def test_basicrmsprop_max_scaling():
a = shared_floatx([1e-6, 1e-6])
cost = (a**2).sum()
step_rule = BasicRMSProp(decay_rate=0.5, max_scaling=1e5)
steps, updates = step_rule.compute_steps(
OrderedDict([(a, tensor.grad(cost, a))]))
f = theano.function([], [steps[a]], updates=updates)
assert_allclose(f()[0], [0.2, 0.2])
def test_basicrmsprop_max_scaling():
a = shared_floatx([1e-6, 1e-6])
cost = (a ** 2).sum()
step_rule = BasicRMSProp(decay_rate=0.5, max_scaling=1e5)
steps, updates = step_rule.compute_steps(
OrderedDict([(a, tensor.grad(cost, a))]))
f = theano.function([], [steps[a]], updates=updates)
assert_allclose(f()[0], [0.2, 0.2])
def test_basicrmsprop():
a = shared_floatx([3, 4])
cost = (a**2).sum()
step_rule = BasicRMSProp(decay_rate=0.5, max_scaling=1e5)
steps, updates = step_rule.compute_steps(
OrderedDict([(a, tensor.grad(cost, a))]))
f = theano.function([], [steps[a]], updates=updates)
assert_allclose(f()[0], [1.41421356, 1.41421356])
a.set_value([2, 3])
assert_allclose(f()[0], [0.9701425, 1.02899151])
a.set_value([1, 1.5])
assert_allclose(f()[0], [0.6172134, 0.64699664])
def test_basicrmsprop():
a = shared_floatx([3, 4])
cost = (a ** 2).sum()
step_rule = BasicRMSProp(decay_rate=0.5, max_scaling=1e5)
steps, updates = step_rule.compute_steps(
OrderedDict([(a, tensor.grad(cost, a))]))
f = theano.function([], [steps[a]], updates=updates)
assert_allclose(f()[0], [1.41421356, 1.41421356])
a.set_value([2, 3])
assert_allclose(f()[0], [0.9701425, 1.02899151])
a.set_value([1, 1.5])
assert_allclose(f()[0], [0.6172134, 0.64699664])
def test_basicrmsprop_broadcastable():
verify_broadcastable_handling(BasicRMSProp(0.5, 1e5))
