Modifying model file

[kobics]

Hi We would like to modify a model file, but few things seems wrong. 1 When we modify shrinkage/leaf values, and load, sometime the kernel dies 2 When we modify the shrinkage, and load the model (And the kernel does not die), the predictions are the same as the original model, before modifying the shrinkage

Thanks

[jameslamb]

Hi @kobics , thanks for using LightGBM! If you can please provide some more information, an appropriate maintainer will help you.

Can you please answer the questions you were asked when you clicked "New issue"? Copy this content (https://raw.githubusercontent.com/microsoft/LightGBM/master/.github/ISSUE_TEMPLATE/BUG_REPORT.md) and fill it out in a comment here.

[guolinke]

@kobics if you change the model file manually, you should remove tree_size field.

[kobics]

https://raw.githubusercontent.com/microsoft/LightGBM/master/.github/ISSUE_TEMPLATE/BUG_REPORT.md

[kobics]

Hi @kobics , thanks for using LightGBM! If you can please provide some more information, an appropriate maintainer will help you.

Can you please answer the questions you were asked when you clicked "New issue"? Copy this content (https://raw.githubusercontent.com/microsoft/LightGBM/master/.github/ISSUE_TEMPLATE/BUG_REPORT.md) and fill it out in a comment here.

Hi @jameslamb thanks for your reply. please explain what you ask me Thanks

[jameslamb]

it looks like you opened #3232 with the details I'm asking for. That wasn't my intention, but that works!

did @guolinke 's suggestion (https://github.com/microsoft/LightGBM/issues/3232#issuecomment-659409428) help you?

[kobics]

@kobics if you change the model file manually, you should remove tree_size field.

works for the leafs (at this point), when i change the shrinkage and load, it still breaks

[kobics]

it looks like you opened #3232 with the details I'm asking for. That wasn't my intention, but that works!

did @guolinke 's suggestion (#3232 (comment)) help you?

partially :

works for the leafs (at this point), when i change the shrinkage and load, it still breaks

[jameslamb]

Ok I've reopend this issue. Let's keep the conversation here and not use #3232 , so it is one conversation.

From #232 , it looks like you are using version 2.3.0 of the Python package. That version is quite old...could you try installing version 2.3.1 from PyPi (pip install lightgbm) or building from source in this repo?

If you need to be on 2.3.0 for stability reasons (totally understand!), is it possible to provide the model file here and the code that you are using to load it?

Python maintainers like @StrikerRUS will be able to be more helpful, but we need more information than what you've provided so far to give you the best possible help.

[guolinke]

@kobics what do "break" mean? it dies?

2 When we modify the shrinkage, and load the model (And the kernel does not die), the predictions are the same as the original model, before modifying the shrinkage

The shrinkage rate is not used in prediction time. The leaf value is already shrinkage in training. It will not apply again.

[kobics]

@kobics what do "break" mean? it dies?

2 When we modify the shrinkage, and load the model (And the kernel does not die), the predictions are the same as the original model, before modifying the shrinkage

The shrinkage rate is not used in prediction time. The leaf value is already shrinkage in training. It will not apply again.

@guolinke Thanks this is very helpful. Can you tell me if there additional parts of the model I can erase and still be able to : 1 load and predict 2 load and continue to train

Thank you

[guolinke]

@kobics For the tree model, you can edit tree structure and tree outputs (but need to remove the tree_size). For the most remaining fields in tree, there are to record the status in training.

[kobics]

@kobics For the tree model, you can edit tree structure and tree outputs (but need to remove the tree_size). For the most remaining fields in tree, there are to record the status in training.

@guolinke Thanks. is the status recorded used for later training (if we use it as an init model)?

[guolinke]

@kobics I don't think so.

[github-actions[bot]]

This issue has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.