As a final test for this chapter, let's visualize some data from an external dataset, such as the digits dataset from scikit-learn.
Specifically, we will need three tools for visualization:
- scikit-learn for the actual data
- NumPy for data processing
- Matplotlib
So, let's start by importing all of these:
In [1]: import numpy as np
... from sklearn import datasets
... import matplotlib.pyplot as plt
... %matplotlib inline
The first step is to actually load the data:
In [2]: digits = datasets.load_digits()
If we remember correctly, digits is supposed to have two different fields: a data field containing the actual image data and a target field containing the image labels. Rather than trusting our memory, we should simply investigate the digits object. We do this by typing out its name, adding a period, and then hitting the Tab key: digits.<TAB>. This will reveal that the digits object also contains some other fields, such as one called images. The two fields, images and data, seem to simply differ by shape:
In [3]: print(digits.data.shape)
... print(digits.images.shape)
Out[3]: (1797, 64)
(1797, 8, 8)
In both cases, the first dimension corresponds to the number of images in the dataset. However, data has all of the pixels lined up in one big vector, whereas images preserves the 8 x 8 spatial arrangement of each image.
Hence, if we wanted to plot a single image, the images field would be more appropriate. First, we grab a single image from the dataset using NumPy's array slicing:
In [4]: img = digits.images[0, :, :]
Here, we are saying that we want to grab the first row in the 1,797-item-long array and all of the corresponding 8 x 8 = 64 pixels. We can then plot the image using the imshow function of plt:
In [5]: plt.imshow(img, cmap='gray')
... plt.savefig('figures/02.04-digit0.png')
Out[5]: <matplotlib.image.AxesImage at 0x7efcd27f30f0>
The preceding command gives the following output. Note that the image is blurred because we have resized it to a larger size. The original image's size is just 8 x 8:
In addition, I also specified a colormap with the cmap argument. By default, Matplotlib uses MATLAB's default colormap jet. However, in the case of grayscale images, the gray colormap makes more sense.
Finally, we can plot a whole number of digit samples using the subplot function of plt. The subplot function is the same as in MATLAB, where we specify the number of rows, number of columns, and current subplot index (starts counting at 1). We will use a for loop to iterate over the first 10 images in the dataset and every image gets assigned its own subplot:
In [6]: plt.figure(figsize=(14,4))
...
... for image_index in range(10):
... # images are 0-indexed, but subplots are 1-indexed
... subplot_index = image_index + 1
... plt.subplot(2, 5, subplot_index)
... plt.imshow(digits.images[image_index, :, :], cmap='gray')
This leads to the following output:
