0
$\begingroup$

I have built a multiclass perceptron, but it has low accuracy (around 80%). I think I'm missing something. One possibility is that I should add a bias, but I'm not sure how to incorporate that.

The task is, given 2 dimensions, predict the class, which is between 0 and 8.

I could use some pointers as to where this code is going wrong. We are not given classes for test data, but rather need to populate it and it is checked elsewhere.

def dot_product(weights_list, training_data):
    return [sum([a * b for a, b in zip(training_data, i)]) for i in weights_list]

def predict(row, weights_list, n_classes):
    # get dimensions * weight
    activations = dot_product(weights_list, row[:-1])
    # get index of argmax
    predicted_label = activations.index(max(activations))
    return predicted_label

def train_weights(train, n_epoch, n_classes):
    # prepopulate just weights
    weights_list = [[0.0 for x in range(len(train[0])-1)] for x in range(n_classes)]
    for epoch in range(n_epoch):
        for row in train:
            actual_class = int(row[-1])
            # argmax from predicting each class
            prediction = int(predict(row, weights_list, n_classes))
            # if incorrect:
            #   lower score of wrong answer by this row's values
            #   raise score of correct answer by this row's values
            if actual_class != prediction:
                weights_list[prediction] = [a - b for a, b in zip(weights_list[prediction], row[:-1])]
                weights_list[actual_class] = [a + b for a, b in zip(weights_list[actual_class], row[:-1])]

    return weights_list

train = [[83.0, -14.0, 6.0],
        [77.0, 15.0, 6.0],
        [93.0, 35.0, 3.0],
        [86.0, -8.0, 6.0],
        [-51.0, -79.0, 1.0],
        [62.0, -73.0, 1.0]]

test = [[36.0, 27.0, -1],
        [6.0, 99.0, -1],
        [-3.0, 16.0, -1],
        [-40.0, -61.0, -1],
        [70.0, 67.0, -1],
        [86.0, -14.0, -1],
        [-92.0, 67.0, -1]]

n_classes = 9
n_epoch = 10000
weights_list = train_weights(train, n_epoch, n_classes)

for row in test:
    prediction = predict(row, weights_list, n_classes)
    row[2] = prediction
$\endgroup$
7
  • $\begingroup$ How many output classes are you expecting? Regardless, 80% is not necessarily bad. You should try your code using some toy data to see if there are any errors in the code. But, 80% is not necessarily a bad result for a multiclass perceptron depending on the complexity of the classification task. $\endgroup$ – JahKnows May 28 '18 at 2:32
  • $\begingroup$ For this project, accuracy needs to be at least 90%. The classes could be from 0 to 8, so 9 classes. $\endgroup$ – Adam_G May 28 '18 at 3:25
  • $\begingroup$ is the list train the entirety of your training set? What are their respective labels? $\endgroup$ – JahKnows May 28 '18 at 3:32
  • $\begingroup$ Can you post the entirety of your code please? What is the data_test variable? $\endgroup$ – JahKnows May 28 '18 at 3:33
  • 1
    $\begingroup$ Thanks! No, the data is the same dimensions as above, 2. The 3rd value is the class. $\endgroup$ – Adam_G May 28 '18 at 3:42
1
$\begingroup$

Firstly, in all the linear separator algorithms such as linear regression, logistic regression and the perceptron, adding the bias is as simple as adding a feature column consisting of all 1's. Then the third weight that will be trained will act as the bias $b$.


I have some working code for a multi-class perceptron

First let's generate some artificial data with 2 features. Each distribution of points is going to be Gaussian with a given mean and variance described as a list of lists for each dimension under the variable params.

def gen_data(params, n):
    dims = len(params[0])
    num_classes = len(params)

    x = np.zeros((n*num_classes, dims))
    y = np.zeros((n*num_classes,))
    for ix, i in enumerate(range(num_classes)):
        inst = np.random.randn(n, dims)

        for dim in range(dims): x[ix*n:(ix+1)*n,dim] = np.random.normal(params[ix][dim][0], 
                                                                        params[ix][dim][1], n)
        y[ix*n:(ix+1)*n] = ix
    return x, y

params = [[[ 5,1],  [ 5,1]], 
          [[ 0,1],  [ 0,1]],
          [[2, 1],  [ 2,1]],
          [[-2, 1],  [ 2,1]]]

n = 300
x, y = gen_data(params, 300)

plt.scatter(x[:,0], x[:,1])
plt.show()

enter image description here

Alright so now we have 4 distributions with different labels. Let's split the data for sanctity's sake.

x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.33)

Let's train the weights using the training data

def get_weights(x, y, n_epochs, verbose = 0):
    # Append a ones column to the feature for the bias
    data = np.ones((x.shape[0], x.shape[1]+1))
    data[:, 0:x.shape[1]] = x
    # Set the targets as integers for comparison
    targets = y.astype(int)

    # Initialize the weights as a matrix 
    # number of classes by number of features
    weights = np.ones((len(set(y)), x.shape[1]+1))
    for epoch in range(n_epochs):
        for i, target in zip(data, targets):
            temp = np.dot(i, weights.T)
            pred = np.argmax(temp)
            # If wrongly predicted update prediction
            if pred != target:
                weights[target, :] = weights[target, :] + i
                weights[pred, :] = weights[pred, :] - i

        if verbose == 1:
            print('Iteration: ', epoch)
            print(weights)
            print('---------------------------------------------')
    return weights

weights = get_weights(x_train, y_train, n_epochs = 30, verbose = 1)

This converges to approximately this

[[ 23.62752045 16.03867499 -111. ]
[ -3.96545848 -8.66924406 47. ]
[ -0.94290763 -1.84413793 33. ]
[ -14.71915434 -1.52529301 35. ]]

We get an accuracy calculated using the score

def predict(x, weights):
    data = np.ones((  x.shape[0], x.shape[1]+1  ))
    data[:, 0:x.shape[1]] = x
    predictions = np.argmax(np.dot(data, weights.T), axis = 1)
    return predictions

def score(x, y, weights):
    pred = predict(x, weights)
    return sum(pred == y_test)/len(pred)

score(x_test, y_test, weights)

0.8686868686868687

We can check our results using a confusion matrix. For the training set

from sklearn.metrics import confusion_matrix

predictions = predict(x_train, weights)
plt.imshow(confusion_matrix(y_train, predictions))
plt.show()

enter image description here

And the testing set

predictions = predict(x_test, weights)
plt.imshow(confusion_matrix(y_test, predictions))
plt.show()

enter image description here

So we see that in fact our algorithm is performing quite well.

We can then plot our points to see how it is classifying them. I will plot the training points as small circles, and the testing points as larger ones. The dark points are those which are misclassified

colors = ['y', 'r', 'b', 'g', 'k']

# Predict training set
predictions = predict(x_train, weights)
for i, t, p in zip(x_train, y_train, predictions):
    if t == p: plt.scatter(i[0], i[1], c=colors[int(t)], alpha = 0.2, s=20)
    else: plt.scatter(i[0], i[1], c=colors[int(t)], alpha = 1)

# Predict test set
predictions = predict(x_test, weights)
for i, t, p in zip(x_test, y_test, predictions):
    if t == p: plt.scatter(i[0], i[1], c=colors[int(t)], alpha = 0.2)
    else: plt.scatter(i[0], i[1], c=colors[int(t)], alpha = 1)

# Plot the linear separators
x1 = np.linspace(np.min(x[:,0]),np.max(x[:,1]),2)
x2 = np.zeros((weights.shape[0], 2))
for ix_w, weight in enumerate(weights):
    x2 = 1 * ( - weight[2] - weight[0]*x1) / weight[1]
    plt.plot(x1, x2, c = colors[ix_w])

plt.xlabel('Feature 1')
plt.ylabel('Feature 2')

plt.xlim([np.min(x[:,0]), np.max(x[:,0])])
plt.ylim([np.min(x[:,1]), np.max(x[:,1])])

plt.show()

enter image description here


This code generalizes to the binary classification task as well

params = [[[ 5,1],  [ 5,1]], 
          [[ 0,1],  [ 0,1]]]

enter image description here

enter image description here


Stop training on convergence

If you want to stop the algorithm based on convergence you can use a stop criteria. For example you can stop training once every weight in your matrix changes by less than a very small number. The very small number we usually choose is machine epsilon 2.220446049250313e-16, which is essentially zero. Sometimes this requirement is too stringent so it can be replaced by any number of significant values.

Change the get_weights code to include the break criteria as

from copy import deepcopy

def get_weights(x, y, n_epochs, verbose = 0):
    # Append a ones column to the feature for the bias
    data = np.ones((x.shape[0], x.shape[1]+1))
    data[:, 0:x.shape[1]] = x
    # Set the targets as integers for comparison
    targets = y.astype(int)

    # Initialize the weights as a matrix 
    # number of classes by number of features
    weights = np.zeros((len(set(y)), x.shape[1]+1))
    past_weights = np.zeros((len(set(y)), x.shape[1]+1))
    for epoch in range(n_epochs):
        for i, target in zip(data, targets):
            temp = np.dot(i, weights.T)
            pred = np.argmax(temp)
            # If wrongly predicted update prediction
            if pred != target:
                weights[target, :] = weights[target, :] + i
                weights[pred, :] = weights[pred, :] - i

        if np.abs(weights - past_weights).all() < np.finfo(float).eps:
            break
        past_weights = deepcopy(weights)

        if verbose == 1:
            print('Iteration: ', epoch)
            print(weights)
            print('---------------------------------------------')
    return weights
$\endgroup$
4
  • $\begingroup$ This was incredibly verbose, and very helpful. Thank you so much! $\endgroup$ – Adam_G May 28 '18 at 14:49
  • $\begingroup$ No problem! Hope it works, let me know if anything goes wrong. $\endgroup$ – JahKnows May 28 '18 at 15:36
  • $\begingroup$ One additional question: I had originally run this for a fixed number of epochs. Now they want it to terminate upon convergence. How do I detect convergence? $\endgroup$ – Adam_G May 29 '18 at 22:54
  • $\begingroup$ I'll update the answer $\endgroup$ – JahKnows May 30 '18 at 3:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.