def test_1d_simples():
"""Test trivial gradient updates."""
LOGGER.info("Testing Addition.")
a = tj.var(np.ones(3) * 5)
b = tj.var(np.ones(3) * 10)
c = tj.add(a, b)
for i in range(5):
g = tj.gradients(c, [a])
LOGGER.info("%s *** a: %s -- b: %s -- g: %s -- dt: 0.1" % (i, a, b, g))
a.update(a.v + g[0] * 1e-1)
assert _true(a.v > 5.0), "A should be larger than 5 but is %s" % a
LOGGER.info("Testing Subtraction.")
a = tj.var(np.ones(3) * 5)
b = tj.var(np.ones(3) * 10)
c = tj.sub(a, b)
for i in range(5):
g = tj.gradients(c, [b])
LOGGER.info("%s *** a: %s -- b: %s -- g: %s -- dt: 0.1" % (i, a, b, g))
b.update(b.v + g[0] * 1e-1)
assert _true(b.v < 10.0), "A should be smaller than 10 but is %s" % a
LOGGER.info("Testing Multiplication.")
a = tj.var(np.ones(3) * 5)
b = tj.var(np.ones(3) * 10)
c = tj.mul(a, b)
for i in range(5):
g = tj.gradients(c, [a])
LOGGER.info("%s *** a: %s -- b: %s -- g: %s -- dt: 0.1" % (i, a, b, g))
a.update(a.v + g[0] * 1e-1)
assert _true(a.v > 5.0), "A should be larger than 5 but is %s" % a
LOGGER.info("Testing Division.")
a = tj.var(np.ones(3) * 5)
b = tj.var(np.ones(3) * 10)
c = tj.div(a, b)
for i in range(5):
g = tj.gradients(c, [b])
LOGGER.info("%s *** a: %s -- b: %s -- g: %s -- dt: 0.1" % (i, a, b, g))
b.update(b.v + g[0] * 1e-1)
assert _true(b.v < 10.0), "B should be smaller than 10 but is %s" % a
def test_linear_regression():
"""Test making simple 1d linear regression and train it."""
x = np.arange(0, 10)
# Co-domain
y = x + 5
# Variables
a = tj.var(np.random.rand())
b = tj.var(np.random.rand())
err = tj.mse(y, a * x + b)
LOGGER.info("Optimizing a simple 2 linear regression.")
LOGGER.info(" X domain is [0, 1 .. 10]")
LOGGER.info(" Y domain is [5, 6, .. 15]")
LOGGER.info("before training: coefficient %s -- bias: %s -- mse: %s" %
(a, b, err.output()))
for i in range(1200):
g = tj.gradients(err, [a, b])
a.update(a.v - np.mean(g[0]) * 1e-2)
b.update(b.v - np.mean(g[1]) * 1e-2)
LOGGER.info("after training: coefficient %s -- bias: %s -- mse: %s" %
(a, b, err.output()))
def test_logistic_regression():
"""Test making simple 1d linear regression and train it."""
x = np.random.rand(10) - 0.5
y = sigmoid(4 * x - 1)
a = tj.var(np.random.rand())
b = tj.var(np.random.rand())
o = tj.sigmoid(a * x + b)
err = tj.mse(y, o)
LOGGER.info("Optimizing a simple logistic regression.")
LOGGER.info(" X domain is (10, 2)")
LOGGER.info(" Y domain is (10, 1)")
LOGGER.info("before training: coefficient %s -- bias: %s -- mse: %s" %
(a, b, err.output()))
for i in range(5000):
g = tj.gradients(err, [a, b])
a.update(a.v - np.mean(g[0]) * 1e-0)
b.update(b.v - np.mean(g[1]) * 1e-0)
LOGGER.info("after training: coefficient %s -- bias: %s -- mse: %s" %
(a, b, err.output()))
LOGGER.info("perfect would be 4 and -1")
LOGGER.info("predictions %s", np.round(o.output()))
LOGGER.info("observations %s", np.round(y))
def minimise(self, nodes: ["node.node"]) -> None:
"""Minimise op."""
cache = list(enumerate(nodes))
for i in range(self.rounds):
grads = tensorjo.gradients(self.master, nodes)
for i, n in cache:
n.update(n.v - np.mean(grads[i]) * self.dt)
def test_convex():
"""Test optimising a simple convex function."""
LOGGER.info("Testing simple convex: x^2")
for s in [1, 3]:
a = tj.var(np.ones(s))
b = tj.mul(a, a)
LOGGER.info("Initial a: %s" % a)
for i in range(100):
g = tj.gradients(b, [a])
a.update(a.v - g[0] * 1e-1)
LOGGER.info("Final a: %s" % a)
assert _true(
abs(a.v) < 1.0), "A should be smaller than 1 but is %s" % a
LOGGER.info("Testing more complex convex: (x * 5 + 3 - x)^2")
for s in [1, 3]:
a = tj.var(np.ones(s))
b = tj.var(5)
c = tj.mul(a, b)
c = tj.add(c, 3)
c = tj.sub(c, a)
c = tj.mul(c, c)
LOGGER.info("Initial a: %s" % a)
for i in range(100):
g = tj.gradients(c, [a])
a.update(a.v - g[0] * 1e-2)
LOGGER.info("Final a: %s" % a.v)
assert _true(
abs(a.v) < 1.0), "A should be smaller than 1 but is %s" % a