A Beginners Guide to Skorch – With Code To Implement Neural Network

Skorch is one of the useful libraries in Pytorch to work on machine learning models especially neural networks. It is a robust library that works in the combination of PyTorch and Scikit-learn. Previously we used the scikit-learn library but its applications are only limited to classical machine learning models while skorch is a compatible library for developing the neural networks.

Topics we cover in the article:

• What is skorch?
• How to install and use skorch?
• Basic Functions we use in skorch
• Neural Network using Skorch
• Hands-on Binary classification model using skorch 

What is SKORCH…? Basic Functions we use in Skorch

Skorch enables programmers to implement code using the customizability scikit-learn and power of PyTorch. Just call model.fit() and you don’t have to worry about writing your own callback functions, skorch handles everything for you. Skorch is an inbuilt library in PyTorch used for Neural networks. In building deep neural networks, we are required to train our model, evaluate, dividing the train & validation data, and many other operations required to build a model. Here skorch library will reduce the effort of the programmer similar to scikit learn library in machine learning. Skorch also provides many features like learning rate schedulers, early stopping, checkpointing, parameter freezing and unfreezing, etc.

So What are callback functions………?

A callback function is passed into another function as an argument and then it is invoked into the outer function and it is used to perform routine actions.

How to Install and Use  Skorch

#Installing skorch
pip install -U skorch

#Using Skorch
import skorch

Basic Functions we use in Skorch 

from skorch import NeuralNet
from skorch.callbacks import EpochTimer
from skorch.callbacks import PrintLog
from skorch.callbacks import EpochScoring
from skorch.callbacks import PassthroughScoring
from skorch.dataset import CVSplit
from skorch.utils import get_dim
from skorch.utils import is_dataset
from skorch.utils import to_numpy

Binary classification model using skorch learn

Here we are importing the required libraries.

import skorch
import numpy as np
from sklearn.datasets import make_classification
from torch import nn
from skorch import NeuralNetClassifier

In the below code snippet we are implementing the basic classification toy model using skorch.

X, y = make_classification(1000, 20, n_informative=10, random_state=0)
X = X.astype(np.float32)
y = y.astype(np.int64)

class MyModule(nn.Module):
    def __init__(self, num_units=10, nonlin=nn.ReLU()):
        super(MyModule, self).__init__()
        self.dense0 = nn.Linear(20, num_units)
        self.nonlin = nonlin
        self.dropout = nn.Dropout(0.5)
        self.dense1 = nn.Linear(num_units, num_units)
        self.output = nn.Linear(num_units, 2)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, X, **kwargs):
        X = self.nonlin(self.dense0(X))
        X = self.dropout(X)
        X = self.nonlin(self.dense1(X))
        X = self.softmax(self.output(X))
        return X

model = NeuralNetClassifier(
    MyModule,
    max_epochs=10,
    lr=0.1,

    # Shuffle training data on each epoch
    iterator_train__shuffle=True,
)

model.fit(X, y)

y_proba = net.predict_proba(X)

#Output

Conclusion

In the above demonstration, we have seen that how the work for programmers can be reduced by implementing only a few lines of code by deploying the skorch library. This library is specially designed for neural network developers to reduce the programming efforts. We can be able to build the neural net model so easily by using a model.fit(), specifying the number of epochs, declaring a few parameters, and so we get a model built.