Arbok
Arbok (Automl wrapper toolbox for openml compatibility) provides wrappers for TPOT and Auto-Sklearn, as a compatibility layer between these tools and OpenML.
The wrapper extends Sklearn's BaseSearchCV and provides all the internal parameters that OpenML needs, such as
cv_results_, best_index_, best_params_, best_score_ and classes_.
Installation
pip install arbok
Simple example
import openml
from arbok import AutoSklearnWrapper, TPOTWrapper
task = openml.tasks.get_task(31)
dataset = task.get_dataset()
# Get the AutoSklearn wrapper and pass parameters like you would to AutoSklearn
clf = AutoSklearnWrapper(
time_left_for_this_task=3600, per_run_time_limit=360
)
# Or get the TPOT wrapper and pass parameters like you would to TPOT
clf = TPOTWrapper(
generations=100, population_size=100, verbosity=2
)
# Execute the task
run = openml.runs.run_model_on_task(task, clf)
run.publish()
print('URL for run: %s/run/%d' % (openml.config.server, run.run_id))Preprocessing data
To make the wrapper more robust, we need to preprocess the data. We can fill the missing values, and one-hot encode categorical data.
First, we get a mask that tells us whether a feature is a categorical feature or not.
dataset = task.get_dataset()
_, categorical = dataset.get_data(return_categorical_indicator=True)
categorical = categorical[:-1] # Remove last index (which is the class)Next, we setup a pipeline for the preprocessing. We are using a ConditionalImputer,
which is an imputer which is able to use different strategies for categorical (nominal) and numerical data.
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import OneHotEncoder
from arbok import ConditionalImputer
preprocessor = make_pipeline(
ConditionalImputer(
categorical_features=categorical,
strategy="mean",
strategy_nominal="most_frequent"
),
OneHotEncoder(
categorical_features=categorical, handle_unknown="ignore", sparse=False
)
)And finally, we put everything together in one of the wrappers.
clf = AutoSklearnWrapper(
preprocessor=preprocessor, time_left_for_this_task=3600, per_run_time_limit=360
)Limitations
- Currently only the classifiers are implemented. Regression is therefore not possible.
- For TPOT, the
config_dictvariable can not be set, because this causes problems with the API.
Benchmarking
Installing the arbok package includes the arbench cli tool.
We can generate a json file like this:
from arbok.bench import Benchmark
bench = Benchmark()
config_file = bench.create_config_file(
# Wrapper parameters
wrapper={"refit": True, "verbose": False, "retry_on_error": True},
# TPOT parameters
tpot={
"max_time_mins": 6, # Max total time in minutes
"max_eval_time_mins": 1 # Max time per candidate in minutes
},
# Autosklearn parameters
autosklearn={
"time_left_for_this_task": 360, # Max total time in seconds
"per_run_time_limit": 60 # Max time per candidate in seconds
}
)And then, we can call arbench like this:
arbench --classifier autosklearn --task-id 31 --config config.jsonOr calling arbok as a python module:
python -m arbok --classifier autosklearn --task-id 31 --config config.jsonRunning a benchmark on batch systems
To run a large scale benchmark, we can create a configuration file like above, and generate and submit jobs to a batch system as follows.
# We create a benchmark setup where we specify the headers, the interpreter we
# want to use, the directory to where we store the jobs (.sh-files), and we give
# it the config-file we created earlier.
bench = Benchmark(
headers="#PBS -lnodes=1:cpu3\n#PBS -lwalltime=1:30:00",
python_interpreter="python3", # Path to interpreter
root="/path/to/project/",
jobs_dir="jobs",
config_file="config.json",
log_file="log.json"
)
# Create the config file like we did in the section above
config_file = bench.create_config_file(
# Wrapper parameters
wrapper={"refit": True, "verbose": False, "retry_on_error": True},
# TPOT parameters
tpot={
"max_time_mins": 6, # Max total time in minutes
"max_eval_time_mins": 1 # Max time per candidate in minutes
},
# Autosklearn parameters
autosklearn={
"time_left_for_this_task": 360, # Max total time in seconds
"per_run_time_limit": 60 # Max time per candidate in seconds
}
)
# Next, we load the tasks we want to benchmark on from OpenML.
# In this case, we load a list of task id's from study 99.
tasks = openml.study.get_study(99).tasks
# Next, we create jobs for both tpot and autosklearn.
bench.create_jobs(tasks, classifiers=["tpot", "autosklearn"])
# And finally, we submit the jobs using qsub
bench.submit_jobs()Preprocessing parameters
from arbok import ParamPreprocessor
import numpy as np
from sklearn.feature_selection import VarianceThreshold
from sklearn.pipeline import make_pipeline
X = np.array([
[1, 2, True, "foo", "one"],
[1, 3, False, "bar", "two"],
[np.nan, "bar", None, None, "three"],
[1, 7, 0, "zip", "four"],
[1, 9, 1, "foo", "five"],
[1, 10, 0.1, "zip", "six"]
], dtype=object)
# Manually specify types, or use types="detect" to automatically detect types
types = ["numeric", "mixed", "bool", "nominal", "nominal"]
pipeline = make_pipeline(ParamPreprocessor(types="detect"), VarianceThreshold())
pipeline.fit_transform(X)Output:
[[-0.4472136 -0.4472136 1.41421356 -0.70710678 -0.4472136 -0.4472136
2.23606798 -0.4472136 -0.4472136 -0.4472136 0.4472136 -0.4472136
-0.85226648 1. ]
[-0.4472136 2.23606798 -0.70710678 -0.70710678 -0.4472136 -0.4472136
-0.4472136 -0.4472136 -0.4472136 2.23606798 0.4472136 -0.4472136
-0.5831297 -1. ]
[ 2.23606798 -0.4472136 -0.70710678 -0.70710678 -0.4472136 -0.4472136
-0.4472136 -0.4472136 2.23606798 -0.4472136 -2.23606798 2.23606798
-1.39054004 -1. ]
[-0.4472136 -0.4472136 -0.70710678 1.41421356 -0.4472136 2.23606798
-0.4472136 -0.4472136 -0.4472136 -0.4472136 0.4472136 -0.4472136
0.49341743 -1. ]
[-0.4472136 -0.4472136 1.41421356 -0.70710678 2.23606798 -0.4472136
-0.4472136 -0.4472136 -0.4472136 -0.4472136 0.4472136 -0.4472136
1.031691 1. ]
[-0.4472136 -0.4472136 -0.70710678 1.41421356 -0.4472136 -0.4472136
-0.4472136 2.23606798 -0.4472136 -0.4472136 0.4472136 -0.4472136
1.30082778 1. ]]
