diff options
| author | Roman Grundkiewicz <rgrundki@exseed.ed.ac.uk> | 2020-11-10 11:28:53 +0300 |
|---|---|---|
| committer | Roman Grundkiewicz <rgrundki@exseed.ed.ac.uk> | 2020-11-10 11:28:53 +0300 |
| commit | 29362e03f56ec2403b601f549c084bd84ab0c9ef (patch) | |
| tree | fb61d460dcacf4f68f267d1ea06758eb749e17ec /tests | |
| parent | 01c568388c2441f722f04a803f3ab76c03ba9f76 (diff) | |
| parent | 5b4adeb50b91a0ef62fc348024632330284415b1 (diff) | |
Merge branch 'master' into quant-test
Diffstat (limited to 'tests')
87 files changed, 198 insertions, 152 deletions
diff --git a/tests/interface/input-tsv/test_tsv_train.sh b/tests/interface/input-tsv/test_tsv_train.sh index 8685004..72b87d8 100644 --- a/tests/interface/input-tsv/test_tsv_train.sh +++ b/tests/interface/input-tsv/test_tsv_train.sh @@ -14,7 +14,7 @@ mkdir -p train # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ -m train/model.npz --tsv -t train.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 10 --disp-freq 2 \ --log train.log diff --git a/tests/interface/input-tsv/test_tsv_train_assume_stdin.sh b/tests/interface/input-tsv/test_tsv_train_assume_stdin.sh index 7fcf2d3..87f8c7c 100644 --- a/tests/interface/input-tsv/test_tsv_train_assume_stdin.sh +++ b/tests/interface/input-tsv/test_tsv_train_assume_stdin.sh @@ -14,7 +14,7 @@ mkdir -p train_stdin2 # Run marian command cat train.tsv | $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ -m train_stdin2/model.npz -t stdin -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 10 --disp-freq 2 \ --log train_stdin2.log diff --git a/tests/interface/input-tsv/test_tsv_train_create_vocab_joint.sh b/tests/interface/input-tsv/test_tsv_train_create_vocab_joint.sh index cf224eb..88b5b62 100644 --- a/tests/interface/input-tsv/test_tsv_train_create_vocab_joint.sh +++ b/tests/interface/input-tsv/test_tsv_train_create_vocab_joint.sh @@ -14,7 +14,7 @@ mkdir -p train_vocab # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ -m train_vocab/model.npz --tsv -t train.tsv -v train_vocab/vocab.spm train_vocab/vocab.spm --dim-vocabs 2000 2000 -T train_vocab \ --after-batches 20 --disp-freq 2 \ --log train_vocab.log diff --git a/tests/interface/input-tsv/test_tsv_train_create_vocabs.sh b/tests/interface/input-tsv/test_tsv_train_create_vocabs.sh index f163900..a23e7ac 100644 --- a/tests/interface/input-tsv/test_tsv_train_create_vocabs.sh +++ b/tests/interface/input-tsv/test_tsv_train_create_vocabs.sh @@ -14,7 +14,7 @@ mkdir -p train_vocabs # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ -m train_vocabs/model.npz --tsv -t train.tsv -v train_vocabs/vocab.de.spm train_vocabs/vocab.en.spm --dim-vocabs 2000 2000 -T train_vocabs \ --after-batches 20 --disp-freq 2 \ --log train_vocabs.log diff --git a/tests/interface/input-tsv/test_tsv_train_create_vocabs_yml.sh b/tests/interface/input-tsv/test_tsv_train_create_vocabs_yml.sh index d627c43..26d20a6 100644 --- a/tests/interface/input-tsv/test_tsv_train_create_vocabs_yml.sh +++ b/tests/interface/input-tsv/test_tsv_train_create_vocabs_yml.sh @@ -14,7 +14,7 @@ mkdir -p train_vocabs_yml # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ -m train_vocabs_yml/model.npz --tsv -t train.bpe.tsv -v train_vocabs_yml/vocab.de.yml train_vocabs_yml/vocab.en.yml --dim-vocabs 2000 2000 -T train_vocabs_yml \ --after-batches 20 --disp-freq 2 \ --log train_vocabs_yml.log diff --git a/tests/interface/input-tsv/test_tsv_train_inputtypes_stdin.sh b/tests/interface/input-tsv/test_tsv_train_inputtypes_stdin.sh index 04831ef..07f2eab 100644 --- a/tests/interface/input-tsv/test_tsv_train_inputtypes_stdin.sh +++ b/tests/interface/input-tsv/test_tsv_train_inputtypes_stdin.sh @@ -14,7 +14,7 @@ mkdir -p train_intypes_stdin # Run marian command cat train.tsv | $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ -m train_intypes_stdin/model.npz --tsv -t stdin --input-types sequence sequence -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 10 --disp-freq 2 \ --log train_intypes_stdin.log diff --git a/tests/interface/input-tsv/test_tsv_train_mini_batch_fit.sh b/tests/interface/input-tsv/test_tsv_train_mini_batch_fit.sh index 7ab4f57..42afeac 100644 --- a/tests/interface/input-tsv/test_tsv_train_mini_batch_fit.sh +++ b/tests/interface/input-tsv/test_tsv_train_mini_batch_fit.sh @@ -14,7 +14,7 @@ mkdir -p train_fit # Run marian command $MRT_MARIAN/marian \ - --mini-batch-fit -w 500 --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 10 --optimizer sgd --sync-sgd \ + --cost-type ce-mean --mini-batch-fit -w 500 --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 10 --optimizer sgd --sync-sgd \ -m train_fit/model.npz --tsv -t train.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 20 --disp-freq 4 \ --log train_fit.log diff --git a/tests/interface/input-tsv/test_tsv_train_mini_batch_fit_stdin.sh b/tests/interface/input-tsv/test_tsv_train_mini_batch_fit_stdin.sh index 8b26362..0b4ba42 100644 --- a/tests/interface/input-tsv/test_tsv_train_mini_batch_fit_stdin.sh +++ b/tests/interface/input-tsv/test_tsv_train_mini_batch_fit_stdin.sh @@ -14,7 +14,7 @@ mkdir -p train_fit_stdin # Run marian command cat train.tsv | $MRT_MARIAN/marian \ - --no-shuffle --mini-batch-fit -w 500 --seed 2222 --dim-emb 32 --dim-rnn 64 --maxi-batch 10 --optimizer sgd \ + --cost-type ce-mean --no-shuffle --mini-batch-fit -w 500 --seed 2222 --dim-emb 32 --dim-rnn 64 --maxi-batch 10 --optimizer sgd \ -m train_fit_stdin/model.npz --tsv -t stdin -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --disp-freq 4 --log train_fit_stdin.log diff --git a/tests/interface/input-tsv/test_tsv_train_restore_from_stdin.sh b/tests/interface/input-tsv/test_tsv_train_restore_from_stdin.sh index 4783e76..f8953ef 100644 --- a/tests/interface/input-tsv/test_tsv_train_restore_from_stdin.sh +++ b/tests/interface/input-tsv/test_tsv_train_restore_from_stdin.sh @@ -11,12 +11,12 @@ test -e vocab.de.yml || $MRT_MARIAN/marian-vocab < train.bpe.de > vocab.de.yml test -e vocab.en.yml || $MRT_MARIAN/marian-vocab < train.bpe.en > vocab.en.yml # TODO: Weight decaying in Adam is disabled, because it gives unstable results on GPU -extra_opts="--no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --mini-batch 16 --dim-emb 128 --dim-rnn 256 --disp-freq 2 --type s2s --sync-sgd --optimizer sgd" +extra_opts="--no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --mini-batch 16 --dim-emb 128 --dim-rnn 256 --disp-freq 2 --type s2s --sync-sgd --optimizer sgd --cost-type ce-mean" # Step 1: Train a model in one go, up to the update no. 70, and save training logs #$MRT_MARIAN/marian \ #-m restore_stdin/model_full.npz -t train.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ - #--after-batches 60 $extra_opts \ + #--cost-type ce-mean --after-batches 60 $extra_opts \ #--log restore_stdin.log #test -e restore_stdin/model_full.npz diff --git a/tests/interface/input-tsv/test_tsv_train_shuffle.sh b/tests/interface/input-tsv/test_tsv_train_shuffle.sh index 502a6a3..e5ca216 100644 --- a/tests/interface/input-tsv/test_tsv_train_shuffle.sh +++ b/tests/interface/input-tsv/test_tsv_train_shuffle.sh @@ -14,7 +14,7 @@ mkdir -p train_shuffle # Run marian command $MRT_MARIAN/marian \ - --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 10 --optimizer sgd --sync-sgd \ + --cost-type ce-mean --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 10 --optimizer sgd --sync-sgd \ -m train_shuffle/model.npz --tsv --tsv-fields 2 -t train.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 20 --disp-freq 4 \ --log train_shuffle.log diff --git a/tests/interface/input-tsv/test_tsv_train_shuffle_in_ram.sh b/tests/interface/input-tsv/test_tsv_train_shuffle_in_ram.sh index d6c20e7..8ce67c3 100644 --- a/tests/interface/input-tsv/test_tsv_train_shuffle_in_ram.sh +++ b/tests/interface/input-tsv/test_tsv_train_shuffle_in_ram.sh @@ -14,7 +14,7 @@ mkdir -p train_shuffle_ram # Run marian command $MRT_MARIAN/marian \ - --shuffle-in-ram --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 10 --optimizer sgd \ + --cost-type ce-mean --shuffle-in-ram --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 10 --optimizer sgd \ -m train_shuffle_ram/model.npz --tsv --tsv-fields 2 -t train.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 20 --disp-freq 4 \ --log train_shuffle_ram.log diff --git a/tests/interface/input-tsv/test_tsv_train_stdin.sh b/tests/interface/input-tsv/test_tsv_train_stdin.sh index 1f9e98b..2a77516 100644 --- a/tests/interface/input-tsv/test_tsv_train_stdin.sh +++ b/tests/interface/input-tsv/test_tsv_train_stdin.sh @@ -14,7 +14,7 @@ mkdir -p train_stdin # Run marian command cat train.tsv | $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ -m train_stdin/model.npz --tsv -t stdin -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 10 --disp-freq 2 \ --log train_stdin.log diff --git a/tests/interface/input-tsv/test_tsv_train_stdin_2_epochs.sh b/tests/interface/input-tsv/test_tsv_train_stdin_2_epochs.sh index 085006c..af35c5c 100644 --- a/tests/interface/input-tsv/test_tsv_train_stdin_2_epochs.sh +++ b/tests/interface/input-tsv/test_tsv_train_stdin_2_epochs.sh @@ -14,7 +14,7 @@ mkdir -p train_stdin_2e # Train for the 1st epoch cat train.tsv | $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ -m train_stdin_2e/model.npz --tsv -t stdin -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --disp-freq 5 \ --log train_stdin_2e.log @@ -25,7 +25,7 @@ test -e train_stdin_2e.log # Train for the 2nd epoch cat train.tsv | $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.002 \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.002 \ -m train_stdin_2e/model.npz --tsv -t stdin -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --disp-freq 5 \ --log train_stdin_2e.log diff --git a/tests/interface/input-tsv/test_tsv_train_stdin_empty_fields.sh b/tests/interface/input-tsv/test_tsv_train_stdin_empty_fields.sh index 4e1aa55..e9b0b33 100644 --- a/tests/interface/input-tsv/test_tsv_train_stdin_empty_fields.sh +++ b/tests/interface/input-tsv/test_tsv_train_stdin_empty_fields.sh @@ -20,7 +20,7 @@ paste train.{de,en} \ # Run marian command cat train_empty_lines.tsv | $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --max-length 200 \ + --cost-type ce-mean --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --max-length 200 \ -m train_empty_lines/model.npz --tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-epochs 1 --disp-freq 2 \ --log train_empty_lines.log diff --git a/tests/interface/input-tsv/test_tsv_train_stdin_lm.sh b/tests/interface/input-tsv/test_tsv_train_stdin_lm.sh index ca22c8f..e9a25c6 100644 --- a/tests/interface/input-tsv/test_tsv_train_stdin_lm.sh +++ b/tests/interface/input-tsv/test_tsv_train_stdin_lm.sh @@ -14,7 +14,7 @@ mkdir -p train_lm # Run marian command cat train.en | $MRT_MARIAN/marian --type lm \ - --no-shuffle --seed 4444 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --cost-type ce-mean --no-shuffle --seed 4444 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ -m train_lm/model.npz -t stdin -v $MRT_MODELS/rnn-spm/vocab.deen.spm \ --after-batches 10 --disp-freq 2 \ --log train_lm.log diff --git a/tests/interface/input-tsv/test_tsv_train_with_align.sh b/tests/interface/input-tsv/test_tsv_train_with_align.sh index 8edf098..c101d51 100644 --- a/tests/interface/input-tsv/test_tsv_train_with_align.sh +++ b/tests/interface/input-tsv/test_tsv_train_with_align.sh @@ -14,7 +14,7 @@ mkdir -p train_align # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ + --cost-type ce-mean --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ -m train_align/model.npz --tsv -t train2.de-en-aln.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 100 --disp-freq 4 \ --guided-alignment 2 --guided-alignment-weight 1.0 \ diff --git a/tests/interface/input-tsv/test_tsv_train_with_align_and_weights.sh b/tests/interface/input-tsv/test_tsv_train_with_align_and_weights.sh index 60c45c3..b02488a 100644 --- a/tests/interface/input-tsv/test_tsv_train_with_align_and_weights.sh +++ b/tests/interface/input-tsv/test_tsv_train_with_align_and_weights.sh @@ -14,7 +14,7 @@ mkdir -p train_align_weights # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 7777 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ + --cost-type ce-mean --no-shuffle --seed 7777 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ -m train_align_weights/model.npz --tsv -t train2.de-w-aln-en.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 60 --disp-freq 4 \ --guided-alignment 2 --guided-alignment-weight 1.0 --data-weighting 1 \ diff --git a/tests/interface/input-tsv/test_tsv_train_with_align_and_weights_inputtypes.sh b/tests/interface/input-tsv/test_tsv_train_with_align_and_weights_inputtypes.sh index a983f22..8653a67 100644 --- a/tests/interface/input-tsv/test_tsv_train_with_align_and_weights_inputtypes.sh +++ b/tests/interface/input-tsv/test_tsv_train_with_align_and_weights_inputtypes.sh @@ -14,7 +14,7 @@ mkdir -p train_align_weights_intypes # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 7777 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ + --cost-type ce-mean --no-shuffle --seed 7777 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ -m train_align_weights_intypes/model.npz --tsv -t train2.de-w-aln-en.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 60 --disp-freq 4 \ --input-types sequence weight alignment sequence --guided-alignment-weight 1.0 \ diff --git a/tests/interface/input-tsv/test_tsv_train_with_align_pos0.sh b/tests/interface/input-tsv/test_tsv_train_with_align_pos0.sh index 85758e3..55f9995 100644 --- a/tests/interface/input-tsv/test_tsv_train_with_align_pos0.sh +++ b/tests/interface/input-tsv/test_tsv_train_with_align_pos0.sh @@ -14,7 +14,7 @@ mkdir -p train_align0 # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ + --cost-type ce-mean --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ -m train_align0/model.npz --tsv -t train2.aln-de-en.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 100 --disp-freq 4 \ --guided-alignment 0 --guided-alignment-weight 1.0 \ diff --git a/tests/interface/input-tsv/test_tsv_train_with_align_shuffle.sh b/tests/interface/input-tsv/test_tsv_train_with_align_shuffle.sh index 3a4168f..d02f8dc 100644 --- a/tests/interface/input-tsv/test_tsv_train_with_align_shuffle.sh +++ b/tests/interface/input-tsv/test_tsv_train_with_align_shuffle.sh @@ -14,7 +14,7 @@ mkdir -p train_align_shuffle # Run marian command $MRT_MARIAN/marian \ - --seed 4444 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 --sync-sgd \ + --cost-type ce-mean --seed 4444 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 --sync-sgd \ -m train_align_shuffle/model.npz --tsv -t train2.aln-de-en.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 100 --disp-freq 4 \ --guided-alignment 0 --guided-alignment-weight 1.0 \ diff --git a/tests/interface/input-tsv/test_tsv_train_with_align_shuffle_in_ram.sh b/tests/interface/input-tsv/test_tsv_train_with_align_shuffle_in_ram.sh index 0934fa1..2c4dc1a 100644 --- a/tests/interface/input-tsv/test_tsv_train_with_align_shuffle_in_ram.sh +++ b/tests/interface/input-tsv/test_tsv_train_with_align_shuffle_in_ram.sh @@ -14,7 +14,7 @@ mkdir -p train_align_shuffle_ram # Run marian command $MRT_MARIAN/marian \ - --shuffle-in-ram --seed 4444 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ + --cost-type ce-mean --shuffle-in-ram --seed 4444 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ -m train_align_shuffle_ram/model.npz --tsv -t train2.aln-de-en.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 100 --disp-freq 4 \ --guided-alignment 0 --guided-alignment-weight 1.0 \ diff --git a/tests/interface/input-tsv/test_tsv_train_with_align_stdin.sh b/tests/interface/input-tsv/test_tsv_train_with_align_stdin.sh index 92adebc..b266f3d 100644 --- a/tests/interface/input-tsv/test_tsv_train_with_align_stdin.sh +++ b/tests/interface/input-tsv/test_tsv_train_with_align_stdin.sh @@ -14,7 +14,7 @@ mkdir -p train_align_stdin # Run marian command cat train2.aln-de-en.tsv | $MRT_MARIAN/marian \ - --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ + --cost-type ce-mean --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ -m train_align_stdin/model.npz -t stdin -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --disp-freq 4 \ --guided-alignment 0 --guided-alignment-weight 1.0 \ diff --git a/tests/interface/input-tsv/test_tsv_train_with_weights.sh b/tests/interface/input-tsv/test_tsv_train_with_weights.sh index 7d6927e..6528713 100644 --- a/tests/interface/input-tsv/test_tsv_train_with_weights.sh +++ b/tests/interface/input-tsv/test_tsv_train_with_weights.sh @@ -14,7 +14,7 @@ mkdir -p train_weights # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ + --cost-type ce-mean --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ -m train_weights/model.npz --tsv -t train2.de-en-w.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 100 --disp-freq 4 \ --data-weighting 2 --data-weighting-type sentence \ diff --git a/tests/interface/input-tsv/test_tsv_train_with_weights_pos0.sh b/tests/interface/input-tsv/test_tsv_train_with_weights_pos0.sh index 430b8c9..ee4522f 100644 --- a/tests/interface/input-tsv/test_tsv_train_with_weights_pos0.sh +++ b/tests/interface/input-tsv/test_tsv_train_with_weights_pos0.sh @@ -14,7 +14,7 @@ mkdir -p train_weights0 # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ + --cost-type ce-mean --no-shuffle --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.1 \ -m train_weights0/model.npz --tsv -t train2.w-de-en.tsv -v $MRT_MODELS/rnn-spm/vocab.deen.{spm,spm} \ --after-batches 100 --disp-freq 4 \ --data-weighting 0 --data-weighting-type sentence \ diff --git a/tests/training/basics/gzip.expected b/tests/training/basics/gzip.expected index 47c8765..e6ad9be 100644 --- a/tests/training/basics/gzip.expected +++ b/tests/training/basics/gzip.expected @@ -1,5 +1,5 @@ -448.35955811 -374.86718750 -324.67779541 -283.97842407 -248.12561035 +10.16832447 +10.15427303 +10.14024067 +10.12231636 +10.10052204 diff --git a/tests/training/basics/test_gzipped_train_sets.sh b/tests/training/basics/test_gzipped_train_sets.sh index 87ae5ef..85e4ddd 100644 --- a/tests/training/basics/test_gzipped_train_sets.sh +++ b/tests/training/basics/test_gzipped_train_sets.sh @@ -22,7 +22,7 @@ test -e gzip/model.npz test -e gzip.log cat gzip.log | $MRT_TOOLS/extract-costs.sh > gzip.out -$MRT_TOOLS/diff-nums.py gzip.out gzip.expected -p 0.1 -o gzip.diff +$MRT_TOOLS/diff-nums.py gzip.out gzip.expected -p 0.05 -o gzip.diff # Exit with success code exit 0 diff --git a/tests/training/basics/test_toy_vocab.sh b/tests/training/basics/test_toy_vocab.sh index 2892fdb..671843f 100644 --- a/tests/training/basics/test_toy_vocab.sh +++ b/tests/training/basics/test_toy_vocab.sh @@ -25,7 +25,7 @@ test -e toy/model.npz.yml test -e toy/model.npz.amun.yml cat toy.log | $MRT_TOOLS/extract-costs.sh > toy.out -$MRT_TOOLS/diff-nums.py toy.out toy.expected -p 0.5 -o toy.diff +$MRT_TOOLS/diff-nums.py toy.out toy.expected -p 0.1 -o toy.diff # Exit with success code exit 0 diff --git a/tests/training/basics/toy.expected b/tests/training/basics/toy.expected index e978d78..f134a62 100644 --- a/tests/training/basics/toy.expected +++ b/tests/training/basics/toy.expected @@ -1,44 +1,44 @@ -387.11929321 -317.48147583 -270.41656494 -229.41304016 -191.78842163 -152.79827881 -117.75115967 -84.71041870 -123.99937439 -301.64492798 -245.78007507 -207.74856567 -172.86341858 -144.79455566 -118.54357910 -95.56129456 -67.47161865 -172.77297974 -264.60137939 -218.77970886 -187.74063110 -157.71662903 -132.04907227 -108.09243011 -87.24269867 -57.82399750 -227.66172791 -249.18188477 -207.59941101 -177.44326782 -149.12503052 -124.39839935 -101.03861237 -79.69851685 -47.27352142 -321.00448608 -260.45831299 -210.67678833 -176.55291748 -146.53402710 -120.28997803 -96.91591644 -75.33888245 -43.18191910 +8.78630924 +8.75282860 +8.69768810 +8.60030174 +8.42331791 +8.08456516 +7.58919859 +7.15421867 +7.07388210 +7.15911722 +7.00602388 +6.87553406 +6.70337963 +6.58344078 +6.49999046 +6.45548153 +6.15827656 +6.46891832 +6.53488016 +6.44239426 +6.40002155 +6.31497908 +6.22505951 +6.17120123 +6.16692400 +5.82793045 +6.37700939 +6.38672018 +6.30941343 +6.23907804 +6.14631748 +6.09236145 +5.99969482 +5.93795681 +5.50292015 +7.16254091 +6.92708254 +6.59395170 +6.40627527 +6.22969198 +6.13137770 +5.96749878 +5.96595860 +5.42677450 diff --git a/tests/training/cost-functions/test_ce-mean.sh b/tests/training/cost-functions/test_ce-mean.sh index 9d49c48..c4109b1 100644 --- a/tests/training/cost-functions/test_ce-mean.sh +++ b/tests/training/cost-functions/test_ce-mean.sh @@ -14,6 +14,7 @@ rm -rf ce-mean ce-mean.log mkdir -p ce-mean $MRT_MARIAN/marian \ + --cost-type ce-mean \ --seed 9999 --sync-sgd \ -m ce-mean/model.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 2 --after-epochs 1 \ diff --git a/tests/training/cost-functions/test_ce-sum.sh b/tests/training/cost-functions/test_ce-sum.sh index dc8bfe0..f22f137 100644 --- a/tests/training/cost-functions/test_ce-sum.sh +++ b/tests/training/cost-functions/test_ce-sum.sh @@ -14,7 +14,7 @@ rm -rf ce-sum ce-sum.log mkdir -p ce-sum $MRT_MARIAN/marian \ - --cost-type ce-sum \ + --cost-type ce-sum --disp-label-counts false \ --seed 9999 --optimizer sgd --sync-sgd \ -m ce-sum/model.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 2 --after-epochs 1 \ diff --git a/tests/training/features/data-weighting/test_maxi_batches_with_sentence_weights.sh b/tests/training/features/data-weighting/test_maxi_batches_with_sentence_weights.sh index 2cb6e29..58e9e82 100644 --- a/tests/training/features/data-weighting/test_maxi_batches_with_sentence_weights.sh +++ b/tests/training/features/data-weighting/test_maxi_batches_with_sentence_weights.sh @@ -18,7 +18,7 @@ test -e vocab.en.yml || $MRT_MARIAN/marian-vocab < $MRT_DATA/europarl.de-en/corp $MRT_MARIAN/marian \ --seed 3333 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd \ -m maxibatch/model.npz -t train.1k.{de,en} -v vocab.{de,en}.yml \ - --log maxibatch.log --disp-freq 10 --after-batches 100 --mini-batch 16 --cost-type ce-sum \ + --log maxibatch.log --disp-freq 10 --after-batches 100 --mini-batch 16 --cost-type ce-sum --disp-label-counts false \ --data-weighting train.1k.inc.txt --data-weighting-type sentence test -e maxibatch/model.npz diff --git a/tests/training/features/data-weighting/test_maxi_batches_with_word_weights.sh b/tests/training/features/data-weighting/test_maxi_batches_with_word_weights.sh index 2ba494f..5cb5592 100644 --- a/tests/training/features/data-weighting/test_maxi_batches_with_word_weights.sh +++ b/tests/training/features/data-weighting/test_maxi_batches_with_word_weights.sh @@ -18,7 +18,7 @@ test -e vocab.en.yml || $MRT_MARIAN/marian-vocab < $MRT_DATA/europarl.de-en/corp $MRT_MARIAN/marian \ --seed 6666 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd \ -m word_maxibatch/model.npz -t train.1k.{de,en} -v vocab.{de,en}.yml \ - --log word_maxibatch.log --disp-freq 10 --after-batches 100 --mini-batch 16 \ + --log word_maxibatch.log --disp-freq 10 --after-batches 100 --mini-batch 16 --cost-type ce-mean \ --data-weighting train.1k.wordinc.txt --data-weighting-type word test -e word_maxibatch/model.npz diff --git a/tests/training/features/data-weighting/test_sentence_weighting_sqlite.sh b/tests/training/features/data-weighting/test_sentence_weighting_sqlite.sh index c0a200a..2295d64 100644 --- a/tests/training/features/data-weighting/test_sentence_weighting_sqlite.sh +++ b/tests/training/features/data-weighting/test_sentence_weighting_sqlite.sh @@ -13,7 +13,7 @@ rm -rf sqlite sqlite.log mkdir -p sqlite $MRT_MARIAN/marian \ - --seed 1111 --no-shuffle --maxi-batch 1 --maxi-batch-sort none --max-length 100 --dim-emb 128 --dim-rnn 256 --optimizer sgd \ + --seed 1111 --no-shuffle --maxi-batch 1 --maxi-batch-sort none --max-length 100 --dim-emb 128 --dim-rnn 256 --optimizer sgd --cost-type ce-mean \ -m sqlite/model.npz -t train.1k.{de,en} -v vocab.{de,en}.yml \ --log sqlite.log --disp-freq 1 --after-batches 100 --mini-batch 1 \ --data-weighting train.1k.weights.txt --data-weighting-type sentence --sqlite sqlite/corpus.sqlite3 diff --git a/tests/training/features/data-weighting/test_sentence_weights_x3.sh b/tests/training/features/data-weighting/test_sentence_weights_x3.sh index 3857814..1b28af8 100644 --- a/tests/training/features/data-weighting/test_sentence_weights_x3.sh +++ b/tests/training/features/data-weighting/test_sentence_weights_x3.sh @@ -18,7 +18,7 @@ test -e vocab.en.yml || $MRT_MARIAN/marian-vocab < $MRT_DATA/europarl.de-en/corp $MRT_MARIAN/marian \ --seed 2222 --no-shuffle --maxi-batch 1 --maxi-batch-sort none --max-length 100 --dim-emb 128 --dim-rnn 256 --optimizer sgd \ -m x3copied/model.npz -t train.x3.{de,en} -v vocab.{de,en}.yml \ - --log x3copied.log --disp-freq 1 --after-batches 100 --mini-batch 4 --cost-type ce-sum + --log x3copied.log --disp-freq 1 --after-batches 100 --mini-batch 4 --cost-type ce-sum --disp-label-counts false test -e x3copied/model.npz test -e x3copied.log @@ -27,7 +27,7 @@ cat x3copied.log | grep 'Cost ' | sed -r 's/.*Cost (.*) : Time.*/\1/' > x3copied $MRT_MARIAN/marian \ --seed 2222 --no-shuffle --maxi-batch 1 --maxi-batch-sort none --max-length 100 --dim-emb 128 --dim-rnn 256 --optimizer sgd \ -m x3weights/model.npz -t train.1k.{de,en} -v vocab.{de,en}.yml \ - --log x3weights.log --disp-freq 1 --after-batches 100 --mini-batch 2 --cost-type ce-sum \ + --log x3weights.log --disp-freq 1 --after-batches 100 --mini-batch 2 --cost-type ce-sum --disp-label-counts false \ --data-weighting train.1k.weights.txt --data-weighting-type sentence test -e x3weights/model.npz diff --git a/tests/training/features/data-weighting/test_validation.sh b/tests/training/features/data-weighting/test_validation.sh index e84c976..125ba34 100644 --- a/tests/training/features/data-weighting/test_validation.sh +++ b/tests/training/features/data-weighting/test_validation.sh @@ -13,7 +13,7 @@ rm -rf valid valid_script.temp mkdir -p valid $MRT_MARIAN/marian \ - --seed 4444 --no-shuffle --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 \ + --seed 4444 --no-shuffle --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --cost-type ce-mean \ -m valid/model.npz -t train.1k.{de,en} -v vocab.{de,en}.yml \ --disp-freq 5 --valid-freq 15 --after-batches 50 \ --data-weighting train.1k.weights.txt --data-weighting-type sentence \ diff --git a/tests/training/features/data-weighting/test_word_weighting_sqlite.sh b/tests/training/features/data-weighting/test_word_weighting_sqlite.sh index a07c344..1658e1b 100644 --- a/tests/training/features/data-weighting/test_word_weighting_sqlite.sh +++ b/tests/training/features/data-weighting/test_word_weighting_sqlite.sh @@ -15,7 +15,7 @@ mkdir -p sqlite_word cat $MRT_DATA/europarl.de-en/toy.bpe.en | sed -r 's/[^ ]+/2/g' > sqlite_word.weights.txt $MRT_MARIAN/marian \ - --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd \ + --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd --cost-type ce-mean \ -m sqlite_word/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{de,en} -v vocab.{de,en}.yml \ --log sqlite_word.log --disp-freq 5 -e 2 --mini-batch-fit -w 500 \ --data-weighting sqlite_word.weights.txt --data-weighting-type word --sqlite sqlite_word/corpus.sqlite3 diff --git a/tests/training/features/data-weighting/test_word_weighting_with_eos.sh b/tests/training/features/data-weighting/test_word_weighting_with_eos.sh index 1242f40..c3d7b93 100644 --- a/tests/training/features/data-weighting/test_word_weighting_with_eos.sh +++ b/tests/training/features/data-weighting/test_word_weighting_with_eos.sh @@ -17,7 +17,7 @@ cat $MRT_DATA/europarl.de-en/toy.bpe.en | sed -r -e 's/[^ ]+/2/g' -e 's/$/ 2/' > # Train $MRT_MARIAN/marian \ - --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd \ + --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd --cost-type ce-mean \ -m word_eos/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{de,en} -v vocab.{de,en}.yml \ --log word_eos.log --disp-freq 5 -e 2 \ --data-weighting word_eos.weights.txt --data-weighting-type word diff --git a/tests/training/features/data-weighting/test_word_weighting_with_twos.sh b/tests/training/features/data-weighting/test_word_weighting_with_twos.sh index 81bc3e5..0660501 100644 --- a/tests/training/features/data-weighting/test_word_weighting_with_twos.sh +++ b/tests/training/features/data-weighting/test_word_weighting_with_twos.sh @@ -18,7 +18,7 @@ cat $MRT_DATA/europarl.de-en/toy.bpe.en | sed -r 's/[^ ]+/2/g' > word_twos.weigh # Train with word weighting $MRT_MARIAN/marian \ - --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd \ + --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd --cost-type ce-mean \ -m word_twos/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{de,en} -v vocab.{de,en}.yml \ --log word_twos.log --disp-freq 5 -e 2 \ --data-weighting word_twos.weights.txt --data-weighting-type word @@ -41,7 +41,7 @@ echo "data-weighting-type: word" >> word_twos.config.yml # Train with word weighting $MRT_MARIAN/marian \ - --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd \ + --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd --cost-type ce-mean \ -m word_twos_cfg/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{de,en} -v vocab.{de,en}.yml \ --log word_twos_cfg.log --disp-freq 5 -e 2 \ -c word_twos.config.yml diff --git a/tests/training/features/data-weighting/test_word_weighting_with_twos_sync.sh b/tests/training/features/data-weighting/test_word_weighting_with_twos_sync.sh index 68154af..675ae8b 100644 --- a/tests/training/features/data-weighting/test_word_weighting_with_twos_sync.sh +++ b/tests/training/features/data-weighting/test_word_weighting_with_twos_sync.sh @@ -18,7 +18,7 @@ cat $MRT_DATA/europarl.de-en/toy.bpe.en | sed -r 's/[^ ]+/2/g' > word_twos_sync. # Train with word weighting $MRT_MARIAN/marian \ - --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd \ + --seed 1111 --no-shuffle --dim-emb 128 --dim-rnn 256 --optimizer sgd --cost-type ce-mean \ -m word_twos_sync/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{de,en} -v vocab.{de,en}.yml --sync-sgd \ --log word_twos_sync.log --disp-freq 5 -e 2 \ --data-weighting word_twos_sync.weights.txt --data-weighting-type word diff --git a/tests/training/features/exp-smoothing/test_expsmooth.sh b/tests/training/features/exp-smoothing/test_expsmooth.sh index f55f64d..b27cee9 100644 --- a/tests/training/features/exp-smoothing/test_expsmooth.sh +++ b/tests/training/features/exp-smoothing/test_expsmooth.sh @@ -8,7 +8,7 @@ rm -rf expsmooth expsmooth*.log mkdir -p expsmooth -opts="--no-shuffle --seed 777 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.5 --valid-sets valid.bpe.en valid.bpe.de --valid-metrics cross-entropy --valid-mini-batch 32" +opts="--no-shuffle --seed 777 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.5 --valid-sets valid.bpe.en valid.bpe.de --valid-metrics cross-entropy --valid-mini-batch 32 --cost-type ce-mean" # No exponential smoothing $MRT_MARIAN/marian \ diff --git a/tests/training/features/guided-alignment/test_guided_alignment_rnn.sh b/tests/training/features/guided-alignment/test_guided_alignment_rnn.sh index e1dea95..a022e5c 100644 --- a/tests/training/features/guided-alignment/test_guided_alignment_rnn.sh +++ b/tests/training/features/guided-alignment/test_guided_alignment_rnn.sh @@ -15,7 +15,7 @@ mkdir -p rnn # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --cost-type ce-mean \ -m rnn/model.npz -t corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --after-batches 100 --disp-freq 10 \ --guided-alignment corpus.bpe.align --guided-alignment-weight 1.0 --learn-rate 0.1 \ diff --git a/tests/training/features/guided-alignment/test_guided_alignment_transformer.sh b/tests/training/features/guided-alignment/test_guided_alignment_transformer.sh index 9694aab..f5f18b9 100644 --- a/tests/training/features/guided-alignment/test_guided_alignment_transformer.sh +++ b/tests/training/features/guided-alignment/test_guided_alignment_transformer.sh @@ -15,7 +15,7 @@ mkdir -p transformer # Run marian command $MRT_MARIAN/marian --type transformer \ - --no-shuffle --seed 2222 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --no-shuffle --seed 2222 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --cost-type ce-mean \ -m transformer/model.npz -t corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --after-batches 100 --disp-freq 10 \ --guided-alignment corpus.bpe.align --guided-alignment-weight 1.0 --learn-rate 0.1 \ diff --git a/tests/training/features/guided-alignment/test_guided_alignment_transformer_sync.sh b/tests/training/features/guided-alignment/test_guided_alignment_transformer_sync.sh index 3bbd5a7..49675e8 100644 --- a/tests/training/features/guided-alignment/test_guided_alignment_transformer_sync.sh +++ b/tests/training/features/guided-alignment/test_guided_alignment_transformer_sync.sh @@ -15,7 +15,7 @@ mkdir -p transformer_sync # Run marian command $MRT_MARIAN/marian --type transformer \ - --no-shuffle --seed 2222 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --sync-sgd \ + --no-shuffle --seed 2222 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --cost-type ce-mean --sync-sgd \ -m transformer_sync/model.npz -t corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --after-batches 100 --disp-freq 10 \ --guided-alignment corpus.bpe.align --guided-alignment-weight 1.0 --learn-rate 0.1 \ diff --git a/tests/training/features/right-left/test_right_left_rnn.sh b/tests/training/features/right-left/test_right_left_rnn.sh index 31c65dc..245125e 100644 --- a/tests/training/features/right-left/test_right_left_rnn.sh +++ b/tests/training/features/right-left/test_right_left_rnn.sh @@ -14,7 +14,7 @@ mkdir -p rnn # Run marian command $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --no-shuffle --seed 1111 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --cost-type ce-mean \ -m rnn/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --after-batches 100 --disp-freq 10 \ --right-left --log rnn.log diff --git a/tests/training/features/right-left/test_right_left_transformer.sh b/tests/training/features/right-left/test_right_left_transformer.sh index 429ef69..8a40bdd 100644 --- a/tests/training/features/right-left/test_right_left_transformer.sh +++ b/tests/training/features/right-left/test_right_left_transformer.sh @@ -14,7 +14,7 @@ mkdir -p transformer # Run marian command $MRT_MARIAN/marian --type transformer \ - --no-shuffle --seed 2222 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ + --no-shuffle --seed 2222 --dim-emb 32 --dim-rnn 64 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --cost-type ce-mean \ -m transformer/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --after-batches 100 --disp-freq 10 \ --right-left --log transformer.log diff --git a/tests/training/models/lm/test_lm-transformer.sh b/tests/training/models/lm/test_lm-transformer.sh index 916d08e..476c2e2 100644 --- a/tests/training/models/lm/test_lm-transformer.sh +++ b/tests/training/models/lm/test_lm-transformer.sh @@ -15,7 +15,7 @@ mkdir -p lm-transformer $MRT_MARIAN/marian \ --seed 1111 --no-shuffle \ - --type lm-transformer --dim-emb 128 --dim-rnn 256 \ + --type lm-transformer --dim-emb 128 --dim-rnn 256 --cost-type ce-mean \ -m lm-transformer/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.en -v vocab.en.yml \ --disp-freq 20 --after-batches 100 \ --log lm-transformer.log diff --git a/tests/training/models/lm/test_lm.sh b/tests/training/models/lm/test_lm.sh index 75503a6..91f94d4 100644 --- a/tests/training/models/lm/test_lm.sh +++ b/tests/training/models/lm/test_lm.sh @@ -15,7 +15,7 @@ mkdir -p lm $MRT_MARIAN/marian \ --seed 1111 --no-shuffle \ - --type lm --dim-emb 128 --dim-rnn 256 \ + --type lm --dim-emb 128 --dim-rnn 256 --cost-type ce-mean \ -m lm/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.en -v vocab.en.yml \ --disp-freq 20 --after-batches 100 \ --log lm.log diff --git a/tests/training/models/multi-source/test_multi-s2s.sh b/tests/training/models/multi-source/test_multi-s2s.sh index e53808e..a1ceef1 100644 --- a/tests/training/models/multi-source/test_multi-s2s.sh +++ b/tests/training/models/multi-source/test_multi-s2s.sh @@ -15,7 +15,7 @@ mkdir -p multi-s2s $MRT_MARIAN/marian \ --seed 1111 --no-shuffle \ - --type multi-s2s --dim-emb 128 --dim-rnn 256 \ + --type multi-s2s --dim-emb 128 --dim-rnn 256 --cost-type ce-mean \ -m multi-s2s/model.npz -t train.bpe.{en,xx,de} -v vocab.en.yml vocab.xx.yml vocab.de.yml \ --disp-freq 20 --after-batches 100 \ --log multi-s2s.log diff --git a/tests/training/models/multi-source/test_multi-transformer.sh b/tests/training/models/multi-source/test_multi-transformer.sh index b263f98..425ebdc 100644 --- a/tests/training/models/multi-source/test_multi-transformer.sh +++ b/tests/training/models/multi-source/test_multi-transformer.sh @@ -15,7 +15,7 @@ mkdir -p multi-transformer $MRT_MARIAN/marian \ --seed 1111 --no-shuffle \ - --type multi-transformer --dim-emb 128 --dim-rnn 256 \ + --type multi-transformer --dim-emb 128 --dim-rnn 256 --cost-type ce-mean \ -m multi-transformer/model.npz -t train.bpe.{en,xx,de} -v vocab.en.yml vocab.xx.yml vocab.de.yml \ --disp-freq 20 --after-batches 100 \ --log multi-transformer.log diff --git a/tests/training/models/nematus/test_encdec_depth.sh b/tests/training/models/nematus/test_encdec_depth.sh index fd08c9d..ed5276e 100644 --- a/tests/training/models/nematus/test_encdec_depth.sh +++ b/tests/training/models/nematus/test_encdec_depth.sh @@ -17,7 +17,7 @@ $MRT_MARIAN/marian \ --type nematus --enc-cell gru-nematus --dec-cell gru-nematus \ --enc-depth 4 --enc-cell-depth 4 --enc-type bidirectional --dec-depth 4 --dec-cell-base-depth 4 --dec-cell-high-depth 1 \ --layer-normalization \ - --no-shuffle --seed 1111 --dim-emb 64 --dim-rnn 128 \ + --no-shuffle --seed 1111 --dim-emb 64 --dim-rnn 128 --cost-type ce-mean \ -m encdec_depth/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{de,en} -v vocab.en.yml vocab.de.yml \ --log encdec_depth.log --disp-freq 2 --after-batches 10 diff --git a/tests/training/models/nematus/test_wmt17_model.sh b/tests/training/models/nematus/test_wmt17_model.sh index 5c06c83..d43ff9c 100644 --- a/tests/training/models/nematus/test_wmt17_model.sh +++ b/tests/training/models/nematus/test_wmt17_model.sh @@ -11,7 +11,7 @@ $MRT_MARIAN/marian \ --type nematus --enc-cell gru-nematus --dec-cell gru-nematus \ --enc-depth 1 --enc-cell-depth 4 --enc-type bidirectional --dec-depth 1 --dec-cell-base-depth 8 --dec-cell-high-depth 1 \ --layer-normalization \ - --no-shuffle --seed 1111 --dim-emb 64 --dim-rnn 128 \ + --no-shuffle --seed 1111 --dim-emb 64 --dim-rnn 128 --cost-type ce-mean \ -m wmt17/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{de,en} -v vocab.en.yml vocab.de.yml \ --log wmt17.log --disp-freq 2 --after-batches 10 diff --git a/tests/training/models/transformer/test_transformer.sh b/tests/training/models/transformer/test_transformer.sh index e9754cf..b1ad881 100644 --- a/tests/training/models/transformer/test_transformer.sh +++ b/tests/training/models/transformer/test_transformer.sh @@ -7,7 +7,7 @@ set -e rm -rf transformer transformer*.log mkdir -p transformer -opts="--no-shuffle --seed 1111 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --dim-emb 64 --dim-rnn 128" +opts="--no-shuffle --seed 1111 --mini-batch 32 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --dim-emb 64 --dim-rnn 128 --cost-type ce-mean" $MRT_MARIAN/marian \ --type transformer -m transformer/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ diff --git a/tests/training/multi-gpu/test_async_sgd_runs.sh b/tests/training/multi-gpu/test_async_sgd_runs.sh index 1094cd2..5043548 100644 --- a/tests/training/multi-gpu/test_async_sgd_runs.sh +++ b/tests/training/multi-gpu/test_async_sgd_runs.sh @@ -13,7 +13,7 @@ rm -rf async_sgd async_sgd.log mkdir -p async_sgd $MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --maxi-batch 1 --maxi-batch-sort none --dim-emb 128 --dim-rnn 256 \ + --no-shuffle --seed 1111 --maxi-batch 1 --maxi-batch-sort none --dim-emb 128 --dim-rnn 256 --cost-type ce-mean \ --devices 0 1 \ -m async_sgd/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 20 --after-batches 400 \ diff --git a/tests/training/multi-gpu/test_sync_sgd.sh b/tests/training/multi-gpu/test_sync_sgd.sh index ac2d009..ca29645 100644 --- a/tests/training/multi-gpu/test_sync_sgd.sh +++ b/tests/training/multi-gpu/test_sync_sgd.sh @@ -15,7 +15,7 @@ mkdir -p sync_sgd $MRT_MARIAN/marian \ --no-shuffle --seed 777 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none \ --dim-rnn 64 --dim-emb 32 --learn-rate 0.001 --clip-norm 0 \ - --devices 0 1 --sync-sgd --optimizer sgd \ + --devices 0 1 --sync-sgd --optimizer sgd --cost-type ce-mean \ -m sync_sgd/model.full.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --after-batches 100 \ --log sync_sgd.log diff --git a/tests/training/multi-gpu/test_sync_sgd_1gpu.sh b/tests/training/multi-gpu/test_sync_sgd_1gpu.sh index e3a8255..2b9c72b 100644 --- a/tests/training/multi-gpu/test_sync_sgd_1gpu.sh +++ b/tests/training/multi-gpu/test_sync_sgd_1gpu.sh @@ -10,7 +10,7 @@ mkdir -p sync_sgd_1gpu $MRT_MARIAN/marian \ --no-shuffle --seed 888 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none \ --dim-rnn 64 --dim-emb 32 --learn-rate 0.1 \ - --devices 0 --sync-sgd --optimizer sgd \ + --devices 0 --sync-sgd --optimizer sgd --cost-type ce-mean \ -m sync_sgd_1gpu/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 5 --save-freq 10 --after-batches 20 \ --log sync_sgd_1gpu.log diff --git a/tests/training/multi-gpu/test_sync_sgd_1gpu_expsmooth.sh b/tests/training/multi-gpu/test_sync_sgd_1gpu_expsmooth.sh index 510aa26..d1d77d3 100644 --- a/tests/training/multi-gpu/test_sync_sgd_1gpu_expsmooth.sh +++ b/tests/training/multi-gpu/test_sync_sgd_1gpu_expsmooth.sh @@ -10,7 +10,7 @@ mkdir -p sync_sgd_1gpu_expsmooth $MRT_MARIAN/marian \ --no-shuffle --seed 888 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none \ --dim-rnn 64 --dim-emb 32 --learn-rate 0.1 \ - --devices 0 --sync-sgd --optimizer sgd --exponential-smoothing \ + --devices 0 --sync-sgd --optimizer sgd --exponential-smoothing --cost-type ce-mean \ -m sync_sgd_1gpu_expsmooth/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 5 --save-freq 10 --after-batches 20 \ --log sync_sgd_1gpu_expsmooth.log diff --git a/tests/training/restarting/test_sgd_for_two_epochs.sh b/tests/training/restarting/test_sgd_for_two_epochs.sh index abef975..ad92b5f 100644 --- a/tests/training/restarting/test_sgd_for_two_epochs.sh +++ b/tests/training/restarting/test_sgd_for_two_epochs.sh @@ -8,6 +8,8 @@ rm -rf sgd_2e sgd_1st_epoch.log sgd_2nd_epoch.log mkdir -p sgd_2e extra_opts="--no-shuffle --seed 1111 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Uncomment to prepare the expected output diff --git a/tests/training/restarting/test_sgd_for_two_epochs_sync.sh b/tests/training/restarting/test_sgd_for_two_epochs_sync.sh index 4808894..d3ee295 100644 --- a/tests/training/restarting/test_sgd_for_two_epochs_sync.sh +++ b/tests/training/restarting/test_sgd_for_two_epochs_sync.sh @@ -8,6 +8,8 @@ rm -rf sgd_sync_2e sgd_sync_*_epoch.log mkdir -p sgd_sync_2e extra_opts="--no-shuffle --seed 1111 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd --sync-sgd" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Uncomment to prepare the expected output diff --git a/tests/training/restoring/corpus/test_corpus_restoration.sh b/tests/training/restoring/corpus/test_corpus_restoration.sh index 9937dff..905d2d0 100644 --- a/tests/training/restoring/corpus/test_corpus_restoration.sh +++ b/tests/training/restoring/corpus/test_corpus_restoration.sh @@ -11,6 +11,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 1111 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd --dim-emb 128 --dim-rnn 256 --disp-freq 4" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Step 1: Train a model in one go, up to the update no. 70, and save training logs diff --git a/tests/training/restoring/corpus/test_corpus_restoration_maxi_batch.sh b/tests/training/restoring/corpus/test_corpus_restoration_maxi_batch.sh index c439cb4..c36213f 100644 --- a/tests/training/restoring/corpus/test_corpus_restoration_maxi_batch.sh +++ b/tests/training/restoring/corpus/test_corpus_restoration_maxi_batch.sh @@ -11,6 +11,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 1111 --maxi-batch 20 --mini-batch 32 --optimizer sgd --dim-emb 128 --dim-rnn 256 --disp-freq 4" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" $MRT_MARIAN/marian \ -m corpus_maxi/model_full.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ diff --git a/tests/training/restoring/corpus/test_corpus_restoration_mini_batch_fit.sh b/tests/training/restoring/corpus/test_corpus_restoration_mini_batch_fit.sh index b595126..f32aac2 100644 --- a/tests/training/restoring/corpus/test_corpus_restoration_mini_batch_fit.sh +++ b/tests/training/restoring/corpus/test_corpus_restoration_mini_batch_fit.sh @@ -11,6 +11,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 5555 --maxi-batch 8 --maxi-batch-sort src --mini-batch 32 --mini-batch-fit -w 100 --optimizer sgd --dim-emb 128 --dim-rnn 256 --disp-freq 4" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Step 1: Train a model in one go, up to the update no. 70, and save training logs diff --git a/tests/training/restoring/corpus/test_corpus_restoration_mini_batch_fit_sync.sh b/tests/training/restoring/corpus/test_corpus_restoration_mini_batch_fit_sync.sh index 4b8a2bb..f1b1815 100644 --- a/tests/training/restoring/corpus/test_corpus_restoration_mini_batch_fit_sync.sh +++ b/tests/training/restoring/corpus/test_corpus_restoration_mini_batch_fit_sync.sh @@ -11,6 +11,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 5555 --maxi-batch 8 --maxi-batch-sort src --mini-batch 32 --mini-batch-fit -w 100 --optimizer sgd --dim-emb 128 --transformer-dim-ffn 256 --disp-freq 4 --type transformer --sync-sgd" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Step 1: Train a model in one go, up to the update no. 70, and save training logs diff --git a/tests/training/restoring/corpus/test_corpus_restoration_no_shuffle.sh b/tests/training/restoring/corpus/test_corpus_restoration_no_shuffle.sh index 63105e4..7b6da5e 100644 --- a/tests/training/restoring/corpus/test_corpus_restoration_no_shuffle.sh +++ b/tests/training/restoring/corpus/test_corpus_restoration_no_shuffle.sh @@ -11,6 +11,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 1234 --no-shuffle --maxi-batch 8 --maxi-batch-sort src --mini-batch 32 --mini-batch-fit -w 100 --optimizer sgd --dim-emb 128 --dim-rnn 256 --disp-freq 4" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" $MRT_MARIAN/marian \ -m corpus_noshuf/model_full.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ diff --git a/tests/training/restoring/corpus/test_corpus_restoration_one_epoch.sh b/tests/training/restoring/corpus/test_corpus_restoration_one_epoch.sh index 1001d74..b9e777c 100644 --- a/tests/training/restoring/corpus/test_corpus_restoration_one_epoch.sh +++ b/tests/training/restoring/corpus/test_corpus_restoration_one_epoch.sh @@ -11,6 +11,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 9999 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd --dim-emb 128 --dim-rnn 256 --disp-freq 4" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" $MRT_MARIAN/marian \ -m corpus_one/model_full.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ diff --git a/tests/training/restoring/corpus/test_corpus_restoration_s2s.sh b/tests/training/restoring/corpus/test_corpus_restoration_s2s.sh index ab9b533..cd658c0 100644 --- a/tests/training/restoring/corpus/test_corpus_restoration_s2s.sh +++ b/tests/training/restoring/corpus/test_corpus_restoration_s2s.sh @@ -11,6 +11,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 1111 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd --dim-emb 128 --dim-rnn 256 --disp-freq 4 --type s2s" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Step 1: Train a model in one go, up to the update no. 70, and save training logs diff --git a/tests/training/restoring/corpus/test_corpus_restoration_s2s_sync.sh b/tests/training/restoring/corpus/test_corpus_restoration_s2s_sync.sh index 87cd06b..e8f01ba 100644 --- a/tests/training/restoring/corpus/test_corpus_restoration_s2s_sync.sh +++ b/tests/training/restoring/corpus/test_corpus_restoration_s2s_sync.sh @@ -12,6 +12,8 @@ test -e vocab.en.yml # TODO: Weight decaying in Adam is disabled, because it gives unstable results on GPU extra_opts="--seed 2222 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --dim-emb 128 --dim-rnn 256 --disp-freq 4 --type s2s --sync-sgd --optimizer adam --optimizer-params 0.9 0.98 0" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Step 1: Train a model in one go, up to the update no. 70, and save training logs diff --git a/tests/training/restoring/corpus/test_finetune.sh b/tests/training/restoring/corpus/test_finetune.sh index d9b01d8..1e99645 100644 --- a/tests/training/restoring/corpus/test_finetune.sh +++ b/tests/training/restoring/corpus/test_finetune.sh @@ -17,6 +17,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 2222 --maxi-batch 1 --maxi-batch-sort none --mini-batch 64 --optimizer sgd --dim-emb 128 --dim-rnn 256 --disp-freq 4" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Train a model on a training corpus diff --git a/tests/training/restoring/corpus/test_sqlite_restoration.sh b/tests/training/restoring/corpus/test_sqlite_restoration.sh index 369cfad..7197de8 100644 --- a/tests/training/restoring/corpus/test_sqlite_restoration.sh +++ b/tests/training/restoring/corpus/test_sqlite_restoration.sh @@ -11,6 +11,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 3333 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd --dim-emb 128 --dim-rnn 256 --disp-freq 4" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" $MRT_MARIAN/marian \ -m sqlite/model_full.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ diff --git a/tests/training/restoring/corpus/test_sqlite_restoration_maxi_batch.sh b/tests/training/restoring/corpus/test_sqlite_restoration_maxi_batch.sh index 77305c5..b45e1fa 100644 --- a/tests/training/restoring/corpus/test_sqlite_restoration_maxi_batch.sh +++ b/tests/training/restoring/corpus/test_sqlite_restoration_maxi_batch.sh @@ -11,6 +11,8 @@ test -e vocab.de.yml test -e vocab.en.yml extra_opts="--seed 4444 --maxi-batch 20 --mini-batch 32 --optimizer sgd --dim-emb 128 --dim-rnn 256 --disp-freq 4" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" $MRT_MARIAN/marian \ -m sqlite_maxi/model_full.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ diff --git a/tests/training/restoring/exp-smoothing/test_expsmooth.sh b/tests/training/restoring/exp-smoothing/test_expsmooth.sh index 912ece8..f048018 100644 --- a/tests/training/restoring/exp-smoothing/test_expsmooth.sh +++ b/tests/training/restoring/exp-smoothing/test_expsmooth.sh @@ -8,7 +8,11 @@ rm -rf expsmooth expsmooth_*.log mkdir -p expsmooth -opts="--no-shuffle --seed 777 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.5 --valid-sets valid.bpe.en valid.bpe.de --valid-metrics cross-entropy --valid-mini-batch 32" +opts="--no-shuffle --seed 777 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none" +opts="$opts --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.5" +opts="$opts --valid-sets valid.bpe.en valid.bpe.de --valid-metrics cross-entropy --valid-mini-batch 32" +# Added because default options has changes +opts="$opts --cost-type ce-mean --disp-label-counts false" opt_disp=20 opt_valid=20 diff --git a/tests/training/restoring/exp-smoothing/test_expsmooth_s2s.sh b/tests/training/restoring/exp-smoothing/test_expsmooth_s2s.sh index 863b1f8..1080546 100644 --- a/tests/training/restoring/exp-smoothing/test_expsmooth_s2s.sh +++ b/tests/training/restoring/exp-smoothing/test_expsmooth_s2s.sh @@ -8,7 +8,11 @@ rm -rf expsmooth_s2s expsmooth_s2s_*.log mkdir -p expsmooth_s2s -opts="--no-shuffle --seed 777 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.5 --valid-sets valid.bpe.en valid.bpe.de --valid-metrics cross-entropy --valid-mini-batch 32 --type s2s" +opts="--no-shuffle --seed 777 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none" +opts="$opts --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.5" +opts="$opts --valid-sets valid.bpe.en valid.bpe.de --valid-metrics cross-entropy --valid-mini-batch 32 --type s2s" +# Added because default options has changes +opts="$opts --cost-type ce-mean --disp-label-counts false" opt_disp=20 opt_valid=20 diff --git a/tests/training/restoring/exp-smoothing/test_expsmooth_sync.sh b/tests/training/restoring/exp-smoothing/test_expsmooth_sync.sh index 813b600..3e26acf 100644 --- a/tests/training/restoring/exp-smoothing/test_expsmooth_sync.sh +++ b/tests/training/restoring/exp-smoothing/test_expsmooth_sync.sh @@ -13,7 +13,11 @@ rm -rf expsmooth_sync expsmooth_sync_*.log mkdir -p expsmooth_sync -opts="--no-shuffle --seed 777 --cost-type ce-sum --disp-label-counts --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.001 --valid-sets valid.bpe.en valid.bpe.de --valid-metrics ce-mean-words --valid-mini-batch 32 --devices 0 1 --sync-sgd --clip-norm 0" +opts="--no-shuffle --seed 777 --cost-type ce-sum --disp-label-counts" +opts="$opts --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none" +opts="$opts --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.001" +opts="$opts --valid-sets valid.bpe.en valid.bpe.de --valid-metrics ce-mean-words --valid-mini-batch 32" +opts="$opts --devices 0 1 --sync-sgd --clip-norm 0" opt_disp=20 opt_valid=20 diff --git a/tests/training/restoring/multi-gpu/test_adam_sync.sh b/tests/training/restoring/multi-gpu/test_adam_sync.sh index f97b319..84079d3 100644 --- a/tests/training/restoring/multi-gpu/test_adam_sync.sh +++ b/tests/training/restoring/multi-gpu/test_adam_sync.sh @@ -15,6 +15,8 @@ mkdir -p adam_sync # TODO: The weight decaying in the Adam optimizer is disabled, because it gives # unstable results even for a single GPU trainig, without training restoration. opts="--no-shuffle --seed 777 --mini-batch 2 --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --learn-rate 0.1 --optimizer adam --optimizer-params 0.9 0.98 0 --sync-sgd --devices 0 1" +# Added because default options has changes +opts="$opts --cost-type ce-mean --disp-label-counts false" # Step 1: Training in one go diff --git a/tests/training/restoring/multi-gpu/test_adam_sync_cpu.sh b/tests/training/restoring/multi-gpu/test_adam_sync_cpu.sh index b85a0be..fcfdcd5 100644 --- a/tests/training/restoring/multi-gpu/test_adam_sync_cpu.sh +++ b/tests/training/restoring/multi-gpu/test_adam_sync_cpu.sh @@ -20,6 +20,8 @@ mkdir -p adam_sync_cpu # TODO: The weight decaying in the Adam optimizer is enabled for CPU, because # it gives stable results, in contrary to the GPU version opts="--no-shuffle --seed 777 --mini-batch 2 --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --learn-rate 0.1 --optimizer adam --optimizer-params 0.9 0.98 0.001 --sync-sgd --devices 0 1 --cpu-threads 1" +# Added because default options has changes +opts="$opts --cost-type ce-mean --disp-label-counts false" # Step 1: Training in one go diff --git a/tests/training/restoring/multi-gpu/test_async.sh b/tests/training/restoring/multi-gpu/test_async.sh index c32050d..8e22f30 100644 --- a/tests/training/restoring/multi-gpu/test_async.sh +++ b/tests/training/restoring/multi-gpu/test_async.sh @@ -13,6 +13,8 @@ rm -rf async async_*.log async.*out async.*expected mkdir -p async opts="--no-shuffle --seed 777 --mini-batch 1 --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.1 --devices 0 1" +# Added because default options has changes +opts="$opts --cost-type ce-mean --disp-label-counts false" opt_disp=1 opt_save=8 diff --git a/tests/training/restoring/multi-gpu/test_sync.sh b/tests/training/restoring/multi-gpu/test_sync.sh index c4cd6cb..ff10d23 100644 --- a/tests/training/restoring/multi-gpu/test_sync.sh +++ b/tests/training/restoring/multi-gpu/test_sync.sh @@ -13,6 +13,8 @@ rm -rf sync sync_*.log mkdir -p sync opts="--no-shuffle --seed 777 --mini-batch 4 --maxi-batch 1 --maxi-batch-sort none --dim-rnn 64 --dim-emb 32 --optimizer sgd --learn-rate 0.1 --devices 0 1 --sync-sgd" +# Added because default options has changes +opts="$opts --cost-type ce-mean --disp-label-counts false" $MRT_MARIAN/marian \ diff --git a/tests/training/restoring/optimizer/test_adagrad_params.sh b/tests/training/restoring/optimizer/test_adagrad_params.sh index 4b5d420..8fca356 100644 --- a/tests/training/restoring/optimizer/test_adagrad_params.sh +++ b/tests/training/restoring/optimizer/test_adagrad_params.sh @@ -10,7 +10,7 @@ mkdir -p adagrad $MRT_MARIAN/marian \ --no-shuffle --seed 7777 --maxi-batch 1 --maxi-batch-sort none --dim-emb 128 --dim-rnn 256 \ -m adagrad/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ - --disp-freq 10 --after-batches 100 --save-freq 60 --optimizer adagrad \ + --disp-freq 10 --after-batches 100 --save-freq 60 --optimizer adagrad --cost-type ce-mean \ --log adagrad.log test -e adagrad/model.npz diff --git a/tests/training/restoring/optimizer/test_adam_params.sh b/tests/training/restoring/optimizer/test_adam_params.sh index e70663f..e0bd76f 100644 --- a/tests/training/restoring/optimizer/test_adam_params.sh +++ b/tests/training/restoring/optimizer/test_adam_params.sh @@ -10,7 +10,7 @@ mkdir -p adam $MRT_MARIAN/marian \ --no-shuffle --seed 7777 --maxi-batch 1 --maxi-batch-sort none --dim-emb 128 --dim-rnn 256 \ -m adam/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ - --disp-freq 10 --after-batches 100 --save-freq 60 \ + --disp-freq 10 --after-batches 100 --save-freq 60 --cost-type ce-mean \ --log adam.log test -e adam/model.npz diff --git a/tests/training/restoring/optimizer/test_adam_params_async.sh b/tests/training/restoring/optimizer/test_adam_params_async.sh index cb335c0..2b2c869 100644 --- a/tests/training/restoring/optimizer/test_adam_params_async.sh +++ b/tests/training/restoring/optimizer/test_adam_params_async.sh @@ -15,7 +15,7 @@ fi $MRT_MARIAN/marian \ --no-shuffle --seed 7777 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --dim-emb 128 --dim-rnn 256 \ -m adam_async/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ - --disp-freq 10 --after-batches 100 --save-freq 60 --cost-type ce-sum \ + --disp-freq 10 --after-batches 100 --save-freq 60 --cost-type ce-sum --disp-label-counts false \ --log adam_async.log --devices 0 1 test -e adam_async/model.npz diff --git a/tests/training/restoring/optimizer/test_adam_params_sync.sh b/tests/training/restoring/optimizer/test_adam_params_sync.sh index 6697e9b..1e2481d 100644 --- a/tests/training/restoring/optimizer/test_adam_params_sync.sh +++ b/tests/training/restoring/optimizer/test_adam_params_sync.sh @@ -16,7 +16,7 @@ $MRT_MARIAN/marian \ --no-shuffle --seed 7777 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --dim-emb 128 --dim-rnn 256 \ -m adam_sync/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --after-batches 100 --save-freq 60 \ - --log adam_sync.log --devices 0 1 --sync-sgd --cost-type ce-sum --clip-norm 0 + --log adam_sync.log --devices 0 1 --sync-sgd --cost-type ce-sum --disp-label-counts false --clip-norm 0 test -e adam_sync/model.npz test -e adam_sync/model.npz.optimizer.npz diff --git a/tests/training/restoring/optimizer/test_loading_adam_params.sh b/tests/training/restoring/optimizer/test_loading_adam_params.sh index 00c0cdc..95a48c8 100644 --- a/tests/training/restoring/optimizer/test_loading_adam_params.sh +++ b/tests/training/restoring/optimizer/test_loading_adam_params.sh @@ -8,6 +8,8 @@ rm -rf adam_load adam_load_?.log mkdir -p adam_load extra_opts="--no-shuffle --seed 7777 --maxi-batch 1 --maxi-batch-sort none --mini-batch 2 --dim-rnn 64 --dim-emb 32" +# Added because default options has changes +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" $MRT_MARIAN/marian \ -m adam_load/model.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ diff --git a/tests/training/restoring/validation/test_adding_validator_after_restart.sh b/tests/training/restoring/validation/test_adding_validator_after_restart.sh index e8ef741..ff95d90 100644 --- a/tests/training/restoring/validation/test_adding_validator_after_restart.sh +++ b/tests/training/restoring/validation/test_adding_validator_after_restart.sh @@ -7,10 +7,11 @@ set -e rm -rf valid_add valid_add_?.log mkdir -p valid_add +extra_opts="--no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd" +extra_opts="$extra_opts --dim-emb 128 --dim-rnn 256 --mini-batch 16" +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" -#$MRT_MARIAN/marian \ - #--no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ - #--dim-emb 128 --dim-rnn 256 --mini-batch 16 \ +#$MRT_MARIAN/marian $extra_opts \ #-m valid_add/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ #--disp-freq 10 --valid-freq 20 --after-batches 200 --early-stopping 5 \ #--valid-metrics cross-entropy perplexity \ @@ -21,9 +22,7 @@ mkdir -p valid_add #exit 1 -$MRT_MARIAN/marian \ - --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ - --dim-emb 128 --dim-rnn 256 --mini-batch 16 \ +$MRT_MARIAN/marian $extra_opts \ -m valid_add/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 20 --after-batches 100 --early-stopping 5 \ --valid-metrics cross-entropy \ @@ -37,9 +36,7 @@ test -e valid_add_1.log cp valid_add/model.npz.progress.yml valid_add/model.npz.progress.yml.bac cat valid_add_1.log | $MRT_TOOLS/strip-timestamps.sh > valid_add.out -$MRT_MARIAN/marian \ - --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd \ - --dim-emb 128 --dim-rnn 256 --mini-batch 16 \ +$MRT_MARIAN/marian $extra_opts \ -m valid_add/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 20 --after-batches 200 --early-stopping 5 \ --valid-metrics cross-entropy ce-mean-words \ diff --git a/tests/training/restoring/validation/test_restoring_newbest_validators.sh b/tests/training/restoring/validation/test_restoring_newbest_validators.sh index f1b8dc4..444599b 100644 --- a/tests/training/restoring/validation/test_restoring_newbest_validators.sh +++ b/tests/training/restoring/validation/test_restoring_newbest_validators.sh @@ -15,7 +15,7 @@ head -n 8 $MRT_DATA/europarl.de-en/toy.bpe.de > valid.mini.bpe.de #$MRT_MARIAN/marian \ #--type s2s --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --quiet-translation \ - #--dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd \ + #--dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd --cost-type ce-mean \ #-m valid_newbest/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ #--disp-freq 5 --valid-freq 10 --after-batches 100 \ #--valid-metrics cross-entropy translation --valid-script-path ./count_bytes.sh \ @@ -29,7 +29,7 @@ head -n 8 $MRT_DATA/europarl.de-en/toy.bpe.de > valid.mini.bpe.de $MRT_MARIAN/marian \ --type s2s --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --quiet-translation \ - --dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd \ + --dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd --cost-type ce-mean \ -m valid_newbest/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 5 --valid-freq 10 --after-batches 50 \ --valid-metrics cross-entropy translation --valid-script-path ./count_bytes.sh \ diff --git a/tests/training/restoring/validation/test_restoring_stalled_validators.sh b/tests/training/restoring/validation/test_restoring_stalled_validators.sh index 358c272..50a6d2e 100644 --- a/tests/training/restoring/validation/test_restoring_stalled_validators.sh +++ b/tests/training/restoring/validation/test_restoring_stalled_validators.sh @@ -7,16 +7,18 @@ set -e rm -rf valid_stalled valid_stalled_?.*log valid_script_?.temp mkdir -p valid_stalled +extra_opts="--no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --quiet-translation" +extra_opts="$extra_opts --dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd" +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" + + head -n 8 $MRT_DATA/europarl.de-en/toy.bpe.en > valid.mini.bpe.en head -n 8 $MRT_DATA/europarl.de-en/toy.bpe.de > valid.mini.bpe.de - # Uncomment to re-generate the expected output #mkdir -p valid_stalled_full -#$MRT_MARIAN/marian \ - #--no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --quiet-translation \ - #--dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd \ +#$MRT_MARIAN/marian $extra_opts \ #-m valid_stalled_full/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ #--disp-freq 10 --valid-freq 20 --after-batches 200 --early-stopping 5 \ #--valid-metrics cross-entropy valid-script translation --valid-script-path ./valid_script_ab.sh \ @@ -33,9 +35,7 @@ head -n 8 $MRT_DATA/europarl.de-en/toy.bpe.de > valid.mini.bpe.de #exit 1 -$MRT_MARIAN/marian \ - --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --quiet-translation \ - --dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd \ +$MRT_MARIAN/marian $extra_opts \ -m valid_stalled/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 20 --after-batches 100 --early-stopping 5 \ --valid-metrics cross-entropy valid-script translation --valid-script-path ./valid_script_ab.sh \ @@ -53,9 +53,7 @@ cat valid_stalled_1.log | $MRT_TOOLS/strip-timestamps.sh \ | head -n -1 > valid_stalled.out -$MRT_MARIAN/marian \ - --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --quiet-translation \ - --dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd \ +$MRT_MARIAN/marian $extra_opts \ -m valid_stalled/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 20 --after-batches 200 --early-stopping 5 \ --valid-metrics cross-entropy valid-script translation --valid-script-path ./valid_script_ab.sh \ diff --git a/tests/training/restoring/validation/test_restoring_validation.sh b/tests/training/restoring/validation/test_restoring_validation.sh index bdc81b7..483e408 100644 --- a/tests/training/restoring/validation/test_restoring_validation.sh +++ b/tests/training/restoring/validation/test_restoring_validation.sh @@ -7,9 +7,12 @@ set -e rm -rf valid valid_?.log valid_script.temp mkdir -p valid -$MRT_MARIAN/marian \ - --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none \ - --dim-emb 128 --dim-rnn 256 --mini-batch 16 \ +extra_opts="--no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none" +extra_opts="$extra_opts --dim-emb 128 --dim-rnn 256 --mini-batch 16" +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" + + +$MRT_MARIAN/marian $extra_opts \ -m valid/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 20 --after-batches 150 --early-stopping 5 \ --valid-metrics valid-script cross-entropy --valid-script-path ./valid_script.sh \ @@ -23,9 +26,7 @@ test -e valid_1.log cp valid/model.npz.progress.yml valid/model.npz.progress.yml.bac cat valid_1.log | $MRT_TOOLS/strip-timestamps.sh | grep "valid-script" > valid.out -$MRT_MARIAN/marian \ - --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none \ - --dim-emb 128 --dim-rnn 256 --mini-batch 16 \ +$MRT_MARIAN/marian $extra_opts \ -m valid/model.npz -t $MRT_DATA/europarl.de-en/corpus.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 20 --after-batches 300 --early-stopping 5 \ --valid-metrics valid-script cross-entropy --valid-script-path ./valid_script.sh \ diff --git a/tests/training/restoring/validation/test_restoring_validation_lower_is_better.sh b/tests/training/restoring/validation/test_restoring_validation_lower_is_better.sh index 33ad8d0..a29e534 100644 --- a/tests/training/restoring/validation/test_restoring_validation_lower_is_better.sh +++ b/tests/training/restoring/validation/test_restoring_validation_lower_is_better.sh @@ -7,10 +7,13 @@ set -e rm -rf valid_lowisbet valid_lowisbet_?.log mkdir -p valid_lowisbet +extra_opts="--no-shuffle --seed 1111 --maxi-batch 1 --maxi-batch-sort none" +extra_opts="$extra_opts --dim-emb 64 --dim-rnn 128 --mini-batch 32" +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" + + # Files for the validation sets are swapped intentionally -$MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --maxi-batch 1 --maxi-batch-sort none \ - --dim-emb 64 --dim-rnn 128 --mini-batch 32 \ +$MRT_MARIAN/marian $extra_opts \ -m valid_lowisbet/model.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 30 --after-batches 160 --early-stopping 2 \ --valid-metrics cross-entropy --valid-sets $MRT_DATA/europarl.de-en/toy.bpe.{de,en} --valid-mini-batch 64 \ @@ -24,9 +27,7 @@ cp valid_lowisbet/model.npz.progress.yml valid_lowisbet/model.npz.progress.yml.b cat valid_lowisbet_1.log | $MRT_TOOLS/strip-timestamps.sh | grep "cross-entropy" > valid_lowisbet.out # Files for the validation sets are swapped intentionally -$MRT_MARIAN/marian \ - --no-shuffle --seed 1111 --maxi-batch 1 --maxi-batch-sort none \ - --dim-emb 64 --dim-rnn 128 --mini-batch 32 \ +$MRT_MARIAN/marian $extra_opts \ -m valid_lowisbet/model.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 30 --after-batches 320 --early-stopping 4 \ --valid-metrics cross-entropy --valid-sets $MRT_DATA/europarl.de-en/toy.bpe.{de,en} --valid-mini-batch 64 \ diff --git a/tests/training/restoring/validation/test_valid_reset_stalled.sh b/tests/training/restoring/validation/test_valid_reset_stalled.sh index 02b24a5..e967a1f 100644 --- a/tests/training/restoring/validation/test_valid_reset_stalled.sh +++ b/tests/training/restoring/validation/test_valid_reset_stalled.sh @@ -16,11 +16,13 @@ mkdir -p valid_reset_stalled test -s valid.mini.bpe.en || head -n 8 $MRT_DATA/europarl.de-en/toy.bpe.en > valid.mini.bpe.en test -s valid.mini.bpe.de || head -n 8 $MRT_DATA/europarl.de-en/toy.bpe.de > valid.mini.bpe.de +extra_opts="--no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --quiet-translation" +extra_opts="$extra_opts --dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd" +extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" + # Train a model for a while and stop -$MRT_MARIAN/marian \ - --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --quiet-translation \ - --dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd \ +$MRT_MARIAN/marian $extra_opts \ -m valid_reset_stalled/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 20 --after-batches 140 --early-stopping 5 \ --valid-metrics translation valid-script cross-entropy --valid-script-path ./valid_script_ab.sh \ @@ -36,9 +38,7 @@ cat valid_reset_stalled_1.log | $MRT_TOOLS/strip-timestamps.sh | grep -P "\[vali # Restart training with --valid-reset-stalled -$MRT_MARIAN/marian \ - --no-shuffle --seed 2222 --maxi-batch 1 --maxi-batch-sort none --quiet-translation \ - --dim-emb 64 --dim-rnn 128 --mini-batch 16 --optimizer sgd \ +$MRT_MARIAN/marian $extra_opts \ -m valid_reset_stalled/model.npz -t $MRT_DATA/europarl.de-en/toy.bpe.{en,de} -v vocab.en.yml vocab.de.yml \ --disp-freq 10 --valid-freq 20 --after-batches 200 --early-stopping 5 --valid-reset-stalled \ --valid-metrics translation valid-script cross-entropy --valid-script-path ./valid_script_ab.sh \ |