VI-ME-BA-BAR on “real” data

Today we will be analyzing a dataset made of images of apples and bananas, with the aim to automatically classify them into images of apples on the one hand and images of bananas on the other hand.

Before proceeding you need to learn some best practices for executable documents that we will follow. When done, execute the next cell and proceed to the following section.


# Display which version of Python you are running
from platform import python_version

# Load general libraries
import os, re
from glob import glob as ls
import numpy as np                    # Matrix algebra library
import pandas as pd                   # Data table (DataFrame) library
import seaborn as sns; sns.set()      # Graphs and visualization library
from PIL import Image                 # Image processing library
import matplotlib.pyplot as plt       # Library to make graphs 
# Command to insert the graphs in line in the notebook:
%matplotlib inline

# Reload code when changes are made
%load_ext autoreload
%autoreload 2

# Import utilities
from utilities import *

# The dataset for today
from intro_science_donnees import data_dir
dataset_dir = os.path.join(data_dir, 'apples_and_bananas_simple')

Step 1: preprocessing and [VI]sualizing

Welcome back! Load the images from our dataset in the variable images, and show them here:

images = load_images(dataset_dir, "*.png")
image_grid(images, titles=images.index)

Look at the images in details. Can you always tell apples and bananas apart? Try to imagine how your brain achieves this!


Data often come already in a feature representation: namely each element in the dataset is described by a collection of features – measurable properties or characteristic of that element; for an animal, it could be its size, its body temperature, etc.

This is also the case in our dataset: an image is described by the color of each of its pixel. However these features are too numerous and too low level for our purposes. We want instead just a few high level features measuring properties about the fruit in the picture, like its average color or shape.

Our first step is therefore to preprocess the data to extract such high level features. To achieve this, you need first to learn some basics of image processing. Then proceed with the feature extraction


Let’s summarize the data about our images in a pandas DataFrame with one column for each feature, and a last column holding the ground truth value: 1 if the picture shows an apple, -1 if it’s a banana.

This dataframe is all we will need in the following. For your convenience its content is also provided together with the dataset. If you have not yet finished the feature extraction and nevertheless wish to proceed with the data analysis, you may instead uncomment and execute the next cell to read the data.

df = pd.DataFrame({'redness':    images.apply(redness),
                   'elongation': images.apply(elongation),
                   'fruit': name: 1 if name[0] == 'a' else -1),
# df = pd.read_csv(os.path.join(dataset_dir, "preprocessed_data.csv"), index_col=0)

You may remember that Pandas dataframes are really convenient to show heat maps:'RdYlGn_r')

Note how the colormap is applied on the range of values in each column independently. Therefore, we can visualize the variation in each column even when column values in different units or in different ranges of values, as is the case here.

Next we want to compute correlations between columns. We have seen in the lecture that the correlation becomes meaningless when columns are of different magnitudes. So we need first to standardize each column to make it have mean \(0\) and standard deviation \(1\).


Standardizing the dataframe – that is, for each column, subtracting the mean and dividing by the standard deviation – is as easy as:

dfstd =  (df - df.mean()) / df.std()

Admittedly, it takes a bit of head scratching to figure out why this actually works thanks to columnwise operations :-)

Now all columns are indeed standardized:


This is not quite what we want however: the fruit value is not really numerical; we want to keep the original values -1 and 1.

dfstd['fruit'] = df['fruit']


Let’s look at the heatmap of the standardized data frame, which is similar to the original one:'RdYlGn_r') # Panda's native heatmap

We see that redness is correlated with fruit type: unsurprisingly, apples tend to be red (with exceptions) and bananas green (with exceptions). Meanwhile elongation is anti-correlated with fruit type: round apples and long bananas. This is confirmed by the correlation matrix:

fig = Figure(figsize=(7,7))
sns.heatmap(dfstd.corr(), fmt='0.2f', annot=True, square=True, cmap='RdYlGn_r', vmin=-1, vmax=1, ax=fig.add_subplot())

We can spot on the heatmap some outliers (e.g. two red bananas). This is confirmed by looking at the scatter plot:

make_scatter_plot(dfstd, images.apply(transparent_background_filter), axis='square')

We can also visualize the dataset with pair plots:

sns.pairplot(dfstd, hue="fruit", diag_kind="hist", palette='Set2');

Notice that a single feature (redness alone or elongation alone) is almost sufficient to perfectly separate apples and bananas: the problem is easy!

Step 2: [ME]asuring performance

Now that we have well understood and prepared our data, we want to determine how to evaluate performance for the task at hand: separating apples from bananas (a classification problem).

Splitting the data into a training set and a test set

Let’s separate the information in our data table according to its nature:

  • X will hold the features from which to make predictions

  • Y will hold the ground truth: what we try to predict: is it actually an apple or a banana?

X = dfstd[['redness', 'elongation']]
Y = dfstd['fruit']

Why the notations \(X\) and \(Y\)? Because we are on a quest for predictive models \(f\); these try, for each index \(i\), to predict the ground truth \(Y_i\) from the features in \(X_i\): ideally, \(Y_i=f(X_i)\), for all \(i\).

Now, we want to split our images into two subsets, a training set (that we will use to adjust the parameters of our predictive models) and a test set (to compute prediction performance without being overly optimistic).

# Make one training-test split in a stratified manner, i.e. same number of apples and bananas in each set.
train_index, test_index = split_data(X, Y, verbose = True, seed=0)
Xtrain, Xtest = X.iloc[train_index], X.iloc[test_index]
Ytrain, Ytest = Y.iloc[train_index], Y.iloc[test_index]

These images will serve for training the predictive models:

image_grid(images.iloc[train_index], titles=train_index)

And these for testing the predictive models:

image_grid(images.iloc[test_index], titles=test_index)

Note that both the training set and the testing set contain the same proportion of apples and bananas as in the original data set. This is guaranteed, and on purpose!

We plot the training and test data as scatter plots. The test data are shown with question marks, because their class identities (apple or banana) are hidden:

make_scatter_plot(dfstd, images, train_index, test_index, filter=transparent_background_filter, axis='square')

Exercise: error rate computation

The error rate is a performance metric for predictions. It is defined as the fraction \(\frac e n\), where \(e\) is the number of incorrect predictions and \(n\) the total number of predictions.

Write a function that computes the error rate, taking as input:

  • A vector solution containing the target values (1 for apples and -1 for bananas).

  • A vector prediction containing the predicted value

def error_rate(solution, prediction):
    '''Compute the error rate between two vectors.'''
    return np.mean( solution != prediction )

Then write unit tests (with assert) that check that:

  • the error rate between solution=Ytrain and prediction=Ytrain is zero (why?)

  • the error rate between solution=Ytrain and prediction=[1,...,1] is 0.5 (why?)

  • the error rate between solution=Ytrain and prediction=[0,...,0] is one (why?)

Hint: you may use np.zeros(Ytrain.shape) to generate an array [0,...,0] of same size as Ytrain, and similarly for [1,...,1].

assert error_rate(Ytrain, Ytrain) == 0
assert error_rate(Ytrain, np.zeros(Ytrain.shape)) == 1
assert error_rate(Ytrain, np.ones(Ytrain.shape)) == 0.5

The Machine Learning library scikit-learn also called sklearn has a function accuracy_score. As an additional verification, we test that error_rate + accuracy_score = 1 using the same examples as above.

from sklearn.metrics import accuracy_score
assert abs(error_rate(Ytrain, Ytrain)                 + accuracy_score(Ytrain, Ytrain)                 - 1) <= .1
assert abs(error_rate(Ytrain, np.zeros(Ytrain.shape)) + accuracy_score(Ytrain, np.zeros(Ytrain.shape)) - 1) <= .1
assert abs(error_rate(Ytrain, np.ones(Ytrain.shape))  + accuracy_score(Ytrain, np.ones (Ytrain.shape)) - 1) <= .1

Step 3: [BA]seline results

Exercise: 1-nearest-neighbor classifier

In a k-nearest neighbor algorithm (KNN), an unlabeled input is classified regarding the proximity of its labeled neighbors. The input will be predicted as belonging to the class C if the majority of the k nearest neighbors belongs to class C. Here, we take k=1 so we only consider the closest labeled image to classify an unlabeled image.

The 1-nearest neighbor classifier is a nice and simple method. It is implemented in scikit-learn along with many others. You may also want to implement it yourself later in the semester at the occasion of your project.

Import the KNeighborsClassifier classifier from sklearn.neighbors. Use it to construct a new model, setting the number of neighbors to one. Train this model with Xtrain by calling the method fit. Then use the trained model to create two vectors of prediction Ytrain_predicted and Ytest_predicted by calling the method predict. Compute e_tr, the training error rate, and e_te the test error rate.
Hint: look up the documentation as often as needed, starting with KNeighborsClassifier?

from sklearn.neighbors import KNeighborsClassifier
neigh = KNeighborsClassifier(n_neighbors=1), Ytrain) 
Ytrain_predicted = neigh.predict(Xtrain)
Ytest_predicted = neigh.predict(Xtest)
e_tr = error_rate(Ytrain, Ytrain_predicted)
e_te = error_rate(Ytest, Ytest_predicted)
print("Training error:", e_tr)
print("Test error:", e_te)

This problem is too easy! We get zero error on training data and one error on test data!

Here we overlay the predictions on test examples on the scatter plot …

# The training examples are shown as white circles and the test examples are black squares.
# The predictions made are shown as letters in the black squares.
make_scatter_plot(X, images.apply(transparent_background_filter),
                  train_index, test_index, 
                  predicted_labels=Ytest_predicted, axis='square')

… then, we show the “ground truth” and compute the error rate

# The training examples are shown as white circles and the test examples are blue squares.
make_scatter_plot(X, images.apply(transparent_background_filter),
                  train_index, test_index, 
                  predicted_labels='GroundTruth', axis='square')

Step 4: [BAR]s of error and test set size

Last but not least, let us evaluate the significance of our results by computing error bars. Obviously, since we have only 10 test examples, we cannot see at least 100 errors (which is the target we gave to ourselves in class). But this is only a toy example.

Exercise: Test set standard error

Compute the 1-sigma error bar of the test error rate e_te using the standard error formula defined in class, and assign it to sigma. How many test examples would we need to divide this error bar by a factor of two?

n_te = len(Ytest)
sigma = np.sqrt(e_te * (1-e_te) / n_te)
print("TEST SET ERROR RATE: {0:.2f}".format(e_te))
print("TEST SET STANDARD ERROR: {0:.2f}".format(sigma))
assert abs( sigma - 0.13 ) < 0.1

Cross-validation (CV) error bar

Another way of computing an error bar is to repeat multiple times the train/test split and compute the mean and standard deviation of the test error. In some sense this is more informative because it involves both the variability of the training set and that of the test set. But is is known to be a biased estimator of the error variability.

n_te = 10
SSS = StratifiedShuffleSplit(n_splits=n_te, test_size=0.5, random_state=5)
E = np.zeros([n_te, 1])
k = 0
for train_index, test_index in SSS.split(X, Y):
    print("TRAIN:", train_index, "TEST:", test_index)
    Xtrain, Xtest = X.iloc[train_index], X.iloc[test_index]
    Ytrain, Ytest = Y.iloc[train_index], Y.iloc[test_index], Ytrain.ravel()) 
    Ytrain_predicted = neigh.predict(Xtrain)
    Ytest_predicted = neigh.predict(Xtest)
    e_te = error_rate(Ytest, Ytest_predicted)
    print("TEST ERROR RATE:", e_te)
    E[k] = e_te
    k = k+1
e_te_ave = np.mean(E)
# It is bad practice to show too many decimal digits:
print("\n\nCV ERROR RATE: {0:.2f}".format(e_te_ave))
print("CV STANDARD DEVIATION: {0:.2f}".format(np.std(E)))

sigma = np.sqrt(e_te_ave * (1-e_te_ave) / n_te)
print("TEST SET STANDARD ERROR (for comparison): {0:.2f}".format(sigma))


This is the end of our first data analysis, where we applied the VI-ME-BA-BAR schema to classify pictures of apples and bananas. We applied a lightweight preprocessing to the images to extract two features: the redness and elongation of the depicted fruits. Then, we [VI]sualized the obtained data, introduced a [ME]tric for the classification problem, namely the error rate for predictions. We proceeded with a [BA]seline method using a simple nearest neighbor classifier. Finally, we estimated the performance of this first classifier by computing error [BAR]s over many samples of training / testing sets.

The obtained predictions are fairly robust, which is unsurprising given that, up to a few gentle outliers, the pictures in the data are well constrained.

Aiming at more complex data sets from real life, we will in the following weeks progressively enrich this schema with more tools.

You have reached the end of this assignment. Congratulations!

All you have to do now is to double check the quality of your code with the code review notebook, submit your work, and fetch the feedback.

If you can’t wait to explore further, you can engage in the next section where you will build your own classifier.

Extra credit: build a oneR classifier

Using the template below, creates a “one rule” (oneR) classifier which:

  • selects the “good” feature G (Redness or Elongation), which is most correlated (in absolute value) to the fruit target values y = +- 1;

  • uses G to classify a new example as an apple or a banana by setting a threshold on its values.

You may follow the template below or try other ideas of your own.

class oneR():
    def __init__(self):
        This constructor is supposed to initialize data members.
        Use triple quotes for function documentation. 
        self.is_trained = False  
        self.ig = 0     # Index of the good feature G
        self.w = 1      # Feature polarity
        self.theta = 0  # Threshold on the good feature

    def fit(self, X, Y):
        This function should train the model parameters.
            X: Training data matrix of dim num_train_samples * num_feat.
            Y: Training label matrix of dim num_train_samples * 1.
        Both inputs are panda dataframes.
        # Compute correlations
        ### BEGIN SOLUTION
        C = np.corrcoef(pd.concat((X, Y), axis=1).T)
        ### END SOLUTION
        np.fill_diagonal(C, 0) # avoid that the max be the diagonal value
        # Select the most correlated feature in absolute value using the last line,
        # and store it in self.ig
        ### BEGIN SOLUTION
        last_line = C[-1]
        self.ig = np.argmax(np.abs(last_line))
        ### END SOLUTION
        # Get feature polarity and store it in self.w
        ### BEGIN SOLUTION
        self.w = np.sign(last_line[self.ig])
        ### END SOLUTION
        # Fetch the feature values and multiply by polarity
        G = X.iloc[:, self.ig] * self.w
        # Compute the threshold as a mid-point between cluster centers
        ### BEGIN SOLUTION
        self.theta = (np.mean(G[Y>0]) + np.mean(G[Y<0]))/2
        ### END SOLUTION
        print("FIT: Training Successful: Feature selected = %d; Polarity = %d; Threshold = %5.2f." % (self.ig, self.w, self.theta))

    def predict(self, X):
        This function should provide predictions of labels on (test) data.
            X: Test data matrix of dim num_test_samples * num_feat.
            Y: Predicted label matrix of dim num_test_samples * 1.
        # Fetch the feature of interest and multiply by polarity
        G = X.iloc[:,self.ig] * self.w
        # Make decisions according to threshold
        Y = G.copy()
        Y[G < self.theta] = -1  
        Y[G >= self.theta] = 1  
        print("PREDICT: Prediction done")
        return Y
# Use this code to test your classifier
clf = oneR(), Ytrain) 
Ytrain_predicted = clf.predict(Xtrain)
Ytest_predicted = clf.predict(Xtest)
e_tr = error_rate(Ytrain, Ytrain_predicted)
e_te = error_rate(Ytest, Ytest_predicted)
print("Training error:", e_tr)
print("Test error:", e_te)
# This is what you get as decision boundary.
# The training examples are shown as white circles and the test examples are blue squares.
make_scatter_plot(X, images.apply(transparent_background_filter),
                  [], test_index, 
                  feat = clf.ig, theta=clf.theta, axis='square')
# Compare with what you would get if you used both features, voting with the same weight.
make_scatter_plot(X, images.apply(transparent_background_filter),
                  [], test_index, 
                  show_diag=True, axis='square')