def test_opt_cloner(self): # Dumb test. Just for coverage
opt_1 = opt.create_opt(opt.SGD())
opt_1.learning_rate = 0.5
opt_2 = opt.create_opt(opt_1, opt_1.learning_rate)
self.assertIsInstance(opt_1, opt.SGD)
self.assertIsInstance(opt_2, opt.SGD)
self.assertTrue(opt_1.learning_rate == opt_2.learning_rate)
def select_optimizer(self):
if self.opt_type == self._Optimizer.MOMENTUM:
return opt.Momentum(momentum=self.momentum)
elif self.opt_type == self._Optimizer.NAG:
return opt.NesterovMomentum(momentum=self.momentum)
elif self.opt_type == self._Optimizer.ADAGRAD:
return opt.AdaGrad()
elif self.opt_type == self._Optimizer.RMSPROP:
return opt.RMSProp(rho=self.rho)
elif self.opt_type == self._Optimizer.ADADELTA:
return opt.AdaDelta(rho=self.rho)
elif self.opt_type == self._Optimizer.ADAM:
return opt.Adam(beta1=self.beta1, beta2=self.beta2)
elif self.opt_type == self._Optimizer.ADAMAX:
return opt.Adamax(beta1=self.beta1, beta2=self.beta2)
else:
return opt.SGD()
import orangecontrib.recommendation.optimizers as opt
import numpy as np
import collections
import unittest
import copy
__optimizers__ = [
opt.SGD(learning_rate=0.1),
opt.Momentum(learning_rate=0.1, momentum=0.5),
opt.NesterovMomentum(learning_rate=0.1, momentum=0.5),
opt.AdaGrad(learning_rate=0.1),
opt.RMSProp(learning_rate=0.01, rho=0.9, epsilon=1e-6),
opt.AdaDelta(learning_rate=1, rho=0.95, epsilon=1e-6),
opt.Adam(learning_rate=0.01, beta1=0.9, beta2=0.999, epsilon=1e-8),
opt.Adamax(learning_rate=0.01, beta1=0.9, beta2=0.999, epsilon=1e-8)
]
def dxf(X):
tensor = [0.1, 0.2, 0.3]
return tensor * X * 2 # f(x) = x^2 -> dx(f(x)) = 2x
def dxf2(X):
return X * 2
class TestOptimizers(unittest.TestCase):
# These tests compare results on a toy problem to values