Update tests with TSV inputs

39 files changed, 190 insertions, 163 deletions

diff --git a/tests/interface/input-tsv/restore_stdin.expected b/tests/interface/input-tsv/restore_stdin.expected
index b9cc6d4..d26f7ff 100644
--- a/tests/interface/input-tsv/restore_stdin.expected
+++ b/tests/interface/input-tsv/restore_stdin.expected
@@ -1,30 +1,30 @@
-Ep. 1 : Up. 2 : Sen. 32 : Cost 214.87408447
-Ep. 1 : Up. 4 : Sen. 64 : Cost 195.88232422
-Ep. 1 : Up. 6 : Sen. 96 : Cost 222.75996399
-Ep. 1 : Up. 8 : Sen. 128 : Cost 232.49481201
-Ep. 1 : Up. 10 : Sen. 160 : Cost 204.78642273
-Ep. 1 : Up. 12 : Sen. 192 : Cost 256.89501953
-Ep. 1 : Up. 14 : Sen. 224 : Cost 237.73818970
-Ep. 1 : Up. 16 : Sen. 256 : Cost 210.21063232
-Ep. 1 : Up. 18 : Sen. 288 : Cost 178.70904541
-Ep. 1 : Up. 20 : Sen. 320 : Cost 224.30038452
-Ep. 1 : Up. 22 : Sen. 352 : Cost 225.22837830
-Ep. 1 : Up. 24 : Sen. 384 : Cost 210.81533813
-Ep. 1 : Up. 26 : Sen. 416 : Cost 202.19320679
-Ep. 1 : Up. 28 : Sen. 448 : Cost 211.53353882
-Ep. 1 : Up. 30 : Sen. 480 : Cost 209.39002991
-Ep. 1 : Up. 32 : Sen. 512 : Cost 206.38954163
-Ep. 1 : Up. 34 : Sen. 544 : Cost 202.88201904
-Ep. 1 : Up. 36 : Sen. 576 : Cost 192.36555481
-Ep. 1 : Up. 38 : Sen. 608 : Cost 179.21670532
-Ep. 1 : Up. 40 : Sen. 640 : Cost 164.29644775
-Ep. 1 : Up. 42 : Sen. 672 : Cost 187.61584473
-Ep. 1 : Up. 44 : Sen. 704 : Cost 244.09938049
-Ep. 1 : Up. 46 : Sen. 736 : Cost 266.25546265
-Ep. 1 : Up. 48 : Sen. 768 : Cost 197.74813843
-Ep. 1 : Up. 50 : Sen. 800 : Cost 187.12585449
-Ep. 1 : Up. 52 : Sen. 832 : Cost 186.14714050
-Ep. 1 : Up. 54 : Sen. 864 : Cost 227.19046021
-Ep. 1 : Up. 56 : Sen. 896 : Cost 210.90580750
-Ep. 1 : Up. 58 : Sen. 928 : Cost 210.68801880
-Ep. 1 : Up. 60 : Sen. 960 : Cost 182.89875793
+Ep. 1 : Up. 2 : Sen. 32 : Cost 214.83363342
+Ep. 1 : Up. 4 : Sen. 64 : Cost 195.70648193
+Ep. 1 : Up. 6 : Sen. 96 : Cost 222.41781616
+Ep. 1 : Up. 8 : Sen. 128 : Cost 231.91462708
+Ep. 1 : Up. 10 : Sen. 160 : Cost 204.08346558
+Ep. 1 : Up. 12 : Sen. 192 : Cost 255.86239624
+Ep. 1 : Up. 14 : Sen. 224 : Cost 236.60090637
+Ep. 1 : Up. 16 : Sen. 256 : Cost 209.00881958
+Ep. 1 : Up. 18 : Sen. 288 : Cost 177.51702881
+Ep. 1 : Up. 20 : Sen. 320 : Cost 222.74383545
+Ep. 1 : Up. 22 : Sen. 352 : Cost 223.34017944
+Ep. 1 : Up. 24 : Sen. 384 : Cost 208.93505859
+Ep. 1 : Up. 26 : Sen. 416 : Cost 200.02706909
+Ep. 1 : Up. 28 : Sen. 448 : Cost 209.29515076
+Ep. 1 : Up. 30 : Sen. 480 : Cost 207.00128174
+Ep. 1 : Up. 32 : Sen. 512 : Cost 203.81817627
+Ep. 1 : Up. 34 : Sen. 544 : Cost 200.10937500
+Ep. 1 : Up. 36 : Sen. 576 : Cost 189.81176758
+Ep. 1 : Up. 38 : Sen. 608 : Cost 176.77787781
+Ep. 1 : Up. 40 : Sen. 640 : Cost 161.60902405
+Ep. 1 : Up. 42 : Sen. 672 : Cost 184.40527344
+Ep. 1 : Up. 44 : Sen. 704 : Cost 239.88012695
+Ep. 1 : Up. 46 : Sen. 736 : Cost 262.33227539
+Ep. 1 : Up. 48 : Sen. 768 : Cost 194.13323975
+Ep. 1 : Up. 50 : Sen. 800 : Cost 183.32736206
+Ep. 1 : Up. 52 : Sen. 832 : Cost 181.78253174
+Ep. 1 : Up. 54 : Sen. 864 : Cost 222.31034851
+Ep. 1 : Up. 56 : Sen. 896 : Cost 206.36886597
+Ep. 1 : Up. 58 : Sen. 928 : Cost 205.69429016
+Ep. 1 : Up. 60 : Sen. 960 : Cost 178.27331543
diff --git a/tests/interface/input-tsv/test_tsv_train.sh b/tests/interface/input-tsv/test_tsv_train.sh
index 72b87d8..822b249 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 87f8c7c..75a2537 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 88b5b62..7cad7d2 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 a23e7ac..0a3da3f 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 26d20a6..b257ea6 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 07f2eab..0f824ac 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 42afeac..401c7f9 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 \
-    --cost-type ce-mean --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 --clip-norm 0 \
     -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 0b4ba42..81e9410 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 f8953ef..36b7685 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
@@ -11,7 +11,7 @@ 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 --cost-type ce-mean"
+extra_opts="--no-shuffle --clip-norm 0 --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 \
diff --git a/tests/interface/input-tsv/test_tsv_train_shuffle.sh b/tests/interface/input-tsv/test_tsv_train_shuffle.sh
index e5ca216..9dd4ac9 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 \
-    --cost-type ce-mean --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 --clip-norm 0 \
     -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 8ce67c3..c6e0421 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 \
-    --cost-type ce-mean --shuffle-in-ram --seed 1111 --dim-emb 32 --dim-rnn 64 --maxi-batch 10 --optimizer sgd \
+    --cost-type ce-mean --shuffle-in-ram --clip-norm 0 --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 2a77516..7050b36 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 af35c5c..654f019 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 \
-    --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 \
+    --cost-type ce-mean --no-shuffle --clip-norm 0 --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 e9b0b33..d76ad80 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 \
-    --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 \
+    --cost-type ce-mean --no-shuffle --clip-norm 0 --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 e9a25c6..8ebd9ce 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 \
-    --cost-type ce-mean --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 --clip-norm 0 --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 c101d51..a95ad9d 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 \
-    --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 \
+    --cost-type ce-mean --no-shuffle --clip-norm 0 --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.001 \
     -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 b02488a..efa1793 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 \
-    --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 \
+    --cost-type ce-mean --no-shuffle --clip-norm 1 --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 8653a67..c0e175e 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 \
-    --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 \
+    --cost-type ce-mean --no-shuffle --clip-norm 1 --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 55f9995..34c829f 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 \
-    --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 \
+    --cost-type ce-mean --no-shuffle --clip-norm 0 --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.001 \
     -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 d02f8dc..7a0503c 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 \
-    --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 \
+    --cost-type ce-mean --clip-norm 1 --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 2c4dc1a..49e74de 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 \
-    --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 \
+    --cost-type ce-mean --shuffle-in-ram --clip-norm 1 --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 b266f3d..6eb5386 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 \
-    --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 \
+    --cost-type ce-mean --no-shuffle --clip-norm 0 --seed 5555 --dim-emb 32 --dim-rnn 64 --maxi-batch 1 --maxi-batch-sort none --optimizer sgd --learn-rate 0.001 \
     -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 6528713..5b08134 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 \
-    --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 \
+    --cost-type ce-mean --no-shuffle --clip-norm 1 --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 ee4522f..5065dd3 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 \
-    --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 \
+    --cost-type ce-mean --no-shuffle --clip-norm 1 --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/interface/input-tsv/test_tsv_valid.sh b/tests/interface/input-tsv/test_tsv_valid.sh
index df72551..94c46f8 100644
--- a/tests/interface/input-tsv/test_tsv_valid.sh
+++ b/tests/interface/input-tsv/test_tsv_valid.sh
@@ -18,7 +18,7 @@ test -e valid/vocab.spm || cp $MRT_MODELS/rnn-spm/vocab.deen.spm valid/vocab.spm
 
 # Train
 $MRT_MARIAN/marian \
-    --seed 2222 --no-shuffle --mini-batch 32 --maxi-batch 1 --optimizer sgd \
+    --seed 2222 --no-shuffle --clip-norm 1 --mini-batch 32 --maxi-batch 1 --optimizer sgd \
     -m valid/model.npz --tsv -t train.tsv -v valid/vocab.{spm,spm} \
     --disp-freq 20 --valid-freq 30 --after-batches 30 \
     --valid-metrics cross-entropy translation --valid-translation-output valid.out \
diff --git a/tests/interface/input-tsv/train.expected b/tests/interface/input-tsv/train.expected
index 1cf2ecf..ed76575 100644
--- a/tests/interface/input-tsv/train.expected
+++ b/tests/interface/input-tsv/train.expected
@@ -1,5 +1,5 @@
-261.83258057
-257.78665161
-262.48895264
-253.13388062
-234.16917419
+261.58084106
+256.52697754
+260.31454468
+250.19743347
+230.69422913
diff --git a/tests/interface/input-tsv/train_align.expected b/tests/interface/input-tsv/train_align.expected
index 2de51f1..6d1c8b3 100644
--- a/tests/interface/input-tsv/train_align.expected
+++ b/tests/interface/input-tsv/train_align.expected
@@ -1,25 +1,25 @@
-272.57867432
-267.45211792
-245.10440063
-243.12583923
-254.65167236
-251.95730591
-259.63885498
-243.55068970
-236.45735168
-215.81468201
-212.01930237
-222.31713867
-220.01065063
-230.49443054
-219.36715698
-214.80720520
-198.64233398
-195.14010620
-205.54002380
-204.59991455
-215.78044128
-205.05665588
-201.84078979
-187.56027222
-184.19506836
+267.55108643
+238.90954590
+203.89498901
+194.40493774
+201.26391602
+196.71656799
+205.61053467
+190.75955200
+190.06002808
+175.82437134
+171.81805420
+183.65437317
+183.88174438
+195.11131287
+181.43255615
+182.07211304
+168.76817322
+167.46075439
+175.70928955
+179.71203613
+187.29899597
+175.05770874
+175.59832764
+165.63943481
+163.86834717
diff --git a/tests/interface/input-tsv/train_align_stdin.expected b/tests/interface/input-tsv/train_align_stdin.expected
index a468d22..e7a4610 100644
--- a/tests/interface/input-tsv/train_align_stdin.expected
+++ b/tests/interface/input-tsv/train_align_stdin.expected
@@ -1,7 +1,7 @@
-272.57867432
-267.45211792
-245.10440063
-243.12583923
-254.65167236
-251.95730591
-259.63885498
+267.55108643
+238.90954590
+203.89497375
+194.40493774
+201.26391602
+196.71656799
+205.61053467
diff --git a/tests/interface/input-tsv/train_empty_lines.expected b/tests/interface/input-tsv/train_empty_lines.expected
index bf4b72e..acc5403 100644
--- a/tests/interface/input-tsv/train_empty_lines.expected
+++ b/tests/interface/input-tsv/train_empty_lines.expected
@@ -1,16 +1,16 @@
-270.58209229
-212.28765869
-285.14007568
-286.86123657
-168.05017090
-281.66876221
-267.08026123
-269.00738525
-262.89984131
-224.43609619
-284.56796265
-281.34075928
-231.55950928
-286.07806396
-249.47781372
-264.77264404
+270.44882202
+211.79843140
+284.02545166
+285.28485107
+166.57519531
+279.17941284
+264.18832397
+265.68725586
+259.24093628
+221.06784058
+279.52667236
+276.05969238
+226.25883484
+279.85809326
+242.84822083
+257.72427368
diff --git a/tests/interface/input-tsv/train_fit.expected b/tests/interface/input-tsv/train_fit.expected
index 1c2d2eb..c577e77 100644
--- a/tests/interface/input-tsv/train_fit.expected
+++ b/tests/interface/input-tsv/train_fit.expected
@@ -1,5 +1,5 @@
-251.22476196
-252.45635986
-252.93251038
-258.67086792
-232.73229980
+250.19946289
+249.29103088
+247.88410950
+250.90098572
+222.61479187
diff --git a/tests/interface/input-tsv/train_fit_stdin.expected b/tests/interface/input-tsv/train_fit_stdin.expected
index b9ce675..8d06b37 100644
--- a/tests/interface/input-tsv/train_fit_stdin.expected
+++ b/tests/interface/input-tsv/train_fit_stdin.expected
@@ -1,3 +1,3 @@
-344.08602905
-173.01716614
-248.64839172
+342.83029175
+170.86856079
+244.25839233
diff --git a/tests/interface/input-tsv/train_lm.expected b/tests/interface/input-tsv/train_lm.expected
index c768644..fdbed00 100644
--- a/tests/interface/input-tsv/train_lm.expected
+++ b/tests/interface/input-tsv/train_lm.expected
@@ -1,5 +1,5 @@
-274.50836182
-274.87689209
-266.24481201
-259.36730957
-235.45114136
+274.36938477
+274.21545410
+265.08605957
+257.78823853
+233.71450806
diff --git a/tests/interface/input-tsv/train_shuffle.expected b/tests/interface/input-tsv/train_shuffle.expected
index 912fc9a..f9d2a72 100644
--- a/tests/interface/input-tsv/train_shuffle.expected
+++ b/tests/interface/input-tsv/train_shuffle.expected
@@ -1,5 +1,5 @@
-216.91867065
-261.72125244
-267.90841675
-266.43109131
-300.19824219
+216.37680054
+259.30813599
+263.75015259
+260.62384033
+291.27304077
diff --git a/tests/interface/input-tsv/train_stdin_2e.expected b/tests/interface/input-tsv/train_stdin_2e.expected
index 279c4ff..f80f79b 100644
--- a/tests/interface/input-tsv/train_stdin_2e.expected
+++ b/tests/interface/input-tsv/train_stdin_2e.expected
@@ -1,6 +1,6 @@
-262.65640259
-245.10810852
-248.28816223
-262.62274170
-245.00149536
-248.09956360
+261.66619873
+242.05940247
+243.13801575
+238.52865601
+317.93377686
+355.64865112
diff --git a/tests/interface/input-tsv/train_vocab.expected b/tests/interface/input-tsv/train_vocab.expected
index 7e5f15e..2196992 100644
--- a/tests/interface/input-tsv/train_vocab.expected
+++ b/tests/interface/input-tsv/train_vocab.expected
@@ -1,10 +1,10 @@
-221.76995850
-213.01387024
-217.94252014
-216.67770386
-209.55206299
-233.91926575
-206.07537842
-231.02885437
-206.19801331
-222.06900024
+221.08058167
+211.53454590
+216.04510498
+213.98677063
+207.70535278
+230.34349060
+201.30010986
+224.49655151
+201.09184265
+214.30133057
diff --git a/tests/interface/input-tsv/train_vocabs.expected b/tests/interface/input-tsv/train_vocabs.expected
index f003061..18492a4 100644
--- a/tests/interface/input-tsv/train_vocabs.expected
+++ b/tests/interface/input-tsv/train_vocabs.expected
@@ -1,10 +1,10 @@
-225.92840576
-209.46032715
-223.60330200
-207.56042480
-223.16098022
-198.65243530
-225.91799927
-209.45040894
-223.59326172
-207.55093384
+225.66580200
+208.27442932
+221.54891968
+204.83132935
+219.20014954
+194.64096069
+220.58721924
+203.36322021
+216.46800232
+199.75990295
diff --git a/tests/interface/input-tsv/train_vocabs_yml.expected b/tests/interface/input-tsv/train_vocabs_yml.expected
index becf7bf..c6b5967 100644
--- a/tests/interface/input-tsv/train_vocabs_yml.expected
+++ b/tests/interface/input-tsv/train_vocabs_yml.expected
@@ -1,10 +1,10 @@
-200.10346985
-199.77453613
-196.38256836
-185.44483948
-168.59661865
-192.99969482
-181.77833557
-177.61343384
-200.09066772
-199.76245117
+199.88481140
+198.79373169
+194.57501221
+183.03530884
+165.68594360
+189.14419556
+177.06405640
+172.28703308
+193.72500610
+192.91064453
diff --git a/tests/interface/input-tsv/update_outputs.sh b/tests/interface/input-tsv/update_outputs.sh
new file mode 100755
index 0000000..82436fb
--- /dev/null
+++ b/tests/interface/input-tsv/update_outputs.sh
@@ -0,0 +1,27 @@
+#!/bin/sh -x
+cp train.out train.expected
+cp train_stdin2.out train.expected
+cp train_vocab.out train_vocab.expected
+cp train_vocabs.out train_vocabs.expected
+cp train_vocabs.de.spm.out train_vocabs.de.spm.expected
+cp train_vocabs.en.spm.out train_vocabs.en.spm.expected
+cp train_vocabs_yml.out train_vocabs_yml.expected
+cp train_intypes_stdin.out train.expected
+cp train_fit.out train_fit.expected
+cp train_fit_stdin.out train_fit_stdin.expected
+cp restore_stdin.out restore_stdin.expected
+cp train_shuffle.out train_shuffle.expected
+cp train_shuffle_ram.out train_shuffle.expected
+cp train_stdin.out train.expected
+cp train_stdin_2e.out train_stdin_2e.expected
+cp train_empty_lines.out train_empty_lines.expected
+cp train_lm.out train_lm.expected
+cp train_align.out train_align.expected
+cp train_align_weights.out train_align_weights.expected
+cp train_align_weights_intypes.out train_align_weights.expected
+cp train_align0.out train_align.expected
+cp train_align_shuffle.out train_align_shuffle.expected
+cp train_align_shuffle_ram.out train_align_shuffle.expected
+cp train_align_stdin.out train_align_stdin.expected
+cp train_weights.out train_weights.expected
+cp train_weights0.out train_weights.expected