mnist_handler.py

from torchvision import transforms
from ts.torch_handler.image_classifier import ImageClassifier
import torch


class MNISTDigitClassifier(ImageClassifier):
    """
    MNISTDigitClassifier handler class. This handler extends class ImageClassifier from image_classifier.py, a
    default handler. This handler takes an image and returns the number in that image.

    Here method postprocess() has been overridden while others are reused from parent class.
    """

    image_processing = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.1307,), (0.3081,))
    ])

    def postprocess(self, data):
        """The post process of MNIST converts the predicted output response to a label.

        Args:
            data (list): The predicted output from the Inference with probabilities is passed
            to the post-process function
        Returns:
            list : A list of dictionaries with predictions and explanations is returned
        """
        return data.argmax(1).tolist()