Welcome to mirror list, hosted at ThFree Co, Russian Federation.

github.com/marian-nmt/marian-regression-tests.git - Unnamed repository; edit this file 'description' to name the repository.
summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorRoman Grundkiewicz <rgrundkiewicz@gmail.com>2021-04-14 15:41:24 +0300
committerRoman Grundkiewicz <rgrundkiewicz@gmail.com>2021-04-14 15:41:24 +0300
commitaed6fcfddbcb370080fea540121179ca8173b353 (patch)
treeea654ddb55b2ec8d8d124379a2b369d30814d653
parent7d612ca5e4b27a76f92584dad76d240e34f216d0 (diff)
Add test for early-stopping with 2 metrics
-rw-r--r--tests/training/validation/.gitignore2
-rw-r--r--tests/training/validation/stop_on_any.expected10
-rwxr-xr-xtests/training/validation/stop_on_script.sh9
-rw-r--r--tests/training/validation/test_valid_metrics_stop_on_any.sh42
4 files changed, 63 insertions, 0 deletions
diff --git a/tests/training/validation/.gitignore b/tests/training/validation/.gitignore
index 36a9cb8..1e7feaf 100644
--- a/tests/training/validation/.gitignore
+++ b/tests/training/validation/.gitignore
@@ -13,3 +13,5 @@ trans_maxlen
compare-trans
compare-trans.bleu
*.temp
+stop_on_any
+devset.??
diff --git a/tests/training/validation/stop_on_any.expected b/tests/training/validation/stop_on_any.expected
new file mode 100644
index 0000000..5246a4f
--- /dev/null
+++ b/tests/training/validation/stop_on_any.expected
@@ -0,0 +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. 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
diff --git a/tests/training/validation/stop_on_script.sh b/tests/training/validation/stop_on_script.sh
new file mode 100755
index 0000000..c140ff9
--- /dev/null
+++ b/tests/training/validation/stop_on_script.sh
@@ -0,0 +1,9 @@
+#!/bin/bash
+
+prefix=stop_on_script
+
+num=$(cat $prefix.temp 2>/dev/null || echo 1)
+((num=(num+1)%6))
+echo $num > $prefix.temp
+((num=6-num))
+echo 111.$num
diff --git a/tests/training/validation/test_valid_metrics_stop_on_any.sh b/tests/training/validation/test_valid_metrics_stop_on_any.sh
new file mode 100644
index 0000000..a3ec7a1
--- /dev/null
+++ b/tests/training/validation/test_valid_metrics_stop_on_any.sh
@@ -0,0 +1,42 @@
+#!/bin/bash -x
+
+#####################################################################
+# SUMMARY: Test early stopping after stalling on any stop_on_anyation metric
+# AUTHOR: snukky
+# TAGS: stop_on_any stop_on_any-script
+#####################################################################
+
+# Exit on error
+set -e
+
+# Test code goes here
+rm -rf stop_on_any stop_on_any.log stop_on_script.temp
+mkdir -p stop_on_any
+
+# Start training with a fully trained model so that validation metrics do not improve easily
+cp $MRT_MODELS/wngt19/model.small.npz stop_on_any/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 64 -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 \
+ --valid-script-path ./stop_on_script.sh \
+ --valid-sets $MRT_DATA/europarl.de-en/toy.bpe.{en,de} \
+ --early-stopping 4 \
+ --valid-log stop_on_any.log
+
+test -e stop_on_any/model.npz
+test -e stop_on_any/model.npz.yml
+test -e stop_on_any.log
+
+$MRT_TOOLS/strip-timestamps.sh < stop_on_any.log | grep '\[valid\]'> stop_on_any.out
+$MRT_TOOLS/diff-nums.py stop_on_any.out stop_on_any.expected -p 0.2 -o stop_on_any.diff
+
+# Exit with success code
+exit 0