def test_initialize_weights_returns_correct_weight_sizes(self):
linear = Linear(name='linear', combination='x,y')
weights = linear.initialize_weights(input_shape=(2, 4, 3))
assert isinstance(weights, list) and len(weights) == 2
weight_vector, bias = weights
assert K.int_shape(weight_vector) == (3 * 2, 1)
assert K.int_shape(bias) == (1, )
def test_compute_similarity_works_with_subtract_combinations(self):
linear = Linear(name='linear', combination='x-y')
linear.weight_vector = K.variable(numpy.asarray([[-.3], [.5]]))
linear.bias = K.variable(numpy.asarray([0]))
a_vectors = numpy.asarray([[1, 1], [-1, -1]])
b_vectors = numpy.asarray([[1, 0], [0, 1]])
result = K.eval(
linear.compute_similarity(K.variable(a_vectors),
K.variable(b_vectors)))
assert result.shape == (2, )
assert_almost_equal(result, [.5, -.7])
def test_compute_similarity_works_with_higher_order_tensors(self):
linear = Linear(name='linear', combination='x,y')
weights = numpy.random.rand(14, 1)
linear.weight_vector = K.variable(weights)
linear.bias = K.variable(numpy.asarray([0]))
a_vectors = numpy.random.rand(5, 4, 3, 6, 7)
b_vectors = numpy.random.rand(5, 4, 3, 6, 7)
result = K.eval(
linear.compute_similarity(K.variable(a_vectors),
K.variable(b_vectors)))
assert result.shape == (5, 4, 3, 6)
combined_vectors = numpy.concatenate(
[a_vectors[3, 2, 1, 3, :], b_vectors[3, 2, 1, 3, :]])
expected_result = numpy.dot(combined_vectors, weights)
assert_almost_equal(result[3, 2, 1, 3], expected_result, decimal=6)
def test_compute_similarity_does_a_weighted_product(self):
linear = Linear(name='linear', combination='x,y')
linear.weight_vector = K.variable(
numpy.asarray([[-.3], [.5], [2.0], [-1.0]]))
linear.bias = K.variable(numpy.asarray([.1]))
a_vectors = numpy.asarray([[[1, 1], [-1, -1]]])
b_vectors = numpy.asarray([[[1, 0], [0, 1]]])
result = K.eval(
linear.compute_similarity(K.variable(a_vectors),
K.variable(b_vectors)))
assert result.shape == (
1,
2,
)
assert_almost_equal(result, [[2.3, -1.1]])