Update tests for multi-gpu training and exp-smoothing

7 files changed, 23 insertions, 3 deletions

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 \