def test_tensor_matmul():
a = sn.ones((3, 4), requires_grad=True)
b = sn.ones((4, 3), requires_grad=True)
c = sn.matmul(a, b)
c.set_retain_grad()
d = c.sum()
d.backward()
def test_tensor_sub():
a = sn.zeros((3, 4), requires_grad=True)
a = a - 1.0
a = 2.0 - a
b = sn.ones((3, 1))
b = b - a
c = sn.ones((1, 4))
c = c - b
d = sn.random((3, 4))
d = d - c
e = d.sum()
e.backward()
def test_tensor_add():
a = sn.zeros((3, 4), requires_grad=True)
a = a + 1.0
a = 2.0 + a
b = sn.ones((3, 1))
b = b + a
c = sn.ones((1, 4))
c = c + b
d = sn.random((3, 4))
d = d + c
e = d.sum()
e.backward()
def test_tensor_opt():
x = sn.ones((2, 1), requires_grad=True)
b = sn.ones((2, 1))
data = b.detach()
data[0] = 0.5
for i in range(1000):
y = sn.sum((x * b - 1)**2)
y.backward()
x.update_data(0.01)
x.zero_grad()
data = x.to_cpu().detach()
assert np.linalg.norm(data - np.array([[2], [1]])) < 1e-2
def test_tensor_asstride():
x = sn.ones((3, 4), requires_grad=True)
b = x[:, 2]
c = x[:, 2] * 0.33
d = c + b
c = d.sum()
c.backward()
print(c, x, x.grad)
def test_tensor_div():
a = 0.5 / sn.ones((3, 4), requires_grad=True)
a = a / 2.0
b = sn.random((3, 1))
c = b / a
d = sn.random((1, 4))
d = d / c
e = d.sum()
e.backward()
def test_tensor_mul():
a = 0.5 * sn.ones((3, 4), requires_grad=True)
a = a * 2.0
b = sn.random((3, 1))
c = b * a
d = sn.random((1, 4))
d = d * c
e = d.sum()
e.backward()
def optimize_and_store(name="sgd"):
storage = np.zeros((n, 2))
x = sn.ones((2, 1), requires_grad=True)
opt = create_optimizer(name, {"x": x}, optim_options)
for i in range(n):
storage[i] = x.data.reshape((-1,))
x1, x2 = x[0], x[1]
y = x1 ** 2 + 2 * x2 ** 2 - 2 * x1 * x2 - 4 * x1
y.backward()
opt.step()
opt.zero_grad()
return storage
def test_tensor_log():
a = sn.ones((3, 4), requires_grad=True)
b = sn.log(a)
c = b.sum()
c.backward()
def test_tensor_pow():
a = sn.ones((3, 4), requires_grad=True)
b = sn.pow(a, 2)
c = b.sum()
c.backward()