Add test sets for early-stopping-on first|all|any

tests/training/validation/.gitignore

@@ -14,4 +14,6 @@ compare-trans
 compare-trans.bleu
 *.temp
 stop_on_any
+stop_on_all
+stop_on_1st
 devset.??

tests/training/validation/stop_on_1st.expected

@@ -0,0 +1,18 @@
+[valid] Ep. 1 : Up. 10 : valid-script : 111.4 : new best
+[valid] Ep. 1 : Up. 10 : ce-mean-words : 2.74949e-41 : new best
+[valid] Ep. 1 : Up. 20 : valid-script : 111.3 : stalled 1 times (last best: 111.4)
+[valid] Ep. 1 : Up. 20 : ce-mean-words : 2.74949e-41 : stalled 1 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 30 : valid-script : 111.2 : stalled 2 times (last best: 111.4)
+[valid] Ep. 1 : Up. 30 : ce-mean-words : 2.74949e-41 : stalled 2 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 40 : valid-script : 111.1 : stalled 3 times (last best: 111.4)
+[valid] Ep. 1 : Up. 40 : ce-mean-words : 2.74949e-41 : stalled 3 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 50 : valid-script : 111.6 : new best
+[valid] Ep. 1 : Up. 50 : ce-mean-words : 2.74949e-41 : stalled 4 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 60 : valid-script : 111.5 : stalled 1 times (last best: 111.6)
+[valid] Ep. 1 : Up. 60 : ce-mean-words : 2.74949e-41 : stalled 5 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 70 : valid-script : 111.4 : stalled 2 times (last best: 111.6)
+[valid] Ep. 1 : Up. 70 : ce-mean-words : 2.74949e-41 : stalled 6 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 80 : valid-script : 111.3 : stalled 3 times (last best: 111.6)
+[valid] Ep. 1 : Up. 80 : ce-mean-words : 2.74949e-41 : stalled 7 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 90 : valid-script : 111.2 : stalled 4 times (last best: 111.6)
+[valid] Ep. 1 : Up. 90 : ce-mean-words : 2.74949e-41 : stalled 8 times (last best: 2.74949e-41)

tests/training/validation/stop_on_all.expected

@@ -0,0 +1,18 @@
+[valid] Ep. 1 : Up. 10 : ce-mean-words : 2.74949e-41 : new best
+[valid] Ep. 1 : Up. 10 : valid-script : 111.4 : new best
+[valid] Ep. 1 : Up. 20 : ce-mean-words : 2.74949e-41 : stalled 1 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 20 : valid-script : 111.3 : stalled 1 times (last best: 111.4)
+[valid] Ep. 1 : Up. 30 : ce-mean-words : 2.74949e-41 : stalled 2 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 30 : valid-script : 111.2 : stalled 2 times (last best: 111.4)
+[valid] Ep. 1 : Up. 40 : ce-mean-words : 2.74949e-41 : stalled 3 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 40 : valid-script : 111.1 : stalled 3 times (last best: 111.4)
+[valid] Ep. 1 : Up. 50 : ce-mean-words : 2.74949e-41 : stalled 4 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 50 : valid-script : 111.6 : new best
+[valid] Ep. 1 : Up. 60 : ce-mean-words : 2.74949e-41 : stalled 5 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 60 : valid-script : 111.5 : stalled 1 times (last best: 111.6)
+[valid] Ep. 1 : Up. 70 : ce-mean-words : 2.74949e-41 : stalled 6 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 70 : valid-script : 111.4 : stalled 2 times (last best: 111.6)
+[valid] Ep. 1 : Up. 80 : ce-mean-words : 2.74949e-41 : stalled 7 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 80 : valid-script : 111.3 : stalled 3 times (last best: 111.6)
+[valid] Ep. 1 : Up. 90 : ce-mean-words : 2.74949e-41 : stalled 8 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 90 : valid-script : 111.2 : stalled 4 times (last best: 111.6)

tests/training/validation/stop_on_any.expected

@@ -1,10 +1,10 @@
-[valid] Ep. 1 : Up. 10 : ce-mean-words : 2.74949e-41 : new best
 [valid] Ep. 1 : Up. 10 : valid-script : 111.4 : new best
-[valid] Ep. 1 : Up. 20 : ce-mean-words : 2.74949e-41 : stalled 1 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 10 : ce-mean-words : 2.74949e-41 : new best
 [valid] Ep. 1 : Up. 20 : valid-script : 111.3 : stalled 1 times (last best: 111.4)
-[valid] Ep. 1 : Up. 30 : ce-mean-words : 2.74949e-41 : stalled 2 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 20 : ce-mean-words : 2.74949e-41 : stalled 1 times (last best: 2.74949e-41)
 [valid] Ep. 1 : Up. 30 : valid-script : 111.2 : stalled 2 times (last best: 111.4)
-[valid] Ep. 1 : Up. 40 : ce-mean-words : 2.74949e-41 : stalled 3 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 30 : ce-mean-words : 2.74949e-41 : stalled 2 times (last best: 2.74949e-41)
 [valid] Ep. 1 : Up. 40 : valid-script : 111.1 : stalled 3 times (last best: 111.4)
-[valid] Ep. 1 : Up. 50 : ce-mean-words : 2.74949e-41 : stalled 4 times (last best: 2.74949e-41)
+[valid] Ep. 1 : Up. 40 : ce-mean-words : 2.74949e-41 : stalled 3 times (last best: 2.74949e-41)
 [valid] Ep. 1 : Up. 50 : valid-script : 111.6 : new best
+[valid] Ep. 1 : Up. 50 : ce-mean-words : 2.74949e-41 : stalled 4 times (last best: 2.74949e-41)

tests/training/validation/test_early_stopping_on_1st.sh

@@ -0,0 +1,42 @@
+#!/bin/bash -x
+
+#####################################################################
+# SUMMARY: Test early stopping after stalling on the first validation metric
+# AUTHOR: snukky
+# TAGS: valid valid-script
+#####################################################################
+
+# Exit on error
+set -e
+
+# Test code goes here
+rm -rf stop_on_1st stop_on_1st.log stop_on_script.temp
+mkdir -p stop_on_1st
+
+# Start training with a fully trained model so that validation metrics do not improve easily
+cp $MRT_MODELS/wngt19/model.small.npz stop_on_1st/model.npz
+
+test -s devset.en || head -n 50 $MRT_MODELS/wngt19/newstest2014.en | sed -r 's/@@ //g' > devset.en
+test -s devset.de || head -n 50 $MRT_MODELS/wngt19/newstest2014.de | sed -r 's/@@ //g' > devset.de
+
+# Training sides are intentionaly reversed to test early stopping
+$MRT_MARIAN/marian \
+    --seed 2222 --no-shuffle --clip-norm 1 --maxi-batch 1 --mini-batch 32 -w 2500 \
+    -m stop_on_1st/model.npz -t $MRT_DATA/europarl.de-en/corpus.small.{de,en}.gz \
+    -v $MRT_MODELS/wngt19/en-de.{spm,spm} \
+    --disp-freq 5 --valid-freq 10 --after-batches 200 \
+    --valid-metrics valid-script ce-mean-words \
+    --valid-script-path ./stop_on_script.sh \
+    --valid-sets $MRT_DATA/europarl.de-en/toy.bpe.{en,de} \
+    --valid-log stop_on_1st.log \
+    --early-stopping 4 --early-stopping-on first
+
+test -e stop_on_1st/model.npz
+test -e stop_on_1st/model.npz.yml
+test -e stop_on_1st.log
+
+$MRT_TOOLS/strip-timestamps.sh < stop_on_1st.log | grep '\[valid\]' > stop_on_1st.out
+$MRT_TOOLS/diff-nums.py stop_on_1st.out stop_on_1st.expected -p 0.2 -o stop_on_1st.diff
+
+# Exit with success code
+exit 0

tests/training/validation/test_early_stopping_on_all.sh

@@ -0,0 +1,42 @@
+#!/bin/bash -x
+
+#####################################################################
+# SUMMARY: Test early stopping after stalling on each validation metric
+# AUTHOR: snukky
+# TAGS: valid valid-script
+#####################################################################
+
+# Exit on error
+set -e
+
+# Test code goes here
+rm -rf stop_on_all stop_on_all.log stop_on_script.temp
+mkdir -p stop_on_all
+
+# Start training with a fully trained model so that validation metrics do not improve easily
+cp $MRT_MODELS/wngt19/model.small.npz stop_on_all/model.npz
+
+test -s devset.en || head -n 50 $MRT_MODELS/wngt19/newstest2014.en | sed -r 's/@@ //g' > devset.en
+test -s devset.de || head -n 50 $MRT_MODELS/wngt19/newstest2014.de | sed -r 's/@@ //g' > devset.de
+
+# Training sides are intentionaly reversed to test early stopping
+$MRT_MARIAN/marian \
+    --seed 2222 --no-shuffle --clip-norm 1 --maxi-batch 1 --mini-batch 32 -w 2500 \
+    -m stop_on_all/model.npz -t $MRT_DATA/europarl.de-en/corpus.small.{de,en}.gz \
+    -v $MRT_MODELS/wngt19/en-de.{spm,spm} \
+    --disp-freq 5 --valid-freq 10 --after-batches 200 \
+    --valid-metrics ce-mean-words valid-script \
+    --valid-script-path ./stop_on_script.sh \
+    --valid-sets $MRT_DATA/europarl.de-en/toy.bpe.{en,de} \
+    --valid-log stop_on_all.log \
+    --early-stopping 4 --early-stopping-on all
+
+test -e stop_on_all/model.npz
+test -e stop_on_all/model.npz.yml
+test -e stop_on_all.log
+
+$MRT_TOOLS/strip-timestamps.sh < stop_on_all.log | grep '\[valid\]' > stop_on_all.out
+$MRT_TOOLS/diff-nums.py stop_on_all.out stop_on_all.expected -p 0.2 -o stop_on_all.diff
+
+# Exit with success code
+exit 0

tests/training/validation/test_valid_metrics_stop_on_any.sh → tests/training/validation/test_early_stopping_on_any.sh

@@ -21,21 +21,21 @@ test -s devset.de || head -n 50 $MRT_MODELS/wngt19/newstest2014.de | sed -r 's/@
 
 # Training sides are intentionaly reversed to test early stopping
 $MRT_MARIAN/marian \
-    --seed 2222 --no-shuffle --clip-norm 1 --maxi-batch 1 --mini-batch 64 -w 2500 \
+    --seed 2222 --no-shuffle --clip-norm 1 --maxi-batch 1 --mini-batch 32 -w 2500 \
     -m stop_on_any/model.npz -t $MRT_DATA/europarl.de-en/corpus.small.{de,en}.gz \
     -v $MRT_MODELS/wngt19/en-de.{spm,spm} \
-    --disp-freq 5 --valid-freq 10 --after-batches 100 \
-    --valid-metrics ce-mean-words valid-script \
+    --disp-freq 5 --valid-freq 10 --after-batches 200 \
+    --valid-metrics valid-script ce-mean-words \
     --valid-script-path ./stop_on_script.sh \
     --valid-sets $MRT_DATA/europarl.de-en/toy.bpe.{en,de} \
-    --early-stopping 4 \
-    --valid-log stop_on_any.log
+    --valid-log stop_on_any.log \
+    --early-stopping 4 --early-stopping-on any
 
 test -e stop_on_any/model.npz
 test -e stop_on_any/model.npz.yml
 test -e stop_on_any.log
 
-$MRT_TOOLS/strip-timestamps.sh < stop_on_any.log | grep '\[valid\]'> stop_on_any.out
+$MRT_TOOLS/strip-timestamps.sh < stop_on_any.log | grep '\[valid\]' > stop_on_any.out
 $MRT_TOOLS/diff-nums.py stop_on_any.out stop_on_any.expected -p 0.2 -o stop_on_any.diff
 
 # Exit with success code