Starspace crashes while testing model about starspace HOT 10 CLOSED

@disdi seems like some problem with the binary saved file in your version. Could you try first if the .tsv file works for you? You can try that by changing the first line to the following in ./starspace test (no need to change anything in ./starspace train):
./starspace test -model "${MODELDIR}"/ag_news.tsv

Please let me know if the .tsv format model works for you or not.

from starspace.

On running the below command -
./starspace test -model "${MODELDIR}"/ag_news.tsv, I get the below error -

Start to evaluate trained model:
Empty test file or model path.

"starspace train ..." or "starspace test ..."

The following arguments are mandatory for train:
-trainFile training file path
-model output model file path

The following arguments are mandatory for test:
-testFile test file path
-model model file path

However I ran the starspace with valgrind with below changes-

diff --git a/examples/classification_ag_news.sh b/examples/classification_ag_news.sh
index 49483cb..d7e5207 100755
--- a/examples/classification_ag_news.sh
+++ b/examples/classification_ag_news.sh
@@ -44,7 +44,7 @@ make

echo "Start to train on ag_news data:"

-./starspace train
+G_SLICE=always-malloc G_DEBUG=gc-friendly valgrind -v --tool=memcheck --leak-check=full --num-callers=40 --log-file=valgrind_train.log ./starspace train
-trainFile "${DATADIR}"/ag_news.train
-model "${MODELDIR}"/ag_news
-initRandSd 0.01
@@ -63,7 +63,7 @@ echo "Start to train on ag_news data:"

echo "Start to evaluate trained model:"

-./starspace test
+G_SLICE=always-malloc G_DEBUG=gc-friendly valgrind -v --tool=memcheck --leak-check=full --num-callers=40 --log-file=valgrind_test.log ./starspace test
-model "${MODELDIR}"/ag_news
-testFile "${DATADIR}"/ag_news.test
-ngrams 1
diff --git a/makefile b/makefile
index f36a01e..fd9f18a 100644

I do not get any crash, though both training and testing took long time to complete.
Attaching valgrind logs.
valgrind_logs.zip

from starspace.

@disdi sorry I did not specify the full command to you. For .tsv format model you should try:

./starspace test 
-model "${MODELDIR}"/ag_news.tsv
-testFile "${DATADIR}"/ag_news.test 
-ngrams 1 

I'll take a look at the logs. How long does it take to complete when you run it with valgrind?

from starspace.

Please find the result of below command
./starspace test -model "${MODELDIR}"/ag_news.tsv -testFile "${DATADIR}"/ag_news.test -ngrams 1

attached in the file.
tsv_format.TXT

Valgrind training alone took more than 6 hours. Running the full example script (both training and testing ) was an overnight thing (10-12 hours) with valgrind.

from starspace.

@ledw Any update on this ?

from starspace.

@disdi thanks. It looks like the tsv model works fine -- i.e. it gives you the evaluation metric. The problem occurs with the binary format with your version. I'll investigate on that, meanwhile you can still train model but only use the .tsv format model instead.

from starspace.

closing as no update for a while (can always reopen)

from starspace.

@ledw I believe this happens because ntokens_ is declared as a int32 but the read op tries to write into a int64.

from starspace.

@pkmital yes, I believe that is fixed in the current version.

from starspace.

Absolutely right thank you!

from starspace.