In XGBoost 1.0.0, we introduced experimental support of using JSON for saving/loading XGBoost models and related
hyper-parameters for training, aiming to replace the old binary internal format with an
open format that can be easily reused. The support for binary format will be continued in
the future until JSON format is no-longer experimental and has satisfying performance.
This tutorial aims to share some basic insights into the JSON serialisation method used in
XGBoost. Without explicitly mentioned, the following sections assume you are using the
JSON format, which can be enabled by providing the file name with .json as file
extension when saving/loading model: booster.save_model('model.json'). More details
below.
Before we get started, XGBoost is a gradient boosting library with focus on tree model, which means inside XGBoost, there are 2 distinct parts:
The model consisting of trees and
Hyperparameters and configurations used for building the model.
If you come from Deep Learning community, then it should be clear to you that there are differences between the neural network structures composed of weights with fixed tensor operations, and the optimizers (like RMSprop) used to train them.
So when one calls booster.save_model (xgb.save in R), XGBoost saves the trees, some model
parameters like number of input columns in trained trees, and the objective function, which combined
to represent the concept of “model” in XGBoost. As for why are we saving the objective as
part of model, that’s because objective controls transformation of global bias (called
base_score in XGBoost). Users can share this model with others for prediction,
evaluation or continue the training with a different set of hyper-parameters etc.
However, this is not the end of story. There are cases where we need to save something more than just the model itself. For example, in distrbuted training, XGBoost performs checkpointing operation. Or for some reasons, your favorite distributed computing framework decide to copy the model from one worker to another and continue the training in there. In such cases, the serialisation output is required to contain enougth information to continue previous training without user providing any parameters again. We consider such scenario as memory snapshot (or memory based serialisation method) and distinguish it with normal model IO operation. Currently, memory snapshot is used in the following places:
Python package: when the Booster object is pickled with the built-in pickle module.
R package: when the xgb.Booster object is persisted with the built-in functions saveRDS
or save.
Other language bindings are still working in progress.
Note
The old binary format doesn’t distinguish difference between model and raw memory serialisation format, it’s a mix of everything, which is part of the reason why we want to replace it with a more robust serialisation method. JVM Package has its own memory based serialisation methods.
To enable JSON format support for model IO (saving only the trees and objective), provide
a filename with .json as file extension:
bst.save_model('model_file_name.json')
xgb.save(bst, 'model_file_name.json')
While for memory snapshot, JSON is the default starting with xgboost 1.3.
We guarantee backward compatibility for models but not for memory snapshots.
Models (trees and objective) use a stable representation, so that models produced in earlier
versions of XGBoost are accessible in later versions of XGBoost. If you’d like to store or archive
your model for long-term storage, use save_model (Python) and xgb.save (R).
On the other hand, memory snapshot (serialisation) captures many stuff internal to XGBoost, and its
format is not stable and is subject to frequent changes. Therefore, memory snapshot is suitable for
checkpointing only, where you persist the complete snapshot of the training configurations so that
you can recover robustly from possible failures and resume the training process. Loading memory
snapshot generated by an earlier version of XGBoost may result in errors or undefined behaviors.
If a model is persisted with pickle.dump (Python) or saveRDS (R), then the model may
not be accessible in later versions of XGBoost.
XGBoost accepts user provided objective and metric functions as an extension. These functions are not saved in model file as they are language dependent features. With Python, user can pickle the model to include these functions in saved binary. One drawback is, the output from pickle is not a stable serialization format and doesn’t work on different Python version nor XGBoost version, not to mention different language environments. Another way to workaround this limitation is to provide these functions again after the model is loaded. If the customized function is useful, please consider making a PR for implementing it inside XGBoost, this way we can have your functions working with different language bindings.
As noted, pickled model is neither portable nor stable, but in some cases the pickled
models are valuable. One way to restore it in the future is to load it back with that
specific version of Python and XGBoost, export the model by calling save_model. To help
easing the mitigation, we created a simple script for converting pickled XGBoost 0.90
Scikit-Learn interface object to XGBoost 1.0.0 native model. Please note that the script
suits simple use cases, and it’s advised not to use pickle when stability is needed. It’s
located in xgboost/doc/python with the name convert_090to100.py. See comments in
the script for more details.
A similar procedure may be used to recover the model persisted in an old RDS file. In R, you are
able to install an older version of XGBoost using the remotes package:
library(remotes)
remotes::install_version("xgboost", "0.90.0.1") # Install version 0.90.0.1
Once the desired version is installed, you can load the RDS file with readRDS and recover the
xgb.Booster object. Then call xgb.save to export the model using the stable representation.
Now you should be able to use the model in the latest version of XGBoost.
XGBoost’s C API, Python API and R API support saving and loading the internal
configuration directly as a JSON string. In Python package:
bst = xgboost.train(...)
config = bst.save_config()
print(config)
or in R:
config <- xgb.config(bst)
print(config)
Will print out something similiar to (not actual output as it’s too long for demonstration):
{
"Learner": {
"generic_parameter": {
"gpu_id": "0",
"gpu_page_size": "0",
"n_jobs": "0",
"random_state": "0",
"seed": "0",
"seed_per_iteration": "0"
},
"gradient_booster": {
"gbtree_train_param": {
"num_parallel_tree": "1",
"predictor": "gpu_predictor",
"process_type": "default",
"tree_method": "gpu_hist",
"updater": "grow_gpu_hist",
"updater_seq": "grow_gpu_hist"
},
"name": "gbtree",
"updater": {
"grow_gpu_hist": {
"gpu_hist_train_param": {
"debug_synchronize": "0",
"gpu_batch_nrows": "0",
"single_precision_histogram": "0"
},
"train_param": {
"alpha": "0",
"cache_opt": "1",
"colsample_bylevel": "1",
"colsample_bynode": "1",
"colsample_bytree": "1",
"default_direction": "learn",
...
"subsample": "1"
}
}
}
},
"learner_train_param": {
"booster": "gbtree",
"disable_default_eval_metric": "0",
"dsplit": "auto",
"objective": "reg:squarederror"
},
"metrics": [],
"objective": {
"name": "reg:squarederror",
"reg_loss_param": {
"scale_pos_weight": "1"
}
}
},
"version": [1, 0, 0]
}
You can load it back to the model generated by same version of XGBoost by:
bst.load_config(config)
This way users can study the internal representation more closely. Please note that some JSON generators make use of locale dependent floating point serialization methods, which is not supported by XGBoost.
XGBoost has a function called dump_model in Booster object, which lets you to export
the model in a readable format like text, json or dot (graphviz). The primary
use case for it is for model interpretation or visualization, and is not supposed to be
loaded back to XGBoost. The JSON version has a schema. See next section for
more info.
Another important feature of JSON format is a documented Schema, based on which one can easily reuse the output model from
XGBoost. Here is the initial draft of JSON schema for the output model (not
serialization, which will not be stable as noted above). It’s subject to change due to
the beta status. For an example of parsing XGBoost tree model, see /demo/json-model.
Please notice the “weight_drop” field used in “dart” booster. XGBoost does not scale tree
leaf directly, instead it saves the weights as a separated array.
{
"$schema": "http://json-schema.org/draft-07/schema#",
"definitions": {
"gbtree": {
"type": "object",
"properties": {
"name": {
"const": "gbtree"
},
"model": {
"type": "object",
"properties": {
"gbtree_model_param": {
"$ref": "#/definitions/gbtree_model_param"
},
"trees": {
"type": "array",
"items": {
"type": "object",
"properties": {
"tree_param": {
"type": "object",
"properties": {
"num_nodes": {
"type": "string"
},
"size_leaf_vector": {
"type": "string"
},
"num_feature": {
"type": "string"
}
},
"required": [
"num_nodes",
"num_feature",
"size_leaf_vector"
]
},
"id": {
"type": "integer"
},
"loss_changes": {
"type": "array",
"items": {
"type": "number"
}
},
"sum_hessian": {
"type": "array",
"items": {
"type": "number"
}
},
"base_weights": {
"type": "array",
"items": {
"type": "number"
}
},
"left_children": {
"type": "array",
"items": {
"type": "integer"
}
},
"right_children": {
"type": "array",
"items": {
"type": "integer"
}
},
"parents": {
"type": "array",
"items": {
"type": "integer"
}
},
"split_indices": {
"type": "array",
"items": {
"type": "integer"
}
},
"split_conditions": {
"type": "array",
"items": {
"type": "number"
}
},
"split_type": {
"type": "array",
"items": {
"type": "integer"
}
},
"default_left": {
"type": "array",
"items": {
"type": "boolean"
}
},
"categories": {
"type": "array",
"items": {
"type": "integer"
}
},
"categories_nodes": {
"type": "array",
"items": {
"type": "integer"
}
},
"categories_segments": {
"type": "array",
"items": {
"type": "integer"
}
},
"categorical_sizes": {
"type": "array",
"items": {
"type": "integer"
}
}
},
"required": [
"tree_param",
"loss_changes",
"sum_hessian",
"base_weights",
"left_children",
"right_children",
"parents",
"split_indices",
"split_conditions",
"default_left",
"categories",
"categories_nodes",
"categories_segments",
"categories_sizes"
]
}
},
"tree_info": {
"type": "array",
"items": {
"type": "integer"
}
}
},
"required": [
"gbtree_model_param",
"trees",
"tree_info"
]
}
},
"required": [
"name",
"model"
]
},
"gbtree_model_param": {
"type": "object",
"properties": {
"num_trees": {
"type": "string"
},
"size_leaf_vector": {
"type": "string"
}
},
"required": [
"num_trees",
"size_leaf_vector"
]
},
"tree_param": {
"type": "object",
"properties": {
"num_nodes": {
"type": "string"
},
"size_leaf_vector": {
"type": "string"
},
"num_feature": {
"type": "string"
}
},
"required": [
"num_nodes",
"num_feature",
"size_leaf_vector"
]
},
"reg_loss_param": {
"type": "object",
"properties": {
"scale_pos_weight": {
"type": "string"
}
}
},
"aft_loss_param": {
"type": "object",
"properties": {
"aft_loss_distribution": {
"type": "string"
},
"aft_loss_distribution_scale": {
"type": "string"
}
}
},
"softmax_multiclass_param": {
"type": "object",
"properties": {
"num_class": { "type": "string" }
}
},
"lambda_rank_param": {
"type": "object",
"properties": {
"num_pairsample": { "type": "string" },
"fix_list_weight": { "type": "string" }
}
}
},
"type": "object",
"properties": {
"version": {
"type": "array",
"items": [
{
"type": "number",
"const": 1
},
{
"type": "number",
"minimum": 0
},
{
"type": "number",
"minimum": 0
}
],
"minItems": 3,
"maxItems": 3
},
"learner": {
"type": "object",
"properties": {
"feature_names": {
"type": "array",
"items": {
"type": "string"
}
},
"feature_types": {
"type": "array",
"items": {
"type": "string"
}
},
"gradient_booster": {
"oneOf": [
{
"$ref": "#/definitions/gbtree"
},
{
"type": "object",
"properties": {
"name": { "const": "gblinear" },
"model": {
"type": "object",
"properties": {
"weights": {
"type": "array",
"items": {
"type": "number"
}
}
}
}
}
},
{
"type": "object",
"properties": {
"name": { "const": "dart" },
"gbtree": {
"$ref": "#/definitions/gbtree"
},
"weight_drop": {
"type": "array",
"items": {
"type": "number"
}
}
},
"required": [
"name",
"gbtree",
"weight_drop"
]
}
]
},
"objective": {
"oneOf": [
{
"type": "object",
"properties": {
"name": { "const": "reg:squarederror" },
"reg_loss_param": { "$ref": "#/definitions/reg_loss_param"}
},
"required": [
"name",
"reg_loss_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "reg:pseudohubererror" },
"reg_loss_param": { "$ref": "#/definitions/reg_loss_param"}
},
"required": [
"name",
"reg_loss_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "reg:squaredlogerror" },
"reg_loss_param": { "$ref": "#/definitions/reg_loss_param"}
},
"required": [
"name",
"reg_loss_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "reg:linear" },
"reg_loss_param": { "$ref": "#/definitions/reg_loss_param"}
},
"required": [
"name",
"reg_loss_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "reg:logistic" },
"reg_loss_param": { "$ref": "#/definitions/reg_loss_param"}
},
"required": [
"name",
"reg_loss_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "binary:logistic" },
"reg_loss_param": { "$ref": "#/definitions/reg_loss_param"}
},
"required": [
"name",
"reg_loss_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "binary:logitraw" },
"reg_loss_param": { "$ref": "#/definitions/reg_loss_param"}
},
"required": [
"name",
"reg_loss_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "count:poisson" },
"poisson_regression_param": {
"type": "object",
"properties": {
"max_delta_step": { "type": "string" }
}
}
},
"required": [
"name",
"poisson_regression_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "reg:tweedie" },
"tweedie_regression_param": {
"type": "object",
"properties": {
"tweedie_variance_power": { "type": "string" }
}
}
},
"required": [
"name",
"tweedie_regression_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "survival:cox" }
},
"required": [ "name" ]
},
{
"type": "object",
"properties": {
"name": { "const": "reg:gamma" }
},
"required": [ "name" ]
},
{
"type": "object",
"properties": {
"name": { "const": "multi:softprob" },
"softmax_multiclass_param": { "$ref": "#/definitions/softmax_multiclass_param"}
},
"required": [
"name",
"softmax_multiclass_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "multi:softmax" },
"softmax_multiclass_param": { "$ref": "#/definitions/softmax_multiclass_param"}
},
"required": [
"name",
"softmax_multiclass_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "rank:pairwise" },
"lambda_rank_param": { "$ref": "#/definitions/lambda_rank_param"}
},
"required": [
"name",
"lambda_rank_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "rank:ndcg" },
"lambda_rank_param": { "$ref": "#/definitions/lambda_rank_param"}
},
"required": [
"name",
"lambda_rank_param"
]
},
{
"type": "object",
"properties": {
"name": { "const": "rank:map" },
"lambda_rank_param": { "$ref": "#/definitions/lambda_rank_param"}
},
"required": [
"name",
"lambda_rank_param"
]
},
{
"type": "object",
"properties": {
"name": {"const": "survival:aft"},
"aft_loss_param": { "$ref": "#/definitions/aft_loss_param"}
}
},
{
"type": "object",
"properties": {
"name": {"const": "binary:hinge"}
}
}
]
},
"learner_model_param": {
"type": "object",
"properties": {
"base_score": { "type": "string" },
"num_class": { "type": "string" },
"num_feature": { "type": "string" }
}
}
},
"required": [
"gradient_booster",
"objective"
]
}
},
"required": [
"version",
"learner"
]
}
Right now using the JSON format incurs longer serialisation time, we have been working on optimizing the JSON implementation to close the gap between binary format and JSON format.