added sample scripts

scripts/compute-perp.sh

@@ -1,3 +1,8 @@
+# get the scores from n-best list
+# assume that the n-best list should follow the format: 
+# line in source language ||| line in target language
+# and all sentence pairs are wrapped up with <s>, </s>, <unk> markers
+./attentional -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -T ../sample-data/50best.de-en.unk.cap -i ./work/exp/params.de-en.AM.l_2_h_512_a_256_lstm_bidirectional --lstm --bidirectional -l 2 -h 512 -a 256 --rescore &>./work/res/scores.50best.de-en.AM.l_2_h_512_a_256_lstm_bidirectional & 
 
----
+# note that if using advanced models (with alignment features), please follow the usage of extra params in train.sh
 

scripts/decode.sh

@@ -1,3 +1,10 @@
+# Beam Search decoding
+./attentional --beam 5 -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -T ../sample-data/test.de-en.unk.cap -i ./work/exp/params.de-en.AM.l_2_h_512_a_256_lstm_bidirectional --lstm --bidirectional -l 2 -h 512 -a 256 &>./work/res/translation.beam.de-en.AM.l_2_h_512_a_256_lstm_bidirectional & #with beam_width=5
+
+# Greedy Viterbi decoding
+./attentional -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -T ../sample-data/test.de-en.unk.cap -i ./work/exp/params.de-en.AM.l_2_h_512_a_256_lstm_bidirectional --lstm --bidirectional -l 2 -h 512 -a 256 &>./work/res/translation.greedy.de-en.AM.l_2_h_512_a_256_lstm_bidirectional &
+
+# Sampling decoding
+
 
----
 

scripts/train.sh

@@ -1,3 +1,37 @@
+## train vanilla attentional models
+
+# print help (no params)
+./attentional
+
+# forward model
+./attentional -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -p ./work/exp/params.de-en.AM.l_2_h_512_a_256_lstm_bidirectional --lstm --bidirectional -l 2 -h 512 -a 256 -v &>/work/exp/log.de-en.AM.train.l_2_h_512_a_256_lstm_bidirectional & 
+
+# backward model (simply add '--swap' param)
+./attentional --swap -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -p /work/exp/params.en-de.AM.l_2_h_512_a_256_lstm_bidirectional --lstm --bidirectional -l 2 -h 512 -a 256 -v &>/work/exp/log.en-de.AM.train.l_2_h_512_a_256_lstm_bidirectional & 
+
+## train attentional model with alignment features (including positional bias, Markov conditioning, and global/local fertilities)
+
+# with positional bias (simple add '--gz-position' param)
+./attentional --gz-position -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -p /work/exp/params.en-de.AM.l_2_h_512_a_256_lstm_bidirectional_giza-position --lstm --bidirectional -l 2 -h 512 -a 256 -v &>/work/exp/log.en-de.AM.train.l_2_h_512_a_256_lstm_bidirectional_giza-position &
+
+# with Markov condition (simple add '--gz-condition' param)
+./attentional --gz-markov -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -p /work/exp/params.en-de.AM.l_2_h_512_a_256_lstm_bidirectional_giza-markov --lstm --bidirectional -l 2 -h 512 -a 256 -v &>/work/exp/log.en-de.AM.train.l_2_h_512_a_256_lstm_bidirectional_giza-markov &
+
+# with local fertility (simple add '--gz-fertility' param)
+./attentional --gz-fertility -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -p /work/exp/params.en-de.AM.l_2_h_512_a_256_lstm_bidirectional_giza-fertility --lstm --bidirectional -l 2 -h 512 -a 256 -v &>/work/exp/log.en-de.AM.train.l_2_h_512_a_256_lstm_bidirectional_giza-fertility &
+
+# with all the above features (simple add '--giza' param)
+./attentional --giza -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -p /work/exp/params.en-de.AM.l_2_h_512_a_256_lstm_bidirectional_giza-all --lstm --bidirectional -l 2 -h 512 -a 256 -v &>/work/exp/log.en-de.AM.train.l_2_h_512_a_256_lstm_bidirectional_giza-all &
+
+## train attentional model with global fertility
+# First, we need a pre-trained model with (better) or without alignment features using the commands mentioned earlier
+# Next, let's train the model with global fertility (simple add '--fertility' param)
+./attentional --giza --fertility -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -i /work/exp/params.en-de.AM.l_2_h_512_a_256_lstm_bidirectional_giza-all --lstm --bidirectional -l 2 -h 512 -a 256 -v &>/work/exp/log.en-de.AM.train.l_2_h_512_a_256_lstm_bidirectional_giza-all_continued_fertility &
+
+# train bi-attentional model
+# First, we need pre-trained models with both directions (forward and backward)
+# e.g. forward model: ./work/exp/params.de-en.AM.l_2_h_512_a_256_lstm_bidirectional
+# and backward model: ./work/exp/params.en-de.AM.l_2_h_512_a_256_lstm_bidirectional
+./biattentional -t ../sample-data/train.de-en.unk.cap -d ../sample-data/dev.de-en.unk.cap -i ./work/exp/params.de-en.AM.l_2_h_512_a_256_lstm_bidirectional -i /work/exp/params.en-de.AM.l_2_h_512_a_256_lstm_bidirectional -p /work/exp/params.de-en_en-de.AM.l_2_h_512_a_256_lstm_bidirectional_symmetry --lstm --bidirectional -l 2 -h 512 -a 256 -v &>/work/exp/log.de-en_en-de.AM.train.l_2_h_512_a_256_lstm_bidirectional_symmetry &
 
----