Does lightgbm_predict4j suit for the model generated by lightgbm v2.0.5 or higher?

[dragonclouder]

Noticed here exists a demo suit for v2, and i can run the UseageTest in v2 package successfully, while when i use a model generared by lightgbm v2.0.5, it produced a error: "[Tree] Tree model string format error, should contain default_value field". The default_value is contained in models generated by lightgbm's earlier version. The model file cause the error is attached below.
Thanks for any kind help in advance!

lgbm_test.txt

[dragonclouder]

For there are some friends meeting the same problem as list above email me to ask whether i solved it, i share the trick i used which is possibly not right but a compromise method: "I initialized the private double[] defaultValue; in the org.lightgbm.predict4j.v2.Tree with 0s , or you can substitute it with other values reasonable". Welcome masters to share perfect manners.

[benman1]

Please see my PR (or fork) for support of v2.

[dragonclouder]

#4
yeah, thanks a lot for your positive contribution, it seems like that the default values are set to zeros as what i tried above. Is that reasonable or explainable from you point?

[benman1]

What you set as default_value, i.e. what the value should take if it's missing, should be up to the modeller really - I haven't looked into the lightgbm implementation in any detail to see how missing values are handled. It seems that default_value is entirely deprecated in the newer lightgbm releases (and therefore it's also not in the model file when saved from lightgbm). lightgbm_predict4j is behind in that it still expects it.

The way, I've included it, is that you can still include default_value in the model if you want to set it. It's set to 0.0 if not set otherwise.

It might be worthwhile to make sure lightgbm_predict4j (particularly v2) reflects the latest version of the lightgbm format.

[mrchor]

@benman1 Hello! Why feed different features to the model and the predicted results are the same?

[benman1]

Please explain clearly what you mean, maybe give an example, show your code, etc; otherwise, I am afraid it's not possible to answer.

[mrchor]

@benman1 What does default_value mean?
17:46:12.531 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.532 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.533 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.533 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.534 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.535 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.535 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.536 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.536 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.536 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.537 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.537 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.537 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.538 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.539 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.539 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.540 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
17:46:12.541 [main] DEBUG c.j.m.lightgbm.predictor.V2.Tree - Assuming default values as 0.0
I suspect this is the problem.

[mrchor]

@benman1 Hello! In addition, I found that this prediction code has scoring errors when making binary prediction, for example, the label-1 sample probability score will be less than that labele-0 sample.

Excuse me, is this problem related to default_value?
Looking forward to your reply~

[benman1]

I asked you for a clearer description of the problem. I still haven't seen it. I can't answer your question unless you explain clearly. For example, could you break it down so I can replicate it?

…

On Fri, 26 Apr 2019, 13:19 周杰, ***@***.***> wrote: @benman1 <https://github.com/benman1> Hello! In addition, I found that this prediction code has scoring errors when making binary prediction, for example, the label-1 sample probability score will be less than that labele-0 sample. [image: image] <https://user-images.githubusercontent.com/10007145/56806977-13368500-6860-11e9-8885-35825eb3eaad.png> Excuse me, is this problem related to default_value? Looking forward to your reply~ — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#3 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACSJO7EAVDARU3DSZCDXAA3PSLXLXANCNFSM4EMKWJOA> .

[mrchor]

@benman1 I'm really sorry for not explaining my problem clearly. At present, I have found the problem, from my code. Thank you very much for your recent help. Thank you very much.~

[benman1]

No problem. I am happy you find the code useful.

…

On Sun, 28 Apr 2019, 03:38 周杰, ***@***.***> wrote: @benman1 <https://github.com/benman1> I'm really sorry for not explaining my problem clearly. At present, I have found the problem, from my code. Thank you very much for your recent help. Thank you very much.~ — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#3 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACSJO7AB7FR4DJHBO6ILZ7TPSUEZVANCNFSM4EMKWJOA> .

[KoreHuang]

兄弟，这个问题有结论了吗？lightgbm_predict4j到底能否适用于lightgbm v2.0.5更高版本的模型？

[kohry]

I was having trouble with lgbm model 2.3.2 (2020 May) and had defaultvalue null array exception.
changing all defaultValue[node] to 0 helped me.

anyways, thank you for the issue raising here!

( LGBM is so huge, and I don't understand why people don't have much interest on developing java version. All of the people are talking about AI stuff and writing a book, but never take themselves into deep level of a production. phew..)

[buaanie]

@benman1 I'm really sorry for not explaining my problem clearly. At present, I have found the problem, from my code. Thank you very much for your recent help. Thank you very much.~

hello, do you remember what the problem is in your code. I got the difference between lightgbm_predict4j and lightgbm_python, and i dont know what's wrong with my code.

[buaanie]

@benman1 I'm really sorry for not explaining my problem clearly. At present, I have found the problem, from my code. Thank you very much for your recent help. Thank you very much.~

hello, do you remember what the problem is in your code. I got the difference between lightgbm_predict4j and lightgbm_python, and i dont know what's wrong with my code.

i know what happened...this java implement does not support "category feature"..