def test_pow_grad():
a = ad.variable(2)
b = ad.variable(3)
c = ad.pow(a, b)
dc_da = c.grad(a)
dc_db = c.grad(b)
with ad.Session() as sess:
dc_da_out = sess.run(dc_da)
dc_db_out = sess.run(dc_db)
assert dc_da_out.value == 3 * 2**2
assert dc_db_out.value == 2**3 * math.log(2)
def test_add_grad():
a = ad.variable(1)
b = ad.variable(2)
c = ad.add(a, b)
dc_da = c.grad(a)
dc_db = c.grad(b)
with ad.Session() as sess:
dc_da_out = sess.run(dc_da)
dc_db_out = sess.run(dc_db)
assert dc_da_out.value == 1
assert dc_db_out.value == 1
def test_reciprocal_grad():
a = ad.variable(2)
c = ad.reciprocal(a)
dc_da = c.grad(a)
with ad.Session() as sess:
dc_da_out = sess.run(dc_da)
assert dc_da_out.value == -1 / 4
def test_log_grad():
a = ad.variable(2)
c = ad.log(a)
dc_da = c.grad(a)
with ad.Session() as sess:
dc_da_out = sess.run(dc_da)
assert dc_da_out.value == 1 / 2
def test_trainable_parameters():
node1 = ad.constant(1)
node2 = ad.variable(2)
node3 = ad.variable(3)
node4 = ad.variable(4)
node5 = ad.constant(5)
node1_node2 = ad.mul(node1, node2)
node1_node3 = ad.mul(node1, node3)
node1_node4 = ad.mul(node1, node4)
node_sum = ad.add(node1_node2,
node1_node3,
node1_node4)
final_node = ad.mul(node_sum, node5)
trainable_nodes = final_node.trainable_parameters()
assert len(trainable_nodes) == 3
for node in trainable_nodes:
assert node.value in [2, 3, 4]
def test_trainable_parameters_grads():
node1 = ad.constant(1)
node2 = ad.variable(2)
node3 = ad.variable(3)
node4 = ad.variable(4)
node5 = ad.constant(5)
node1_node2 = ad.mul(node1, node2)
node1_node3 = ad.mul(node1, node3)
node1_node4 = ad.mul(node1, node4)
node_sum = ad.add(node1_node2, node1_node3, node1_node4)
final_node = ad.mul(node_sum, node5)
trainable_nodes = final_node.trainable_parameters()
with Session() as sess:
out = sess.run(final_node)
assert out.value == 45
# test grad
trainable_node_grads = final_node.grad(trainable_nodes)
assert len(trainable_node_grads) == len(trainable_nodes)
with Session() as sess:
trainable_node_grads = sess.run(trainable_node_grads)
for node, grad in zip(trainable_nodes, trainable_node_grads):
print(grad.value, node.value)
continue
if node.name == '2':
assert grad.value == 5
elif node.name == '3':
assert grad.value == 5
elif node.name == '4':
assert grad.value == 5
else:
raise ValueError('Not a trainable node')
assert False