topic modelling crashing in 3.39

[nicolespooner-aws]

Topic Modelling crashes in newer versions (3.39+). I used 3.37 but Installing add-ons like Text via pip silently upgrades Orange, breaking previously working workflows. I had to download old text add on version and install via command prompt...

[ajdapretnar]

Thank you for the report. Could you clarify a bit?
You were running Orange 3.37 and an older version of Text add-on (do you remember which)? Where topic modeling worked normally?
Then you upgraded the Text add-on and with it Orange? Which caused the old workflows to crash?

[nicolespooner-aws]

I was originally running orange 3.39 - the current download on the orange site (I am an MBA student using it for a social analytics course). Then as soon as the topic modelling would reach 100% it would crash the entire computer (black screen) not just the workflow, even with very small data sets - I tried many adjustments / changes to my own environment but it crashed every single time. So I finally downloaded 3.37 but had to then install and older version of text and network add in to stop it crashing. I used this version for the network add on orange3_network-1.10.1.tar.gz
and this for the text add on orange3_text-1.16.3.tar.gz and then to stop it auto updating run - python.exe -m pip install orange3==3.37. Now it runs perfectly.

[nicolespooner-aws]

Just adding my machine specs where the topic modelling crash was happening:

Processor: 12th Gen Intel(R) Core(TM) i5-12500H, 2500 Mhz, 12 Core(s), 16 Logical Processor(s)
Intel(R) Iris(R) Xe Graphics
BaseBoard Product: X1505ZA
BaseBoard Manufacturer: ASUSTeK COMPUTER INC.
Installed Physical Memory (RAM): 16.0 GB

[ajdapretnar]

Thank you so much for this report! It helps a lot.

So I finally downloaded 3.37 but had to then install and older version of text and network add in to stop it crashing. I used this version for the network add on orange3_network-1.10.1.tar.gz
and this for the text add on orange3_text-1.16.3.tar.gz

Do I understand correctly that you are able to run Orange 3.37 and still use the latest version of Text, i.e. 1.16.3?

[nicolespooner-aws]

Yes - I guess I got mixed up that the add on was an older version I see it is the latest version - so yes, IF I install the text add on inside Orange then the program automatically upgrades to 3.39 and then the topic modelling glitch returns. When I install the text add on 1.16.3 directly using Command Prompt to my 3.37 Orange version it does not upgrade and I have not had the Topic Modeling glitch since....

[ajdapretnar]

I've tested this and the only obvious thing that changes is scikit-learn downgrades from 1.7.2 back to 1.4.2. I am currently absolutely unable to find how this would affect topic modeling, since topic modeling is devoid of any scikit-learn dependencies, nor was I able to find how scikit-learn is used in gensim (which is the primary dependency for topic modeling). I will look into this further, but I'm quite stuck at the moment.

[nicolespooner-aws]

So sorry you are unable to replicate it. Several people in my course had the same issue and had to use 3.37 version. We were using a very small data set so the file size and format were not the issue. I can't risk going back to 3.39 to test it yet since I have more work to do over the next three weeks. But once I am finished I will be happy to try again with 3.39 and see if there are more insights I can share when the Topic Modelling crashes. I will keep the processes window open and see what is running before it goes to black screen if that will help. The flow for our work was / File - Corpus - Pre-process Text - Topic Modeling / if that helps at all.