One common example for four-dimensional tensor types is a batch of images. Modern CPUs and GPUs are optimized to perform the same operations on multiple examples faster. So, they take a similar time to process one image or a batch of images. So, it is common to use a batch of examples rather than use a single image at a time. Choosing the batch size is not straightforward; it depends on several factors. One major restriction for using a bigger batch or the complete dataset is GPU memory limitations—16, 32, and 64 are commonly used batch sizes.

Let's look at an example where we load a batch of cat images of size 64 x 224 x 224 x 3 where 64 represents the batch size or the number of images, 244 represents height and width, and 3 represents channels:

#Read cat images from disk
cats = glob(data_path+'*.jpg')
#Convert images into numpy arrays
cat_imgs = np.array([np.array(Image.open(cat).resize((224,224))) for cat in cats[:64]]) 
cat_imgs = cat_imgs.reshape(-1,224,224,3)
cat_tensors = torch.from_numpy(cat_imgs)
cat_tensors.size()

Output - torch.Size([64, 224, 224, 3])