def test_variable_dot():
a_array = np.arange(5)
b_array = np.array([3, 8, 5, 6, 8])
a = Variable(np.arange(5, dtype=float))
b = Variable(np.array([3, 8, 5, 6, 8], dtype=float))
expected_result = 68
assert np.all(a_array == a.data)
assert np.all(b_array == b.data)
assert id(a_array) != id(a.data)
assert id(b_array) != id(b.data)
def validate_result(result, expected_type: type) -> None:
assert result == expected_result
assert isinstance(result, expected_type)
return
# Variable + Variable
validate_result(a.dot(b), Variable)
validate_result(np.dot(a, b), Variable)
# numpy + numpy
validate_result(np.dot(a_array, b_array), np.int64)
validate_result(np.ndarray.dot(a_array, b_array), np.int64)
validate_result(a_array.dot(b_array), np.int64)
# Variable + numpy
validate_result(a.dot(b_array), Variable)
validate_result(np.dot(a, b_array), Variable)
# numpy + Variable
validate_result(np.dot(a_array, b), Variable)
validate_result(np.ndarray.dot(a_array, b), Variable)
validate_result(a_array.dot(b), Variable)
# Verify Derivative
sgd = autograd.optimizer.SGD(learning_rate=1e-3)
dot_product = a.dot(b)
variable_to_gradient = sgd.take_training_step(dot_product)
assert all(
isinstance(var, Variable) and isinstance(grad, np.ndarray)
for var, grad in variable_to_gradient.items())
assert np.all(variable_to_gradient[a] == b_array)
assert np.all(variable_to_gradient[b] == a_array)
# Verify Trainability
x = Variable(np.random.rand(2))
y = 0
sgd = autograd.optimizer.SGD(learning_rate=1e-4)
for training_step_index in range(50):
y_hat = x.dot(np.array([-10, 50]))
diff = np.subtract(y, y_hat)
loss = diff**2
if 0 < loss < 1e-3:
break
sgd.take_training_step(loss)
assert 0 < loss < 1e-3
def test_variable_depended_on_variables():
'''
The computation graph is:
f
/|
/ |
e |
| |
d |
/ | |
| | |
| c |
| / \ |
|/ \|
a b
'''
a = Variable(np.random.rand(8))
b = Variable(np.random.rand(8))
c = a.dot(b)
d = a - c
e = d ** 2
f = b - e
topologically_sorted_variables = list(f.depended_on_variables())
assert len(topologically_sorted_variables) == len(set(topologically_sorted_variables)) == 6
topologically_sorted_variable_ids = eager_map(id, topologically_sorted_variables)
a_index = topologically_sorted_variable_ids.index(id(a))
b_index = topologically_sorted_variable_ids.index(id(b))
c_index = topologically_sorted_variable_ids.index(id(c))
d_index = topologically_sorted_variable_ids.index(id(d))
e_index = topologically_sorted_variable_ids.index(id(e))
f_index = topologically_sorted_variable_ids.index(id(f))
assert f_index < e_index
assert f_index < b_index
assert e_index < d_index
assert d_index < a_index
assert d_index < c_index
assert c_index < a_index
assert c_index < b_index
