Different results after coverting to treelite shared library

[winstonzhao]

My XgBoost model is giving different results after converting to a treelite model.

Minimal example:

import numpy as np
import pandas as pd
import pickle
import treelite
import treelite_runtime
import xgboost as xgb
sample = pd.read_parquet("sample.parquet")
features = list(sample.columns)[1:-3]
xgb1 = xgb.Booster(model_file="model_1s_1week.booster")
print(xgb1.predict(xgb.DMatrix(sample[features])))
model = treelite.Model.from_xgboost(xgb1)
model.export_lib(toolchain='gcc', libpath='./model_1s_1week.so', params={'parallel_comp': 32}, verbose=False)
predictor = treelite_runtime.Predictor('./model_1s_1week.so', verbose=False)
print(predictor.predict(treelite_runtime.DMatrix(sample[features], dtype='float32')))
>>>
[-0.00013971  0.00022221 -0.00022852 -0.00027931 -0.00024623 -0.00030208
 -0.00035697 -0.00042081 -0.00032556]
[21:50:54] ../src/compiler/ast/split.cc:29: Parallel compilation enabled; member trees will be divided into 32 translation units.
[-0.01349026 -0.01085109 -0.00022864 -0.00027949 -0.00024635 -0.0003022
  0.00029168 -0.00042087 -0.0003258 ]

Referenced files model_1s_1week.booster and sample.parquet have been attached.
files.zip

I feel like I'm probably doing something extremely basic incorrectly.
Modelled trained with the following:

xgb1 = xgb.XGBRegressor(learning_rate =0.1,
                     n_estimators=1000,
                     max_depth=3,
                     min_child_weight=1,
                     gamma=0.1,
                     subsample=0.75,
                     colsample_bytree=0.8,
                     reg_alpha=0.01,
                     reg_lambda=1.0,
                     predictor='gpu_predictor',
                     # tree_method="exact",
                     grow_policy="lossguide",
                     tree_method="hist",
                     objective= 'reg:squarederror',
                    feature_selector="greedy",
                    top_k=4,
                     seed=27)

xgb1.fit(train_df[features], train_df[target], eval_set=[(forward_df[features], forward_df[target])], early_stopping_rounds=50)

Versions:

xgboost = 0.90
treelite = 3.0.0

[winstonzhao]

Duplicate of dmlc/tl2cgen#11

This can be fixed by changing:
print(predictor.predict(treelite_runtime.DMatrix(sample[features], dtype='float32')))
to
print(predictor.predict(treelite_runtime.DMatrix(sample[features].to_numpy(), dtype='float32')))