Initial commit

training-scripts/dgs-baseline-rnn

@@ -0,0 +1,34 @@
+#!/bin/sh
+srun \
+-K \
+--container-mounts=/netscratch:/netscratch,/ds:/ds,$HOME:$HOME \
+--container-workdir=/netscratch/angelova/experiments/sl-translation \
+--container-image=/netscratch/avramidis/containers/lefterav+marian-nmt+1.10.0_sentencepiece_cuda-11.3.0_1.sqsh \
+--ntasks=1 \
+--nodes=1 \
+--gpus=4 \
+-p V100-16GB \
+/marian/build/marian train \
+--devices 0 1 2 3 \
+--model /netscratch/angelova/experiments/sl-translation/model/dgs_baseline/dgs_baseline.npz \
+--type s2s \
+--dim-rnn 1024 \
+--dim-emb 512 \
+--layer-normalization \
+--mini-batch 64 \
+--learn-rate 0.0005 \
+--optimizer adam \
+--early-stopping 10 \
+--cost-type=ce-mean-words \
+--valid-metrics bleu-detok perplexity ce-mean-words translation \
+--dim-vocabs 3000 3000 \
+--sentencepiece-options '--model_type=word --vocab_size=3000 --hard_vocab_limit=true' \
+--train-sets /netscratch/angelova/data/sl-translation/glosses_train.txt /netscratch/angelova/data/sl-translation/german_train.txt \
+--vocabs /netscratch/angelova/experiments/sl-translation/model/dgs_baseline/vocab.gl_baseline.spm /netscratch/angelova/experiments/sl-translation/model/dgs_baseline/vocab.de_baseline.spm \
+--valid-sets /netscratch/angelova/experiments/sl-translation/glosses_dev.txt /netscratch/angelova/experiments/sl-translation/german_dev.txt \
+--log /netscratch/angelova/experiments/sl-translation/model/dgs_baseline/valid_dgs_baseline.log \
+--lr-warmup 16000 \
+--keep-best \
+--valid-freq 400 \
+--disp-freq 1000 \
+--optimizer-params 0.9 0.98 1e-09 \

training-scripts/dgs-baseline-rnn-improved

@@ -0,0 +1,47 @@
+#!/bin/sh
+srun \
+-K \
+--container-mounts=/netscratch:/netscratch,/ds:/ds,$HOME:$HOME \
+--container-workdir=/netscratch/angelova/experiments/sl-translation \
+--container-image=/netscratch/avramidis/containers/lefterav+marian-nmt+1.10.0_sentencepiece_cuda-11.3.0_1.sqsh \
+--ntasks=1 \
+--nodes=1 \
+--gpus=4 \
+-p RTXA6000 \
+/marian/build/marian train \
+--devices 0 1 2 3 \
+--model /netscratch/angelova/experiments/sl-translation/model/dgs_nem_4/dgs_nem_4.npz \
+--type s2s \
+--dim-rnn 1024 \
+--dim-emb 512 \
+--enc-depth 1 \
+--dec-depth 2 \
+--enc-cell lstm \
+--enc-cell-depth 2 \
+--dec-cell-base-depth 2 \
+--dec-cell lstm \
+--layer-normalization \
+--dropout-rnn 0.5 \
+--dropout-src 0.4 \
+--dropout-trg 0.4 \
+--label-smoothing 0.2 \
+--max-length 200 \
+--mini-batch 32 \
+--learn-rate 0.0005 \
+--lr-report \
+--optimizer adam \
+--early-stopping 10 \
+--beam-size 5 \
+--cost-type=ce-mean-words \
+--valid-metrics bleu-detok perplexity ce-mean-words \
+--dim-vocabs 2600 2600 \
+--sentencepiece-options '--model_type=word --vocab_size=2600 --hard_vocab_limit=true' \
+--train-sets /netscratch/angelova/data/sl-translation/glosses_train.txt /netscratch/angelova/data/sl-translation/german_train.txt \
+--vocabs /netscratch/angelova/experiments/sl-translation/model/dgs_nem_4/vocab.gl_dgs_nem_4.spm /netscratch/angelova/experiments/sl-translation/model/dgs_nem_4/vocab.de_dgs_nem_4.spm \
+--valid-sets /netscratch/angelova/experiments/sl-translation/glosses_dev.txt /netscratch/angelova/experiments/sl-translation/german_dev.txt \
+--log /netscratch/angelova/experiments/sl-translation/model/dgs_nem_4/valid_dgs_nem_4.log \
+--lr-warmup 16000 \
+--keep-best \
+--valid-freq 500 \
+--disp-freq 1000 \
+--optimizer-params 0.9 0.98 1e-09 \

training-scripts/dgs-baseline-transformer

@@ -0,0 +1,33 @@
+#!/bin/sh
+srun \
+-K \
+--container-mounts=/netscratch:/netscratch,/ds:/ds,$HOME:$HOME \
+--container-workdir=/netscratch/angelova/experiments/sl-translation \
+--container-image=/netscratch/avramidis/containers/lefterav+marian-nmt+1.10.0_sentencepiece_cuda-11.3.0_1.sqsh \
+--ntasks=1 \
+--nodes=1 \
+--gpus=4 \
+-p RTX2080Ti \
+/marian/build/marian train \
+--devices 0 1 2 3 \
+--model /netscratch/angelova/experiments/sl-translation/model/dgs_baseline_transf/dgs_baseline_transf.npz \
+--type transformer \
+--dim-emb 512 \
+--layer-normalization \
+--mini-batch 64 \
+--learn-rate 0.0005 \
+--optimizer adam \
+--early-stopping 10 \
+--cost-type=ce-mean-words \
+--valid-metrics bleu-detok perplexity ce-mean-words translation \
+--dim-vocabs 3000 3000 \
+--sentencepiece-options '--model_type=word --vocab_size=3000 --hard_vocab_limit=true' \
+--train-sets /netscratch/angelova/data/sl-translation/glosses_train.txt /netscratch/angelova/data/sl-translation/german_train.txt \
+--vocabs /netscratch/angelova/experiments/sl-translation/model/dgs_baseline_transf/vocab.gl_baseline_transf.spm /netscratch/angelova/experiments/sl-translation/model/dgs_baseline_transf/vocab.de_baseline_transf.spm \
+--valid-sets /netscratch/angelova/experiments/sl-translation/glosses_dev.txt /netscratch/angelova/experiments/sl-translation/german_dev.txt \
+--log /netscratch/angelova/experiments/sl-translation/model/dgs_baseline_transf/valid_dgs_baseline_transf.log \
+--lr-warmup 16000 \
+--keep-best \
+--valid-freq 400 \
+--disp-freq 1000 \
+--optimizer-params 0.9 0.98 1e-09 \

training-scripts/dgs-baseline-transformer-improved

@@ -0,0 +1,42 @@
+#!/bin/sh
+srun \
+-K \
+--container-mounts=/netscratch:/netscratch,/ds:/ds,$HOME:$HOME \
+--container-workdir=/netscratch/angelova/experiments/sl-translation \
+--container-image=/netscratch/avramidis/containers/lefterav+marian-nmt+1.10.0_sentencepiece_cuda-11.3.0_1.sqsh \
+--ntasks=1 \
+--nodes=1 \
+--gpus=4 \
+-p RTX6000 \
+/marian/build/marian train \
+--devices 0 1 2 3 \
+--model /netscratch/angelova/experiments/sl-translation/model/dgs_transf_improved/dgs_transf_improved.npz \
+--type transformer \
+--dim-emb 512 \
+--enc-depth 1 \
+--dec-depth 2 \
+--layer-normalization \
+--dropout-rnn 0.5 \
+--dropout-src 0.4 \
+--dropout-trg 0.4 \
+--label-smoothing 0.2 \
+--max-length 200 \
+--mini-batch 32 \
+--learn-rate 0.0005 \
+--lr-report \
+--optimizer adam \
+--early-stopping 10 \
+--beam-size 5 \
+--cost-type=ce-mean-words \
+--valid-metrics bleu-detok perplexity ce-mean-words \
+--dim-vocabs 2600 2600 \
+--sentencepiece-options '--model_type=word --vocab_size=2600 --hard_vocab_limit=true' \
+--train-sets /netscratch/angelova/data/sl-translation/glosses_train.txt /netscratch/angelova/data/sl-translation/german_train.txt \
+--vocabs /netscratch/angelova/experiments/sl-translation/model/dgs_nem_4/vocab.gl_dgs_transf_improved.spm /netscratch/angelova/experiments/sl-translation/model/dgs_nem_4/vocab.de_dgs_transf_improved.spm \
+--valid-sets /netscratch/angelova/experiments/sl-translation/glosses_dev.txt /netscratch/angelova/experiments/sl-translation/german_dev.txt \
+--log /netscratch/angelova/experiments/sl-translation/model/dgs_transf_improved/valid_dgs_transf_improved.log \
+--lr-warmup 16000 \
+--keep-best \
+--valid-freq 500 \
+--disp-freq 1000 \
+--optimizer-params 0.9 0.98 1e-09 \

training-scripts/dgs-bpe-tok

@@ -0,0 +1,46 @@
+#!/bin/sh
+srun \
+-K \
+--container-mounts=/netscratch:/netscratch,/ds:/ds,$HOME:$HOME \
+--container-workdir=/netscratch/angelova/experiments/sl-translation \
+--container-image=/netscratch/avramidis/containers/lefterav+marian-nmt+1.10.0_sentencepiece_cuda-11.3.0_1.sqsh \
+--ntasks=1 \
+--nodes=1 \
+--gpus=4 \
+-p RTX6000 \
+/marian/build/marian train \
+--devices 0 1 2 3 \
+--model /netscratch/angelova/experiments/sl-translation/model/model_nematus_bpe_7.npz \
+--type s2s \
+--dim-rnn 1024 \
+--dim-emb 512 \
+--enc-depth 1 \
+--dec-depth 2 \
+--enc-cell lstm \
+--enc-cell-depth 2 \
+--dec-cell-base-depth 2 \
+--dec-cell lstm \
+--layer-normalization \
+--dropout-rnn 0.5 \
+--dropout-src 0.4 \
+--dropout-trg 0.4 \
+--label-smoothing 0.2 \
+--max-length 200 \
+--mini-batch 32 \
+--learn-rate 0.0005 \
+--optimizer adam \
+--early-stopping 10 \
+--beam-size 5 \
+--cost-type=ce-mean-words \
+--valid-metrics ce-mean-words perplexity translation bleu-detok \
+--dim-vocabs 2600 2600 \
+--sentencepiece-options '--model_type=bpe --vocab_size=2600 --hard_vocab_limit=true' \
+--train-sets /netscratch/angelova/data/sl-translation/glosses_train.txt /netscratch/angelova/data/sl-translation/sentences_train.txt \
+--vocabs /netscratch/angelova/experiments/sl-translation/model/vocab.gl_nem_bpe_7.spm /netscratch/angelova/experiments/sl-translation/model/vocab.de_nem_bpe_7.spm \
+--valid-sets /netscratch/angelova/experiments/sl-translation/glosses_dev.txt /netscratch/angelova/experiments/sl-translation/sentences_dev.txt \
+--log /netscratch/angelova/experiments/sl-translation/model/valid_nem_bpe_7.log \
+--lr-warmup 16000 \
+--keep-best \
+--valid-freq 500 \
+--disp-freq 1000 \
+--optimizer-params 0.9 0.98 1e-09 \

training-scripts/dgs-bpe-tok-stemmed

@@ -0,0 +1,46 @@
+#!/bin/sh
+srun \
+-K \
+--container-mounts=/netscratch:/netscratch,/ds:/ds,$HOME:$HOME \
+--container-workdir=/netscratch/angelova/experiments/sl-translation \
+--container-image=/netscratch/avramidis/containers/lefterav+marian-nmt+1.10.0_sentencepiece_cuda-11.3.0_1.sqsh \
+--ntasks=1 \
+--nodes=1 \
+--gpus=4 \
+-p RTX6000 \
+/marian/build/marian train \
+--devices 0 1 2 3 \
+--model /netscratch/angelova/experiments/sl-translation/model/dgs_nem_1.npz \
+--type s2s \
+--dim-rnn 1024 \
+--dim-emb 512 \
+--enc-depth 1 \
+--dec-depth 2 \
+--enc-cell lstm \
+--enc-cell-depth 2 \
+--dec-cell-base-depth 2 \
+--dec-cell lstm \
+--layer-normalization \
+--dropout-rnn 0.5 \
+--dropout-src 0.4 \
+--dropout-trg 0.4 \
+--label-smoothing 0.2 \
+--max-length 200 \
+--mini-batch 32 \
+--learn-rate 0.0005 \
+--optimizer adam \
+--early-stopping 10 \
+--beam-size 5 \
+--cost-type=ce-mean-words \
+--valid-metrics ce-mean-words perplexity translation bleu-detok \
+--dim-vocabs 2600 2600 \
+--sentencepiece-options '--model_type=bpe --vocab_size=2600 --hard_vocab_limit=true' \
+--train-sets /netscratch/angelova/data/sl-translation/glosses_stem_train.txt /netscratch/angelova/data/sl-translation/german_stripped_train.txt \
+--vocabs /netscratch/angelova/experiments/sl-translation/model/vocab.gl_dgs_nem_1.spm /netscratch/angelova/experiments/sl-translation/model/vocab.de_dgs_nem_1.spm \
+--valid-sets /netscratch/angelova/experiments/sl-translation/glosses_stem_dev.txt /netscratch/angelova/experiments/sl-translation/german_stripped_dev.txt \
+--log /netscratch/angelova/experiments/sl-translation/model/valid_dgs_nem_1.log \
+--lr-warmup 16000 \
+--keep-best \
+--valid-freq 400 \
+--disp-freq 1000 \
+--optimizer-params 0.9 0.98 1e-09 \

training-scripts/dgs-custom-tok

@@ -0,0 +1,47 @@
+#!/bin/sh
+srun \
+-K \
+--container-mounts=/netscratch:/netscratch,/ds:/ds,$HOME:$HOME \
+--container-workdir=/netscratch/angelova/experiments/sl-translation \
+--container-image=/netscratch/avramidis/containers/lefterav+marian-nmt+1.10.0_sentencepiece_cuda-11.3.0_1.sqsh \
+--ntasks=1 \
+--nodes=1 \
+--gpus=4 \
+-p V100-16GB \
+/marian/build/marian train \
+--devices 0 1 2 3 \
+--model /netscratch/angelova/experiments/sl-translation/model/dgs_nem_5/dgs_nem_5.npz \
+--type s2s \
+--dim-rnn 1024 \
+--dim-emb 512 \
+--enc-depth 1 \
+--dec-depth 2 \
+--enc-cell lstm \
+--enc-cell-depth 2 \
+--dec-cell-base-depth 2 \
+--dec-cell lstm \
+--layer-normalization \
+--dropout-rnn 0.5 \
+--dropout-src 0.4 \
+--dropout-trg 0.4 \
+--label-smoothing 0.2 \
+--max-length 200 \
+--mini-batch 32 \
+--learn-rate 0.0005 \
+--lr-report \
+--optimizer adam \
+--early-stopping 10 \
+--beam-size 5 \
+--cost-type=ce-mean-words \
+--valid-metrics bleu-detok perplexity ce-mean-words \
+--dim-vocabs 2600 2600 \
+--sentencepiece-options '--model_type=word --vocab_size=2600 --hard_vocab_limit=true' \
+--train-sets /netscratch/angelova/data/sl-translation/glosses_tok_train.txt /netscratch/angelova/data/sl-translation/german_train.txt \
+--vocabs /netscratch/angelova/experiments/sl-translation/model/dgs_nem_5/vocab.gl_dgs_nem_5.spm /netscratch/angelova/experiments/sl-translation/model/dgs_nem_5/vocab.de_dgs_nem_5.spm \
+--valid-sets /netscratch/angelova/experiments/sl-translation/glosses_tok_dev.txt /netscratch/angelova/experiments/sl-translation/german_dev.txt \
+--log /netscratch/angelova/experiments/sl-translation/model/dgs_nem_5/valid_dgs_nem_5.log \
+--lr-warmup 16000 \
+--keep-best \
+--valid-freq 400 \
+--disp-freq 1000 \
+--optimizer-params 0.9 0.98 1e-09 \

training-scripts/dgs-unigram-tok

@@ -0,0 +1,46 @@
+#!/bin/sh
+srun \
+-K \
+--container-mounts=/netscratch:/netscratch,/ds:/ds,$HOME:$HOME \
+--container-workdir=/netscratch/angelova/experiments/sl-translation \
+--container-image=/netscratch/avramidis/containers/lefterav+marian-nmt+1.10.0_sentencepiece_cuda-11.3.0_1.sqsh \
+--ntasks=1 \
+--nodes=1 \
+--gpus=4 \
+-p RTX6000 \
+/marian/build/marian train \
+--devices 0 1 2 3 \
+--model /netscratch/angelova/experiments/sl-translation/model/model_nematus_unigram_1.npz \
+--type s2s \
+--dim-rnn 1024 \
+--dim-emb 512 \
+--enc-depth 1 \
+--dec-depth 2 \
+--enc-cell lstm \
+--enc-cell-depth 2 \
+--dec-cell-base-depth 2 \
+--dec-cell lstm \
+--layer-normalization \
+--dropout-rnn 0.5 \
+--dropout-src 0.4 \
+--dropout-trg 0.4 \
+--label-smoothing 0.2 \
+--max-length 200 \
+--mini-batch 32 \
+--learn-rate 0.0005 \
+--optimizer adam \
+--early-stopping 10 \
+--beam-size 5 \
+--cost-type=ce-mean-words \
+--valid-metrics ce-mean-words perplexity translation bleu-detok \
+--dim-vocabs 2600 2600 \
+--sentencepiece-options '--model_type=unigram --vocab_size=2600 --hard_vocab_limit=true' \
+--train-sets /netscratch/angelova/data/sl-translation/glosses_train.txt /netscratch/angelova/data/sl-translation/sentences_train.txt \
+--vocabs /netscratch/angelova/experiments/sl-translation/model/vocab.gl_nem_unigram_1.spm /netscratch/angelova/experiments/sl-translation/model/vocab.de_nem_unigram_1.spm \
+--valid-sets /netscratch/angelova/experiments/sl-translation/glosses_dev.txt /netscratch/angelova/experiments/sl-translation/sentences_dev.txt \
+--log /netscratch/angelova/experiments/sl-translation/model/valid_nem_unigram_1.log \
+--lr-warmup 16000 \
+--keep-best \
+--valid-freq 500 \
+--disp-freq 1000 \
+--optimizer-params 0.9 0.98 1e-09 \

training-scripts/phoenix-baseline-rnn

@@ -0,0 +1,34 @@
+#!/bin/sh
+srun \
+-K \
+--container-mounts=/netscratch:/netscratch,/ds:/ds,$HOME:$HOME \
+--container-workdir=/netscratch/angelova/experiments/sl-translation \
+--container-image=/netscratch/avramidis/containers/lefterav+marian-nmt+1.10.0_sentencepiece_cuda-11.3.0_1.sqsh \
+--ntasks=1 \
+--nodes=1 \
+--gpus=4 \
+-p RTX2080Ti \
+/marian/build/marian train \
+--devices 0 1 2 3 \
+--model /netscratch/angelova/experiments/sl-translation/model/phoenix_baseline/phoenix_baseline.npz \
+--type s2s \
+--dim-rnn 1024 \
+--dim-emb 512 \
+--layer-normalization \
+--mini-batch 64 \
+--learn-rate 0.0005 \
+--optimizer adam \
+--early-stopping 10 \
+--cost-type=ce-mean-words \
+--valid-metrics bleu-detok perplexity ce-mean-words translation \
+--dim-vocabs 1000 1000 \
+--sentencepiece-options '--model_type=word --vocab_size=1000 --hard_vocab_limit=true' \
+--train-sets /netscratch/angelova/data/sl-translation/phoenix_train_glosses.txt /netscratch/angelova/data/sl-translation/phoenix_train_sentences.txt \
+--vocabs /netscratch/angelova/experiments/sl-translation/model/phoenix_baseline/vocab.ph_gl_baseline.spm /netscratch/angelova/experiments/sl-translation/model/phoenix_baseline/vocab.ph_de_baseline.spm \
+--valid-sets /netscratch/angelova/experiments/sl-translation/phoenix_dev_glosses.txt /netscratch/angelova/experiments/sl-translation/phoenix_dev_sentences.txt \
+--log /netscratch/angelova/experiments/sl-translation/model/phoenix_baseline/valid_ph_baseline.log \
+--lr-warmup 16000 \
+--keep-best \
+--valid-freq 400 \
+--disp-freq 1000 \
+--optimizer-params 0.9 0.98 1e-09 \