在R环境中部署Amazon sagemaker生成的XGBoost模型

Question

I'm trying to deploy an XGBoost model , which is trained using Amazon sagemaker, in an R environment. The sagemaker-generated model is stored as a Python pickle object.

Using the {reticulate} package in R, I'm able to import the model into R. However, using the model locally in R gives very different predictions as compared to using the model directly on Amazon sagemaker , on the same testing dataset. I suspect there might be issues converting a XGBoost model stored by python into a model usable in R. Here is the relevant code I used to make the conversion:

library(reticulate)
library(xgboost)

model <- py_load_object("sagemaker-model")
# save the model locally, to be reload into R
model$save_model("local-model")
model_R = xgb.load("local-model")

The reason I first save the "sagemaker-model" locally and then use R to read it back is because I want to use native xgboost in R to make predictions, and not rely on reticulate for predictions. However, the predictions are clearly not correct.

Answer 1

The problem is that in Python, xgboost requires np.array as input. So you have to convert the input with the Dmatrix function.

Something like this:

dtrain <- xgb.DMatrix(train$data, label=train$label)