returning scikit-learn object while using Joblib

user_9916514 0
Viewed 95

I have a numpy array, and I am using sklearn to transform the array along the first axis. I also want to save the transformer object in a dict to use later in the code. Here is my code:

scalers_dict = {}
for i in range(train_data_numpy.shape[1]):
    for j in range(train_data_numpy.shape[2]):
        scaler = QuantileTransformer(n_quantiles=60000, output_distribution='uniform')
        train_data_numpy[:,i,j] = scaler.fit_transform(train_data_numpy[:,i,j].reshape(-1,1)).reshape(-1)
        scalers_dict[(i,j)] = scaler

My train_data_numpy is of shape (60000, 28,28). The problem is that this takes a very long time to process (train_data_numpy is MNIST dataset). I have an AMD Ryzen 5950X with 16 cores and I would like to parallelize this piece of code.

I know for example I could write something like this:

Parallel(n_jobs=16)(delayed(QuantileTransformer(n_quantiles=60000, output_distribution='uniform').fit_transform)(train_data_numpy[:,i,j].reshape(-1,1)) for j in range(train_data_numpy.shape[2]))

But this doesn't return the scaler object, and I don't know how to utilize Joblib for this task.

joblib scikit-learn numpy-ndarray parallel-processing python
1 Answers

You can use Dask-ML which is implemented on the top of Dask Library, yet it is compatible with scikit-learn.

Installation:

conda install -c conda-forge dask-ml

or

pip install dask-ml

Example

import time
from sklearn.datasets import make_classification
from sklearn.preprocessing import QuantileTransformer as skQT
from dask_ml.preprocessing import QuantileTransformer as daskQT

# toy big dataset for testing
X, y = make_classification(n_samples=1000000, n_features=100, random_state=2021)

# Comparison

scaler = skQT()
start_ = time.time()
scaler.fit_transform(X)
end_ = time.time() - start_
print("No Parallelism -- Time Elapsed: {}".format(end_))


# Using Dask ML
scaler = daskQT()
start_ = time.time()
scaler.fit_transform(X)
end_ = time.time() - start_
print("With Parallelism -- Time Elapsed: {}".format(end_))

Result

No Parallelism -- Time Elapsed: 18.680
With Parallelism -- Time Elapsed: 2.982

My Device Specs:

Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz

Number of Cores: 12
user_63651120
Related