOfKIterationsCallbackFunction() {
   LatentCrfAligner &aligner = *( (LatentCrfAligner*) model );
 
   // fix learningInfo.test_size
-  cerr << "firstKExamplesToLabel = " << aligner.learningInfo.firstKExamplesToLabel << endl;
   if(aligner.learningInfo.firstKExamplesToLabel <= 0) {
     aligner.learningInfo.firstKExamplesToLabel = aligner.examplesCount;
-    cerr << "firstKExamplesToLabel = " << aligner.learningInfo.firstKExamplesToLabel << endl;
   }
 
   // find viterbi alignment for the top K examples of the training set (i.e. our test set)

core/LatentCrfModel.h

@@ -187,6 +187,9 @@ class LatentCrfModel : public UnsupervisedSequenceTaggingModel {
   void ReduceMleAndMarginals(MultinomialParams::ConditionalMultinomialParam<int64_t> &mleGivenOneLabel,
                              boost::unordered_map<int64_t, double> &mleMarginalsGivenOneLabel);
 
+  void AllReduceMleAndMarginals(MultinomialParams::ConditionalMultinomialParam<int64_t> &mleGivenOneLabel,
+                                boost::unordered_map<int64_t, double> &mleMarginalsGivenOneLabel);
+
   void BroadcastTheta(unsigned rankId);
 
   // filenames

ducttape-files/en-fi-exp000.tconf

@@ -0,0 +1,24 @@
+import en-fi.tconf
+
+global {
+    # Output directory:
+    ducttape_output="/usr3/home/wammar/mt-systems/en-fi-exp000"
+
+    # initial autoencoder params
+    init_theta_fwd=""
+    init_lambda_fwd="" 
+    init_theta_bwd=""
+    init_lambda_bwd=""
+
+    # general
+    cores=16
+
+    # prefix; must be important for concurrent runs
+    output_prefix="exp000"
+
+}
+
+plan Full {
+#reach AutoencoderAlignT2S
+    reach Evaluate via (TuneOrTest: tune test) * (UseCustomLM: no) * (Aligner: giza) * (L2: point_o_one) * (DirichletAlpha: one_point_five) * (PrecomputedFeatures: dyer11) * (OptimizeLambdasFirst: yes) * (UseOtherAligners: yes) * (EmItercount: one) * (SymHeuristic: grow_diag_final_and ) * (MaxLambdaEpochCount: one) * (LambdaOptimizer: lbfgs sgd) * (ThetaOptimizer: em)
+}

ducttape-files/en-fi-paths.tconf

@@ -0,0 +1,37 @@
+global {
+
+    train_corpus="/usr3/home/wammar/corpora/parallel/en-fi.short.en-fi"
+    tune_corpus="/usr3/home/wammar/corpora/parallel/en-fi.wmt-dev"
+    test_corpus="/usr3/home/wammar/corpora/parallel/en-fi.wmt-devtest"
+    src_brown_clusters="/usr1/home/wammar/parallel/english/news-commentary10.cz-en.en.tok.brown80"
+    tgt_brown_clusters="/usr3/home/wammar/brown-clusters/wmt_mono+parallel+dev+devtest-c100-p1.out/paths"
+
+    # only specify when you want to reuse a previously built LM 
+    language_model=""
+    # only use these two parameters if you want to build a language model. if you have a language model already built, specify "language_model=" instead
+    lm_order=4
+    lm_data="/usr3/home/wammar/corpora/monolingual/finnish/wmt_mono+parallel+dev+devtest.tok"
+
+    # tool paths
+    cdec_dir="/home/wammar/cdec/"
+    multeval="/home/wammar/git/multeval/multeval.sh"
+    giza_bin="/opt/tools/mgizapp-0.7.2/bin"
+    moses_train_script="/home/wammar/git/mosesdecoder/scripts/training/train-model.perl"
+    mkcls_bin="/mal0/tools/mosesdecoder/bin/mkcls"
+    wammar_utils_dir="/home/wammar/wammar-utils"
+    alignment_with_openfst_dir="/home/wammar/online_em/alignment-with-openfst/"
+    kenlm_dir="/home/wammar/git/kenlm"
+
+    # aer
+    conv_pharaoh_script="/home/wammar/alignment-with-openfst/data/hansards/conv-pharaoh.pl"
+    aer_eval_script=""
+    gold_alignment=""
+
+    # other aligner outputs
+    fwd_giza_alignments=""
+    bwd_giza_alignments=""
+    sym_giza_alignments=""
+    fwd_fast_alignments=""
+    bwd_fast_alignments=""
+    sym_fast_alignments=""
+}

ducttape-files/en-fi.tconf

@@ -0,0 +1,23 @@
+import en-fi-paths.tconf
+
+global {
+
+    # UseCustomAlignment controls whether to retrain word alignment from the train_corpus, or just use 
+    # alignments at 'alignment'
+    alignment="/dev/null"
+
+    # preprocess data
+    tokenize_corpus="false"
+    lowercase_corpus="false"
+    max_sentence_length=0
+
+    # language specific
+    src="en"
+    tgt="fi"
+    meteor_lang="english"
+
+    # aer
+    test_sents_count=0
+
+}
+

ducttape-files/hiero-local.tape

@@ -22,6 +22,7 @@ global {
                       intersect="intersect"
                       union="union")
   lambda_optimizer=(LambdaOptimizer: sgd="sgd" lbfgs="lbfgs")
+  theta_optimizer=(ThetaOptimizer: em="em" online_em="online_em")
 }
 
 #import ../submitters.tape
@@ -239,6 +240,7 @@ task AutoencoderAlignS2T
     :: tgt_brown_clusters=@
     :: output_prefix=@
     :: lambda_optimizer=@
+    :: theta_optimizer=@
     < executable=$executable@BuildLatentCrfAligner
     < wordpair_feats_file=$wordpair_feats_file@GenerateWordpairFeatsS2T 
     > alignment 
@@ -257,6 +259,7 @@ task AutoencoderAlignS2T
   --test-size $test_sents_count
   --max-model1-iter-count $model1_itercount
   --lambda-optimizer $lambda_optimizer
+  --theta-optimizer $theta_optimizer
   --max-iter-count 10"
 
   # --feat SRC0_TGT0
@@ -369,6 +372,7 @@ task AutoencoderAlignT2S
     :: use_src_bigrams=@
     :: src_brown_clusters=@
     :: lambda_optimizer=@
+    :: theta_optimizer=@
     < wordpair_feats_file=$wordpair_feats_file@GenerateWordpairFeatsT2S 
     < executable=$executable@BuildLatentCrfAligner
     > alignment 
@@ -377,7 +381,7 @@ task AutoencoderAlignT2S
 {
 
   model1_itercount="5"
-  variational="true"
+  variational=""
 
   # the latent-CRF word alignment mode
   command="mpirun -np $procs $executable 
@@ -388,7 +392,9 @@ task AutoencoderAlignT2S
   --reverse true
   --max-model1-iter-count $model1_itercount
   --lambda-optimizer $lambda_optimizer 
+  --theta-optimizer $theta_optimizer
   --max-iter-count 10"
+
   #--tgt-word-classes-filename $src_brown_clusters
   #   --feat SRC0_TGT0
   #   --feat SYNC_START --feat SYNC_END
@@ -579,7 +585,6 @@ task Tune
    # :: .submitter=torque_normal .walltime="48:00:00" .cpus=32 .vmem=60g .q=normal 
 {
 
-
   ref_count=$(head -n 1 $tune_set | grep -o '|||' | wc -l)
   #cut -f 1 -d '|' $tune_set > ./tune_source
   for i in `seq 1 $ref_count`; do
@@ -606,8 +611,8 @@ task Tune
   echo "Glue 0.0"              >> $initial_weights
   echo "WordPenalty 0.0"       >> $initial_weights
 
-  $cdec_dir/training/mira/mira.py -j $cores --update-size 250 -o $mira_work --weights ./initial_weights \
-    --devset $tune_set --config $cdec_ini
+  $cdec_dir/training/mira/mira.py --jobs $cores --kbest-size 500 -o $mira_work --weights ./initial_weights \
+    -d $tune_set --config $cdec_ini --step-size 0.001
 
   index=$(ls $mira_work | grep '^weights\.[0-9]*$' | sed 's/^weights\.//' | sort -n | tail -n 1)
   ln -s $mira_work/weights.$index $optimized_weights