Predicting Columns in a Table - Deployment Optimization

This tutorial will cover how to perform the end-to-end AutoML process to create an optimized and deployable AutoGluon artifact for production usage.

This tutorial assumes you have already read Predicting Columns in a Table - Quick Start and Predicting Columns in a Table - In Depth.

Fitting a TabularPredictor

We will again use the AdultIncome dataset as in the previous tutorials and train a predictor to predict whether the person’s income exceeds $50,000 or not, which is recorded in the class column of this table.

from autogluon.tabular import TabularDataset, TabularPredictor
train_data = TabularDataset('https://autogluon.s3.amazonaws.com/datasets/Inc/train.csv')
label = 'class'
subsample_size = 500  # subsample subset of data for faster demo, try setting this to much larger values
train_data = train_data.sample(n=subsample_size, random_state=0)
train_data.head()

	age	workclass	fnlwgt	education	education-num	marital-status	occupation	relationship	race	sex	capital-gain	capital-loss	hours-per-week	native-country	class
6118	51	Private	39264	Some-college	10	Married-civ-spouse	Exec-managerial	Wife	White	Female	0	0	40	United-States	>50K
23204	58	Private	51662	10th	6	Married-civ-spouse	Other-service	Wife	White	Female	0	0	8	United-States	<=50K
29590	40	Private	326310	Some-college	10	Married-civ-spouse	Craft-repair	Husband	White	Male	0	0	44	United-States	<=50K
18116	37	Private	222450	HS-grad	9	Never-married	Sales	Not-in-family	White	Male	0	2339	40	El-Salvador	<=50K
33964	62	Private	109190	Bachelors	13	Married-civ-spouse	Exec-managerial	Husband	White	Male	15024	0	40	United-States	>50K

save_path = 'agModels-predictClass-deployment'  # specifies folder to store trained models
predictor = TabularPredictor(label=label, path=save_path).fit(train_data)

Verbosity: 2 (Standard Logging)
=================== System Info ===================
AutoGluon Version:  1.4.1b20250805
Python Version:     3.12.10
Operating System:   Linux
Platform Machine:   x86_64
Platform Version:   #1 SMP Wed Mar 12 14:53:59 UTC 2025
CPU Count:          8
Memory Avail:       28.90 GB / 30.95 GB (93.4%)
Disk Space Avail:   206.84 GB / 255.99 GB (80.8%)
===================================================
No presets specified! To achieve strong results with AutoGluon, it is recommended to use the available presets. Defaulting to `'medium'`...
	Recommended Presets (For more details refer to https://auto.gluon.ai/stable/tutorials/tabular/tabular-essentials.html#presets):
	presets='extreme' : New in v1.4: Massively better than 'best' on datasets <30000 samples by using new models meta-learned on https://tabarena.ai: TabPFNv2, TabICL, Mitra, and TabM. Absolute best accuracy. Requires a GPU. Recommended 64 GB CPU memory and 32+ GB GPU memory.
	presets='best'    : Maximize accuracy. Recommended for most users. Use in competitions and benchmarks.
	presets='high'    : Strong accuracy with fast inference speed.
	presets='good'    : Good accuracy with very fast inference speed.
	presets='medium'  : Fast training time, ideal for initial prototyping.
Using hyperparameters preset: hyperparameters='default'
Beginning AutoGluon training ...
AutoGluon will save models to "/home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment"
Train Data Rows:    500
Train Data Columns: 14
Label Column:       class
AutoGluon infers your prediction problem is: 'binary' (because only two unique label-values observed).
	2 unique label values:  [' >50K', ' <=50K']
	If 'binary' is not the correct problem_type, please manually specify the problem_type parameter during Predictor init (You may specify problem_type as one of: ['binary', 'multiclass', 'regression', 'quantile'])
Problem Type:       binary
Preprocessing data ...
Selected class <--> label mapping:  class 1 =  >50K, class 0 =  <=50K
	Note: For your binary classification, AutoGluon arbitrarily selected which label-value represents positive ( >50K) vs negative ( <=50K) class.
	To explicitly set the positive_class, either rename classes to 1 and 0, or specify positive_class in Predictor init.
Using Feature Generators to preprocess the data ...
Fitting AutoMLPipelineFeatureGenerator...
	Available Memory:                    29596.02 MB
	Train Data (Original)  Memory Usage: 0.25 MB (0.0% of available memory)
	Inferring data type of each feature based on column values. Set feature_metadata_in to manually specify special dtypes of the features.
	Stage 1 Generators:
		Fitting AsTypeFeatureGenerator...
			Note: Converting 1 features to boolean dtype as they only contain 2 unique values.
	Stage 2 Generators:
		Fitting FillNaFeatureGenerator...
	Stage 3 Generators:
		Fitting IdentityFeatureGenerator...
		Fitting CategoryFeatureGenerator...
			Fitting CategoryMemoryMinimizeFeatureGenerator...
	Stage 4 Generators:
		Fitting DropUniqueFeatureGenerator...
	Stage 5 Generators:
		Fitting DropDuplicatesFeatureGenerator...
	Types of features in original data (raw dtype, special dtypes):
		('int', [])    : 6 | ['age', 'fnlwgt', 'education-num', 'capital-gain', 'capital-loss', ...]
		('object', []) : 8 | ['workclass', 'education', 'marital-status', 'occupation', 'relationship', ...]
	Types of features in processed data (raw dtype, special dtypes):
		('category', [])  : 7 | ['workclass', 'education', 'marital-status', 'occupation', 'relationship', ...]
		('int', [])       : 6 | ['age', 'fnlwgt', 'education-num', 'capital-gain', 'capital-loss', ...]
		('int', ['bool']) : 1 | ['sex']
	0.1s = Fit runtime
	14 features in original data used to generate 14 features in processed data.
	Train Data (Processed) Memory Usage: 0.03 MB (0.0% of available memory)
Data preprocessing and feature engineering runtime = 0.08s ...
AutoGluon will gauge predictive performance using evaluation metric: 'accuracy'
	To change this, specify the eval_metric parameter of Predictor()
Automatically generating train/validation split with holdout_frac=0.2, Train Rows: 400, Val Rows: 100
User-specified model hyperparameters to be fit:
{
	'NN_TORCH': [{}],
	'GBM': [{'extra_trees': True, 'ag_args': {'name_suffix': 'XT'}}, {}, {'learning_rate': 0.03, 'num_leaves': 128, 'feature_fraction': 0.9, 'min_data_in_leaf': 3, 'ag_args': {'name_suffix': 'Large', 'priority': 0, 'hyperparameter_tune_kwargs': None}}],
	'CAT': [{}],
	'XGB': [{}],
	'FASTAI': [{}],
	'RF': [{'criterion': 'gini', 'ag_args': {'name_suffix': 'Gini', 'problem_types': ['binary', 'multiclass']}}, {'criterion': 'entropy', 'ag_args': {'name_suffix': 'Entr', 'problem_types': ['binary', 'multiclass']}}, {'criterion': 'squared_error', 'ag_args': {'name_suffix': 'MSE', 'problem_types': ['regression', 'quantile']}}],
	'XT': [{'criterion': 'gini', 'ag_args': {'name_suffix': 'Gini', 'problem_types': ['binary', 'multiclass']}}, {'criterion': 'entropy', 'ag_args': {'name_suffix': 'Entr', 'problem_types': ['binary', 'multiclass']}}, {'criterion': 'squared_error', 'ag_args': {'name_suffix': 'MSE', 'problem_types': ['regression', 'quantile']}}],
}
Fitting 11 L1 models, fit_strategy="sequential" ...
Fitting model: LightGBMXT ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.9 GB
	0.83	 = Validation score   (accuracy)
	0.39s	 = Training   runtime
	0.0s	 = Validation runtime
Fitting model: LightGBM ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.9 GB
	0.85	 = Validation score   (accuracy)
	0.22s	 = Training   runtime
	0.0s	 = Validation runtime
Fitting model: RandomForestGini ...
	Fitting with cpus=8, gpus=0
	0.84	 = Validation score   (accuracy)
	0.56s	 = Training   runtime
	0.05s	 = Validation runtime
Fitting model: RandomForestEntr ...
	Fitting with cpus=8, gpus=0
	0.83	 = Validation score   (accuracy)
	0.52s	 = Training   runtime
	0.05s	 = Validation runtime
Fitting model: CatBoost ...
	Fitting with cpus=4, gpus=0
	0.85	 = Validation score   (accuracy)
	0.8s	 = Training   runtime
	0.0s	 = Validation runtime
Fitting model: ExtraTreesGini ...
	Fitting with cpus=8, gpus=0
	0.82	 = Validation score   (accuracy)
	0.53s	 = Training   runtime
	0.05s	 = Validation runtime
Fitting model: ExtraTreesEntr ...
	Fitting with cpus=8, gpus=0
	0.81	 = Validation score   (accuracy)
	0.53s	 = Training   runtime
	0.06s	 = Validation runtime
Fitting model: NeuralNetFastAI ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.8 GB
	0.83	 = Validation score   (accuracy)
	2.69s	 = Training   runtime
	0.01s	 = Validation runtime
Fitting model: XGBoost ...
	Fitting with cpus=4, gpus=0
	0.85	 = Validation score   (accuracy)
	0.38s	 = Training   runtime
	0.01s	 = Validation runtime
Fitting model: NeuralNetTorch ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.5 GB
	0.83	 = Validation score   (accuracy)
	2.21s	 = Training   runtime
	0.01s	 = Validation runtime
Fitting model: LightGBMLarge ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.4 GB
	0.83	 = Validation score   (accuracy)
	0.66s	 = Training   runtime
	0.0s	 = Validation runtime
Fitting model: WeightedEnsemble_L2 ...
	Ensemble Weights: {'LightGBM': 1.0}
	0.85	 = Validation score   (accuracy)
	0.07s	 = Training   runtime
	0.0s	 = Validation runtime
AutoGluon training complete, total runtime = 10.01s ... Best model: WeightedEnsemble_L2 | Estimated inference throughput: 19879.2 rows/s (100 batch size)
Disabling decision threshold calibration for metric `accuracy` due to having fewer than 10000 rows of validation data for calibration, to avoid overfitting (100 rows).
	`accuracy` is generally not improved through threshold calibration. Force calibration via specifying `calibrate_decision_threshold=True`.
TabularPredictor saved. To load, use: predictor = TabularPredictor.load("/home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment")

Next, load separate test data to demonstrate how to make predictions on new examples at inference time:

test_data = TabularDataset('https://autogluon.s3.amazonaws.com/datasets/Inc/test.csv')
y_test = test_data[label]  # values to predict
test_data.head()

Loaded data from: https://autogluon.s3.amazonaws.com/datasets/Inc/test.csv | Columns = 15 / 15 | Rows = 9769 -> 9769

	age	workclass	fnlwgt	education	education-num	marital-status	occupation	relationship	race	sex	capital-gain	capital-loss	hours-per-week	native-country	class
0	31	Private	169085	11th	7	Married-civ-spouse	Sales	Wife	White	Female	0	0	20	United-States	<=50K
1	17	Self-emp-not-inc	226203	12th	8	Never-married	Sales	Own-child	White	Male	0	0	45	United-States	<=50K
2	47	Private	54260	Assoc-voc	11	Married-civ-spouse	Exec-managerial	Husband	White	Male	0	1887	60	United-States	>50K
3	21	Private	176262	Some-college	10	Never-married	Exec-managerial	Own-child	White	Female	0	0	30	United-States	<=50K
4	17	Private	241185	12th	8	Never-married	Prof-specialty	Own-child	White	Male	0	0	20	United-States	<=50K

We use our trained models to make predictions on the new data:

predictor = TabularPredictor.load(save_path)  # unnecessary, just demonstrates how to load previously-trained predictor from file

y_pred = predictor.predict(test_data)
y_pred

0        <=50K
1        <=50K
2         >50K
3        <=50K
4        <=50K
         ...  
9764     <=50K
9765     <=50K
9766     <=50K
9767     <=50K
9768     <=50K
Name: class, Length: 9769, dtype: object

We can use leaderboard to evaluate the performance of each individual trained model on our labeled test data:

predictor.leaderboard(test_data)

	model	score_test	score_val	eval_metric	pred_time_test	pred_time_val	fit_time	pred_time_test_marginal	pred_time_val_marginal	fit_time_marginal	stack_level	can_infer	fit_order
0	RandomForestGini	0.842870	0.84	accuracy	0.098668	0.046491	0.557529	0.098668	0.046491	0.557529	1	True	3
1	CatBoost	0.842461	0.85	accuracy	0.008353	0.003541	0.799393	0.008353	0.003541	0.799393	1	True	5
2	RandomForestEntr	0.841130	0.83	accuracy	0.098077	0.046339	0.523186	0.098077	0.046339	0.523186	1	True	4
3	XGBoost	0.839902	0.85	accuracy	0.035705	0.005714	0.376254	0.035705	0.005714	0.376254	1	True	9
4	LightGBM	0.839799	0.85	accuracy	0.015203	0.004305	0.218295	0.015203	0.004305	0.218295	1	True	2
5	WeightedEnsemble_L2	0.839799	0.85	accuracy	0.016673	0.005030	0.288410	0.001470	0.000726	0.070114	2	True	12
6	NeuralNetTorch	0.837138	0.83	accuracy	0.045333	0.010311	2.213367	0.045333	0.010311	2.213367	1	True	10
7	LightGBMXT	0.836421	0.83	accuracy	0.007769	0.004407	0.391984	0.007769	0.004407	0.391984	1	True	1
8	ExtraTreesGini	0.833862	0.82	accuracy	0.097154	0.050375	0.525635	0.097154	0.050375	0.525635	1	True	6
9	ExtraTreesEntr	0.833862	0.81	accuracy	0.099956	0.057210	0.530676	0.099956	0.057210	0.530676	1	True	7
10	NeuralNetFastAI	0.830791	0.83	accuracy	0.122786	0.008419	2.692888	0.122786	0.008419	2.692888	1	True	8
11	LightGBMLarge	0.817074	0.83	accuracy	0.011738	0.003245	0.661469	0.011738	0.003245	0.661469	1	True	11

Snapshot a Predictor with .clone()

Now that we have a working predictor artifact, we may want to alter it in a variety of ways to better suite our needs. For example, we may want to delete certain models to reduce disk usage via .delete_models(), or train additional models on top of the ones we already have via .fit_extra().

While you can do all of these operations on your predictor, you may want to be able to be able to revert to a prior state of the predictor in case something goes wrong. This is where predictor.clone() comes in.

predictor.clone() allows you to create a snapshot of the given predictor, cloning the artifacts of the predictor to a new location. You can then freely play around with the predictor and always load the earlier snapshot in case you want to undo your actions.

All you need to do to clone a predictor is specify a new directory path to clone to:

save_path_clone = save_path + '-clone'
# will return the path to the cloned predictor, identical to save_path_clone
path_clone = predictor.clone(path=save_path_clone)

Cloned TabularPredictor located in '/home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment' to 'agModels-predictClass-deployment-clone'.
	To load the cloned predictor: predictor_clone = TabularPredictor.load(path="agModels-predictClass-deployment-clone")

Note that this logic doubles disk usage, as it completely clones every predictor artifact on disk to make an exact replica.

Now we can load the cloned predictor:

predictor_clone = TabularPredictor.load(path=path_clone)
# You can alternatively load the cloned TabularPredictor at the time of cloning:
# predictor_clone = predictor.clone(path=save_path_clone, return_clone=True)

We can see that the cloned predictor has the same leaderboard and functionality as the original:

y_pred_clone = predictor.predict(test_data)
y_pred_clone

0        <=50K
1        <=50K
2         >50K
3        <=50K
4        <=50K
         ...  
9764     <=50K
9765     <=50K
9766     <=50K
9767     <=50K
9768     <=50K
Name: class, Length: 9769, dtype: object

y_pred.equals(y_pred_clone)

True

predictor_clone.leaderboard(test_data)

	model	score_test	score_val	eval_metric	pred_time_test	pred_time_val	fit_time	pred_time_test_marginal	pred_time_val_marginal	fit_time_marginal	stack_level	can_infer	fit_order
0	RandomForestGini	0.842870	0.84	accuracy	0.095832	0.046491	0.557529	0.095832	0.046491	0.557529	1	True	3
1	CatBoost	0.842461	0.85	accuracy	0.007397	0.003541	0.799393	0.007397	0.003541	0.799393	1	True	5
2	RandomForestEntr	0.841130	0.83	accuracy	0.096292	0.046339	0.523186	0.096292	0.046339	0.523186	1	True	4
3	XGBoost	0.839902	0.85	accuracy	0.034519	0.005714	0.376254	0.034519	0.005714	0.376254	1	True	9
4	LightGBM	0.839799	0.85	accuracy	0.014982	0.004305	0.218295	0.014982	0.004305	0.218295	1	True	2
5	WeightedEnsemble_L2	0.839799	0.85	accuracy	0.016423	0.005030	0.288410	0.001441	0.000726	0.070114	2	True	12
6	NeuralNetTorch	0.837138	0.83	accuracy	0.044763	0.010311	2.213367	0.044763	0.010311	2.213367	1	True	10
7	LightGBMXT	0.836421	0.83	accuracy	0.007699	0.004407	0.391984	0.007699	0.004407	0.391984	1	True	1
8	ExtraTreesGini	0.833862	0.82	accuracy	0.097291	0.050375	0.525635	0.097291	0.050375	0.525635	1	True	6
9	ExtraTreesEntr	0.833862	0.81	accuracy	0.098950	0.057210	0.530676	0.098950	0.057210	0.530676	1	True	7
10	NeuralNetFastAI	0.830791	0.83	accuracy	0.117042	0.008419	2.692888	0.117042	0.008419	2.692888	1	True	8
11	LightGBMLarge	0.817074	0.83	accuracy	0.011572	0.003245	0.661469	0.011572	0.003245	0.661469	1	True	11

Now let’s do some extra logic with the clone, such as calling refit_full:

predictor_clone.refit_full()

predictor_clone.leaderboard(test_data)

Refitting models via `predictor.refit_full` using all of the data (combined train and validation)...
	Models trained in this way will have the suffix "_FULL" and have NaN validation score.
	This process is not bound by time_limit, but should take less time than the original `predictor.fit` call.
	To learn more, refer to the `.refit_full` method docstring which explains how "_FULL" models differ from normal models.
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: LightGBMXT_FULL ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.4 GB
	0.21s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: LightGBM_FULL ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.4 GB
	0.23s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: RandomForestGini_FULL ...
	Fitting with cpus=8, gpus=0
	0.54s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: RandomForestEntr_FULL ...
	Fitting with cpus=8, gpus=0
	0.53s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: CatBoost_FULL ...
	Fitting with cpus=4, gpus=0
	0.02s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: ExtraTreesGini_FULL ...
	Fitting with cpus=8, gpus=0
	0.52s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: ExtraTreesEntr_FULL ...
	Fitting with cpus=8, gpus=0
	0.53s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: NeuralNetFastAI_FULL ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.3 GB
No improvement since epoch 0: early stopping
	0.32s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: XGBoost_FULL ...
	Fitting with cpus=4, gpus=0
	0.05s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: NeuralNetTorch_FULL ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.3 GB
	0.43s	 = Training   runtime
Fitting 1 L1 models, fit_strategy="sequential" ...
Fitting model: LightGBMLarge_FULL ...
	Fitting with cpus=4, gpus=0, mem=0.0/28.3 GB
	0.24s	 = Training   runtime
Fitting model: WeightedEnsemble_L2_FULL | Skipping fit via cloning parent ...
	Ensemble Weights: {'LightGBM': 1.0}
	0.07s	 = Training   runtime
Updated best model to "WeightedEnsemble_L2_FULL" (Previously "WeightedEnsemble_L2"). AutoGluon will default to using "WeightedEnsemble_L2_FULL" for predict() and predict_proba().
Refit complete, total runtime = 3.95s ... Best model: "WeightedEnsemble_L2_FULL"

	model	score_test	score_val	eval_metric	pred_time_test	pred_time_val	fit_time	pred_time_test_marginal	pred_time_val_marginal	fit_time_marginal	stack_level	can_infer	fit_order
0	CatBoost_FULL	0.842870	NaN	accuracy	0.006056	NaN	0.022880	0.006056	NaN	0.022880	1	True	17
1	RandomForestGini	0.842870	0.84	accuracy	0.096052	0.046491	0.557529	0.096052	0.046491	0.557529	1	True	3
2	CatBoost	0.842461	0.85	accuracy	0.007634	0.003541	0.799393	0.007634	0.003541	0.799393	1	True	5
3	RandomForestEntr	0.841130	0.83	accuracy	0.096811	0.046339	0.523186	0.096811	0.046339	0.523186	1	True	4
4	LightGBM_FULL	0.840823	NaN	accuracy	0.017952	NaN	0.227156	0.017952	NaN	0.227156	1	True	14
5	WeightedEnsemble_L2_FULL	0.840823	NaN	accuracy	0.019378	NaN	0.297271	0.001426	NaN	0.070114	2	True	24
6	XGBoost	0.839902	0.85	accuracy	0.034147	0.005714	0.376254	0.034147	0.005714	0.376254	1	True	9
7	LightGBM	0.839799	0.85	accuracy	0.015268	0.004305	0.218295	0.015268	0.004305	0.218295	1	True	2
8	WeightedEnsemble_L2	0.839799	0.85	accuracy	0.016708	0.005030	0.288410	0.001440	0.000726	0.070114	2	True	12
9	RandomForestGini_FULL	0.839390	NaN	accuracy	0.097366	NaN	0.538620	0.097366	NaN	0.538620	1	True	15
10	RandomForestEntr_FULL	0.839185	NaN	accuracy	0.097762	NaN	0.534288	0.097762	NaN	0.534288	1	True	16
11	LightGBMXT_FULL	0.837957	NaN	accuracy	0.008525	NaN	0.205105	0.008525	NaN	0.205105	1	True	13
12	NeuralNetTorch	0.837138	0.83	accuracy	0.044931	0.010311	2.213367	0.044931	0.010311	2.213367	1	True	10
13	LightGBMXT	0.836421	0.83	accuracy	0.008042	0.004407	0.391984	0.008042	0.004407	0.391984	1	True	1
14	XGBoost_FULL	0.836319	NaN	accuracy	0.031770	NaN	0.045425	0.031770	NaN	0.045425	1	True	21
15	ExtraTreesEntr_FULL	0.835705	NaN	accuracy	0.100702	NaN	0.531995	0.100702	NaN	0.531995	1	True	19
16	ExtraTreesGini	0.833862	0.82	accuracy	0.097739	0.050375	0.525635	0.097739	0.050375	0.525635	1	True	6
17	ExtraTreesEntr	0.833862	0.81	accuracy	0.098340	0.057210	0.530676	0.098340	0.057210	0.530676	1	True	7
18	ExtraTreesGini_FULL	0.833453	NaN	accuracy	0.098816	NaN	0.521801	0.098816	NaN	0.521801	1	True	18
19	NeuralNetFastAI	0.830791	0.83	accuracy	0.125398	0.008419	2.692888	0.125398	0.008419	2.692888	1	True	8
20	LightGBMLarge	0.817074	0.83	accuracy	0.011631	0.003245	0.661469	0.011631	0.003245	0.661469	1	True	11
21	NeuralNetTorch_FULL	0.815641	NaN	accuracy	0.046524	NaN	0.431732	0.046524	NaN	0.431732	1	True	22
22	LightGBMLarge_FULL	0.809704	NaN	accuracy	0.011571	NaN	0.241738	0.011571	NaN	0.241738	1	True	23
23	NeuralNetFastAI_FULL	0.768246	NaN	accuracy	0.116563	NaN	0.323847	0.116563	NaN	0.323847	1	True	20

We can see that we were able to fit additional models, but for whatever reason we may want to undo this operation.

Luckily, our original predictor is untouched!

predictor.leaderboard(test_data)

	model	score_test	score_val	eval_metric	pred_time_test	pred_time_val	fit_time	pred_time_test_marginal	pred_time_val_marginal	fit_time_marginal	stack_level	can_infer	fit_order
0	RandomForestGini	0.842870	0.84	accuracy	0.097736	0.046491	0.557529	0.097736	0.046491	0.557529	1	True	3
1	CatBoost	0.842461	0.85	accuracy	0.007975	0.003541	0.799393	0.007975	0.003541	0.799393	1	True	5
2	RandomForestEntr	0.841130	0.83	accuracy	0.096985	0.046339	0.523186	0.096985	0.046339	0.523186	1	True	4
3	XGBoost	0.839902	0.85	accuracy	0.037916	0.005714	0.376254	0.037916	0.005714	0.376254	1	True	9
4	LightGBM	0.839799	0.85	accuracy	0.015207	0.004305	0.218295	0.015207	0.004305	0.218295	1	True	2
5	WeightedEnsemble_L2	0.839799	0.85	accuracy	0.016694	0.005030	0.288410	0.001487	0.000726	0.070114	2	True	12
6	NeuralNetTorch	0.837138	0.83	accuracy	0.045044	0.010311	2.213367	0.045044	0.010311	2.213367	1	True	10
7	LightGBMXT	0.836421	0.83	accuracy	0.008049	0.004407	0.391984	0.008049	0.004407	0.391984	1	True	1
8	ExtraTreesEntr	0.833862	0.81	accuracy	0.098754	0.057210	0.530676	0.098754	0.057210	0.530676	1	True	7
9	ExtraTreesGini	0.833862	0.82	accuracy	0.099693	0.050375	0.525635	0.099693	0.050375	0.525635	1	True	6
10	NeuralNetFastAI	0.830791	0.83	accuracy	0.122739	0.008419	2.692888	0.122739	0.008419	2.692888	1	True	8
11	LightGBMLarge	0.817074	0.83	accuracy	0.012002	0.003245	0.661469	0.012002	0.003245	0.661469	1	True	11

We can simply clone a new predictor from our original, and we will no longer be impacted by the call to refit_full on the prior clone.

Snapshot a deployment optimized Predictor via .clone_for_deployment()

Instead of cloning an exact copy, we can instead clone a copy which has the minimal set of artifacts needed to do prediction.

Note that this optimized clone will have very limited functionality outside of calling predict and predict_proba. For example, it will be unable to train more models.

save_path_clone_opt = save_path + '-clone-opt'
# will return the path to the cloned predictor, identical to save_path_clone_opt
path_clone_opt = predictor.clone_for_deployment(path=save_path_clone_opt)

Cloned TabularPredictor located in '/home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment' to 'agModels-predictClass-deployment-clone-opt'.
	To load the cloned predictor: predictor_clone = TabularPredictor.load(path="agModels-predictClass-deployment-clone-opt")
Clone: Keeping minimum set of models required to predict with best model 'WeightedEnsemble_L2'...
Deleting model LightGBMXT. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/LightGBMXT will be removed.
Deleting model RandomForestGini. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/RandomForestGini will be removed.
Deleting model RandomForestEntr. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/RandomForestEntr will be removed.
Deleting model CatBoost. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/CatBoost will be removed.
Deleting model ExtraTreesGini. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/ExtraTreesGini will be removed.
Deleting model ExtraTreesEntr. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/ExtraTreesEntr will be removed.
Deleting model NeuralNetFastAI. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/NeuralNetFastAI will be removed.
Deleting model XGBoost. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/XGBoost will be removed.
Deleting model NeuralNetTorch. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/NeuralNetTorch will be removed.
Deleting model LightGBMLarge. All files under /home/ci/autogluon/docs/tutorials/tabular/advanced/agModels-predictClass-deployment-clone-opt/models/LightGBMLarge will be removed.
Clone: Removing artifacts unnecessary for prediction. NOTE: Clone can no longer fit new models, and most functionality except for predict and predict_proba will no longer work

predictor_clone_opt = TabularPredictor.load(path=path_clone_opt)

To avoid loading the model in every prediction call, we can persist the model in memory by:

predictor_clone_opt.persist()

Persisting 2 models in memory. Models will require 0.0% of memory.

['WeightedEnsemble_L2', 'LightGBM']

We can see that the optimized clone still makes the same predictions:

y_pred_clone_opt = predictor_clone_opt.predict(test_data)
y_pred_clone_opt

0        <=50K
1        <=50K
2         >50K
3        <=50K
4        <=50K
         ...  
9764     <=50K
9765     <=50K
9766     <=50K
9767     <=50K
9768     <=50K
Name: class, Length: 9769, dtype: object

y_pred.equals(y_pred_clone_opt)

True

predictor_clone_opt.leaderboard(test_data)

	model	score_test	score_val	eval_metric	pred_time_test	pred_time_val	fit_time	pred_time_test_marginal	pred_time_val_marginal	fit_time_marginal	stack_level	can_infer	fit_order
0	LightGBM	0.839799	0.85	accuracy	0.014848	0.004305	0.218295	0.014848	0.004305	0.218295	1	True	1
1	WeightedEnsemble_L2	0.839799	0.85	accuracy	0.015635	0.005030	0.288410	0.000787	0.000726	0.070114	2	True	2

We can check the disk usage of the optimized clone compared to the original:

size_original = predictor.disk_usage()
size_opt = predictor_clone_opt.disk_usage()
print(f'Size Original:  {size_original} bytes')
print(f'Size Optimized: {size_opt} bytes')
print(f'Optimized predictor achieved a {round((1 - (size_opt/size_original)) * 100, 1)}% reduction in disk usage.')

Size Original:  18351007 bytes
Size Optimized: 182055 bytes
Optimized predictor achieved a 99.0% reduction in disk usage.

We can also investigate the difference in the files that exist in the original and optimized predictor.

Original:

predictor.disk_usage_per_file()

/models/ExtraTreesGini/model.pkl                        5065911
/models/ExtraTreesEntr/model.pkl                        5024141
/models/RandomForestGini/model.pkl                      3408886
/models/RandomForestEntr/model.pkl                      3267285
/models/XGBoost/xgb.ubj                                  506961
/models/LightGBMLarge/model.pkl                          310869
/models/NeuralNetTorch/model.pkl                         253993
/models/NeuralNetFastAI/model-internals.pkl              170508
/models/LightGBM/model.pkl                               147792
/models/CatBoost/model.pkl                                52238
/models/LightGBMXT/model.pkl                              43138
/utils/data/X.pkl                                         27584
/learner.pkl                                              10352
/models/WeightedEnsemble_L2/model.pkl                     10340
/metadata.json                                             9707
/utils/data/X_val.pkl                                      8350
/utils/data/y.pkl                                          7462
/models/XGBoost/model.pkl                                  6180
/models/trainer.pkl                                        5229
/models/NeuralNetFastAI/model.pkl                          2714
/utils/data/y_val.pkl                                      2355
/models/WeightedEnsemble_L2/utils/model_template.pkl       1269
/predictor.pkl                                              903
/models/WeightedEnsemble_L2/utils/oof.pkl                   765
/utils/attr/LightGBM/y_pred_proba_val.pkl                   551
/utils/attr/LightGBMLarge/y_pred_proba_val.pkl              551
/utils/attr/XGBoost/y_pred_proba_val.pkl                    551
/utils/attr/NeuralNetTorch/y_pred_proba_val.pkl             551
/utils/attr/ExtraTreesGini/y_pred_proba_val.pkl             551
/utils/attr/RandomForestGini/y_pred_proba_val.pkl           551
/utils/attr/LightGBMXT/y_pred_proba_val.pkl                 551
/utils/attr/NeuralNetFastAI/y_pred_proba_val.pkl            551
/utils/attr/CatBoost/y_pred_proba_val.pkl                   551
/utils/attr/ExtraTreesEntr/y_pred_proba_val.pkl             551
/utils/attr/RandomForestEntr/y_pred_proba_val.pkl           551
/version.txt                                                 14
Name: size, dtype: int64

Optimized:

predictor_clone_opt.disk_usage_per_file()

/models/LightGBM/model.pkl               147820
/models/WeightedEnsemble_L2/model.pkl     10390
/learner.pkl                              10352
/metadata.json                             9707
/models/trainer.pkl                        2869
/predictor.pkl                              903
/version.txt                                 14
Name: size, dtype: int64

Compile models for maximized inference speed

In order to further improve inference efficiency, we can call .compile() to automatically convert sklearn function calls into their ONNX equivalents. Note that this is currently an experimental feature, which only improves RandomForest and TabularNeuralNetwork models. The compilation and inference speed acceleration require installation of skl2onnx and onnxruntime packages. To install supported versions of these packages automatically, we can call pip install autogluon.tabular[skl2onnx] on top of an existing AutoGluon installation, or pip install autogluon.tabular[all,skl2onnx] on a new AutoGluon installation.

It is important to make sure the predictor is cloned, because once the models are compiled, it won’t support fitting.

predictor_clone_opt.compile()

Compiling 2 Models ...
Skipping compilation for WeightedEnsemble_L2 ... (No config specified)
Skipping compilation for LightGBM ... (No config specified)
Finished compiling models, total runtime = 0s.

With the compiled predictor, the prediction results might not be exactly the same but should be very close.

y_pred_compile_opt = predictor_clone_opt.predict(test_data)
y_pred_compile_opt

0        <=50K
1        <=50K
2         >50K
3        <=50K
4        <=50K
         ...  
9764     <=50K
9765     <=50K
9766     <=50K
9767     <=50K
9768     <=50K
Name: class, Length: 9769, dtype: object

Now all that is left is to upload the optimized predictor to a centralized storage location such as S3. To use this predictor in a new machine / system, simply download the artifact to local disk and load the predictor. Ensure that when loading a predictor you use the same Python version and AutoGluon version used during training to avoid instability.