Scaling an image classification model on ImageNet

In Chapter 5, Training Computer Vision Models, we trained the image classification algorithm on a small dataset with dog and cat images (25,000 training images). This time, let's go for something a little bigger.

We're going to train a ResNet-50 network from scratch on the ImageNet dataset – the reference dataset for many computer vision applications (http://www.image-net.org). The 2012 version contains 1,281,167 training images (140 GB) and 50,000 validation images (6.4 GB) from 1,000 classes.

If you want to experiment at a smaller scale, you can work with 5-10% of the dataset. Final accuracy won't be as good, but it doesn't matter for our purposes.

Preparing the ImageNet dataset

This requires a lot of storage – the dataset is 150 GB, so please make sure you have at least 500 GB available to store it in ZIP and processed formats. You're also going to need a lot of bandwidth and a...