Update tests to --clip-norm 0

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

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

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

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

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

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

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

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

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
 

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 \

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

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

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

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

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

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

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 \

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 \

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 \

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 \

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 \

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 \

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 \

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 \

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 \

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 \

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

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

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