Hello, Thank you for the great library. I have an issue, I train a model in python using xgboost, but when I try to use "from_xgboost_dump" I get the following error: thread 'main' panicked at 'failed to load the model: ParseIntError { kind: Empty }'

I was wondering if it would be related to the version of the used xgboost library?!

Thank you

Hi, I am a user of this wonderful lib, a huge big thanks first!

recently, I try to use anyhow in my application and causing this error

E0277: `dyn std::error::Error` cannot be shared between threads safely  `dyn std::error::Error` cannot be shared bet
ween threads safely  help: the trait `std::marker::Sync` is not implemented for `dyn std::error::Error` note: requir
ed because of the requirements on the impl of `std::marker::Sync` for `std::ptr::Unique<dyn std::error::Error>` note
: required because it appears within the type `std::boxed::Box<dyn std::error::Error>` note: required because of the
 requirements on the impl of `std::convert::From<std::boxed::Box<dyn std::error::Error>>` for `anyhow::Error` note:
required by `std::convert::From::from`

so I modify all the return error type from

Box<dyn std::error::Error>

into

Box<dyn std::error::Error + 'static + Sync + Send>

this may lead to some benefits:

1. able to use this lib in multithreads
2. able to use anyhow this modern application library

hope you will accept this, thank you.

Here is the get_impurity method in decision_tree.rs: for pair in sorted_data.iter() { let (index, feature_value) = *pair; if feature_value == VALUE_TYPE_UNKNOWN { let cv: &CacheValue = &cache.cache_value[index]; s += cv.s; ss += cv.ss; c += cv.c; unknown += 1; } else { break; } } I wanna ask why use break instead of continue to break the loop. I think the purpose of this code is to count and calculate the unknown value of samples.

This PR add auto code coverage into CI.

• https://ci.mesalock-linux.org/mesalock-linux/gbdt-rs/31/25/2

Since all stdout will be written into database. If possible, please remove some logs of examples into stdout to reduce the execution time of the pipeline.

Hey, i would like to change the ValueType from f32 to f64. Is there any way to do it from myself or you have to implement new stuffs ?

I saw in your code source that you defined in this way :

///! For now we only support std::$t using this macro.
/// We will generalize ValueType in future.
macro_rules! def_value_type {
    ($t: tt) => {
        pub type ValueType = $t;
        pub const VALUE_TYPE_MAX: ValueType = std::$t::MAX;
        pub const VALUE_TYPE_MIN: ValueType = std::$t::MIN;
        pub const VALUE_TYPE_UNKNOWN: ValueType = VALUE_TYPE_MIN;
    };
}

// use continous variables for decision tree
def_value_type!(f32);

Can you bind the ValueType to an f64 behind the feature flag ?

Thanks, Alexis D.

Hi,

I'm experiencing small delta between prediction (same model, same inputs), of gdbt-rs and rust-xgboost, using xbtree and logistic regression, (https://github.com/davechallis/rust-xgboost) which is based on the C++ implementation.

I'm researching this at the moment and suspect a few causers:

1. floating point precision differences native to C++ vs Rust
2. different XGB implementation
3. I'm training on python and loading into Rust via the convert script -- so maybe a problem in reading the dump on the Rust side (I assume the save side is OK because its using the C++ lib)

From your experience is this a known issue? or maybe you can point me into a more specific direction to research from what I listed above?

Thanks

UPDATE: I have now narrowed it down to initializing parameters on the Python side vs Rust side. Looks like some of the parameters are not loaded or taking into account differently. When both models in Python and Rust sides are loaded with no parameters - results are equal.

Do you have a way to do this?

Or do I have to split the data - fine to do - but then can I call model.fit() multiple times to update or will it overwrite the model?

Just some small changes

For gbdt::config::Config,

• a ConfigBuilder (following the builder pattern) that has a build() method to create a Config() object is more Rust like.
• The to_string() method is not necessary if you implement Display.
• The config.set_loss function takes str rather than an enum.

Are you open to PRs?

Feature importance shows how valuable each feature was in training a model. It can be used to select features.

It will be a good feature for gbdt-rs to provide feature importance.