def FRQ_3():
train_features, test_features, train_targets, test_targets = \
load_data('mnist-multiclass', fraction=0.75)
learner = MultiClassGradientDescent(loss='squared', regularization='l1')
learner.fit(train_features, train_targets, batch_size=20, max_iter=700)
predictions = learner.predict(test_features)
print("predictionsL: ", predictions.shape)
print("test_targets: ", test_targets.shape)
c_m = confusion_matrix(test_targets, predictions.astype(int))
visualized_c_m = np.append(np.arange(-1,5).reshape(1,6), np.append(np.arange(5).reshape(5,1), c_m, axis = 1), axis=0 )
print("confusion matrix: ", visualized_c_m)
def test_multiclass_gradient_descent_blobs():
"""
Tests that the multiclass classifier also works on binary tasks
"""
from your_code import MultiClassGradientDescent
np.random.seed(0)
features, _, targets, _ = load_data('blobs')
learner = MultiClassGradientDescent(loss='squared', regularization=None,
learning_rate=0.01, reg_param=0.05)
learner.fit(features, targets, batch_size=None, max_iter=1000)
predictions = learner.predict(features)
print("predictions: ", predictions)
print("targets: ", targets)
def test_multiclass_gradient_descent_mnist():
"""
Tests the ability of the multiclass gradient descent classifier to classify
a non-trivial problem with a reasonable accuracy.
"""
from your_code import MultiClassGradientDescent, accuracy
np.random.seed(0)
train_features, test_features, train_targets, test_targets = \
load_data('mnist-multiclass', fraction=0.8)
learner = MultiClassGradientDescent(loss='squared',
regularization=None,
learning_rate=0.01,
reg_param=0.05)
learner.fit(train_features, train_targets, batch_size=None, max_iter=1000)
predictions = learner.predict(test_features)
assert accuracy(test_targets, predictions) >= 0.85
def test_multiclass_gradient_descent_separable():
"""
Tests the ability of the OVA multiclass gradient descent classifier to
classify a simple, separable dataset.
"""
from your_code import MultiClassGradientDescent
np.random.seed(0)
features = np.identity(4)
targets = np.array(range(4))
learner = MultiClassGradientDescent(loss='squared',
regularization=None,
learning_rate=0.01,
reg_param=0.05)
learner.fit(features, targets, batch_size=None, max_iter=1000)
predictions = learner.predict(features)
assert np.all(predictions == targets)