class NNTest(unittest.TestCase):
def setUp(self):
self.nn = NeuralNetwork(NUM_OUTPUT_LABELS, NUM_FEATURES)
def test_initialize_weights(self):
w1, w2 = self.nn._initialize_weights()
#rows of w1 should = NUM_HIDDEN_UNITS, cols = NUM_FEATURES + 1 (bias)
self.assertEqual(w1.shape[0], NUM_HIDDEN_UNITS)
self.assertEqual(w1.shape[1], NUM_FEATURES + 1)
#rows of w2 = NUM_OUTPUT_LABELS, cols = NUM_HIDDEN_UNITS + 1 (bias)
self.assertEqual(w2.shape[0], NUM_OUTPUT_LABELS)
self.assertEqual(w2.shape[1], NUM_HIDDEN_UNITS + 1)
def test_one_hot_encoding(self):
#create a labels array of size num_samples, where y[i] is the corresponding label of that instance.
y = [1, 5, 3, 0, 2, 5, 9,
4] #ie, training instance 2 has label y[i=2] = 3
#max(y) <= NUM_OUTPUT_LABELS -1
encoded = self.nn._encode_labels(np.asarray(y), NUM_OUTPUT_LABELS)
for i in range(encoded.shape[1]):
#get encoded row vector
row = encoded[:, i]
#get index where this vector is one, which should be the label
label = np.where(row == 1)[0]
self.assertEqual(y[i], label)
def test_add_bias_unit(self):
#set up testing arrays
X = np.zeros((5, 4))
X_new = self.nn._add_bias_unit(X)
arr_ones = np.ones(X_new.shape[1])
arr_zeros = np.zeros(X_new.shape[1])
#add bias unit in first column
#first column should now be all ones, rest should be all zeros
self.assertEqual(X_new[:, 0].tolist(), arr_ones.tolist())
for i in range(1, X.shape[1]):
self.assertEqual(X_new[:, i].tolist(), arr_zeros.tolist())
X_new_row = self.nn._add_bias_unit(X, column=False)
self.assertEqual(X_new_row[0, :].tolist(),
np.ones(X_new_row.shape[1]).tolist())
