Zoom Data pre-processing

Last updated: January 4, 2023

 Table of contents

Data often needs to be pre-processed before it can be fed to a model. We already saw an example with the MNIST data: we used transform to pre-process the features.

The TorchVision datasets also have the argument target_transform to modify the labels.

Both transform and target_transform accept callables.

Here is an example with the FashionMNIST datasets.

Load packages 

import torch
from torchvision import datasets
from torchvision.transforms import ToTensor, Lambda

Pre-process the FashionMNIST data

The FashionMNIST contains pairs of images and labels.

The features (images) need to be transformed into normalized tensors. For this we use ToTensor which turn a PIL image (the case here) a NumPy’s ndarray into FloatTensor and scales the image’s pixel intensity values in the range [0., 1.].

The labels are integers. Representing categorical variables as integers can create problem in training models, so we transform them into one-hot encoded tensors. For this, we use Lambda to apply a custom function which creates a zero tensor of size 10 (the number of labels) and calls scatter_ which assigns a value=1 on the index as given by the label y.

dataset = datasets.FashionMNIST(
    root="/project/def-sponsor00/data/",
    train=True,
    download=True,
    transform=ToTensor(),
    target_transform=Lambda(lambda y: torch.zeros(10, dtype=torch.float).scatter_(0, torch.tensor(y), value=1))
)

Comments & questions