def test_FANN_recurrent_gradient_multisample():
rc = ForwardAndRecurrentConnection(4, 1)
nn = FANN([rc])
theta = 2 * np.ones((nn.get_param_dim()))
grad_c = nn.calculate_gradient(theta, X, T)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, X, T)
assert_allclose(grad_c, grad_e, rtol=1e-3, atol=1e-5)
def test_FANN_recurrent_gradient_multisample():
rc = ForwardAndRecurrentConnection(4,1)
nn = FANN([rc])
theta = 2 * np.ones((nn.get_param_dim()))
grad_c = nn.calculate_gradient(theta, X, T)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, X, T)
assert_allclose(grad_c, grad_e, rtol=1e-3, atol=1e-5)
def test_FANN_gradient_multisample():
fc = FullConnection(4, 1)
sig = SigmoidLayer(1)
nn = FANN([fc, sig])
theta = np.random.randn(nn.get_param_dim())
grad_c = nn.calculate_gradient(theta, X, T)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, X, T)
assert_almost_equal(grad_c, grad_e)
def test_FANN_with_bias_gradient_multisample():
fc = FullConnectionWithBias(3, 1)
sig = SigmoidLayer(1)
nn = FANN([fc, sig])
theta = np.random.randn(nn.get_param_dim())
grad_c = nn.calculate_gradient(theta, X_nb, T)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, X_nb, T)
assert_almost_equal(grad_c, grad_e)
def test_FANN_recurrent_gradient_single_sample():
rc = ForwardAndRecurrentConnection(1, 1)
nn = FANN([rc])
theta = 2 * np.ones((nn.get_param_dim()))
for x, t in [[0, 1], [1, 1], [0, 0]]:
x = np.array([[x]])
grad_c = nn.calculate_gradient(theta, x, t)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, x, t)
assert_almost_equal(grad_c, grad_e)
def test_FANN_with_bias_gradient_single_sample():
fc = FullConnectionWithBias(3, 1)
sig = SigmoidLayer(1)
nn = FANN([fc, sig])
theta = np.random.randn(nn.get_param_dim())
for x, t in zip(X_nb, T) :
grad_c = nn.calculate_gradient(theta, x, t)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, x, t)
assert_almost_equal(grad_c, grad_e)
def test_FANN_recurrent_gradient_single_sample():
rc = ForwardAndRecurrentConnection(1,1)
nn = FANN([rc])
theta = 2 * np.ones((nn.get_param_dim()))
for x, t in [[0, 1], [1, 1], [0, 0]] :
x = np.array([[x]])
grad_c = nn.calculate_gradient(theta, x, t)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, x, t)
assert_almost_equal(grad_c, grad_e)
def test_FANN_gradient_single_sample():
fc = FullConnection(4, 1)
sig = SigmoidLayer(1)
nn = FANN([fc, sig])
theta = np.random.randn(nn.get_param_dim())
for x, t in zip(X, T):
grad_c = nn.calculate_gradient(theta, x, t)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, x, t)
assert_almost_equal(grad_c, grad_e)
def test_FANN_multilayer_gradient_multisample():
fc0 = FullConnectionWithBias(4, 2)
fc1 = FullConnectionWithBias(2, 1)
sig0 = SigmoidLayer(2)
sig1 = SigmoidLayer(1)
nn = FANN([fc0, sig0, fc1, sig1])
theta = np.random.randn(nn.get_param_dim())
grad_c = nn.calculate_gradient(theta, X, T)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, X, T)
assert_almost_equal(grad_c, grad_e)
def test_FANN_multilayer_with_bias_gradient_multisample():
fc0 = FullConnectionWithBias(3, 2)
fc1 = FullConnectionWithBias(2, 1)
sig0 = SigmoidLayer(2)
sig1 = SigmoidLayer(1)
nn = FANN([fc0, sig0, fc1, sig1])
theta = np.random.randn(nn.get_param_dim())
grad_c = nn.calculate_gradient(theta, X_nb, T)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, X_nb, T)
assert_almost_equal(grad_c, grad_e)
def test_FANN_converges_on_vote_problem():
fc = FullConnectionWithBias(9, 1)
sig = SigmoidLayer(1)
nn = FANN([fc, sig])
vote = generate_majority_vote()
theta = np.zeros((10, ))
for i in range(500):
g = nn.calculate_gradient(theta, vote.data, vote.target)
theta -= g * 1
error = nn.calculate_error(theta, vote.data, vote.target)
assert_less(error, 0.2)
def test_FANN_converges_on_vote_problem():
fc = FullConnectionWithBias(9, 1)
sig = SigmoidLayer(1)
nn = FANN([fc, sig])
vote = generate_majority_vote()
theta = np.zeros((10,))
for i in range(500):
g = nn.calculate_gradient(theta, vote.data, vote.target)
theta -= g * 1
error = nn.calculate_error(theta, vote.data, vote.target)
assert_less(error, 0.2)
def test_FANN_converges_on_and_problem():
fc = FullConnection(2, 1)
sig = SigmoidLayer(1)
nn = FANN([fc, sig])
and_ = load_and()
theta = np.array([-0.1, 0.1])
for i in range(100):
g = nn.calculate_gradient(theta, and_.data, and_.target)
theta -= g * 1
error = nn.calculate_error(theta, and_.data, and_.target)
assert_less(error, 0.2)
def test_FANN_converges_on_and_problem():
fc = FullConnection(2, 1)
sig = SigmoidLayer(1)
nn = FANN([fc, sig])
and_ = load_and()
theta = np.array([-0.1, 0.1])
for i in range(100):
g = nn.calculate_gradient(theta, and_.data, and_.target)
theta -= g * 1
error = nn.calculate_error(theta, and_.data, and_.target)
assert_less(error, 0.2)
def test_FANN_multilayer_gradient_single_sample():
fc0 = FullConnection(4, 2)
fc1 = FullConnection(2, 1)
sig0 = SigmoidLayer(2)
sig1 = SigmoidLayer(1)
nn = FANN([fc0, sig0, fc1, sig1])
theta = np.random.randn(nn.get_param_dim())
for x, t in zip(X, T) :
grad_c = nn.calculate_gradient(theta, x, t)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, x, t)
assert_almost_equal(grad_c, grad_e)
def test_FANN_with_bias_multilayer_gradient_single_sample():
fc0 = FullConnectionWithBias(3, 2)
fc1 = FullConnectionWithBias(2, 1)
sig0 = SigmoidLayer(2)
sig1 = SigmoidLayer(1)
nn = FANN([fc0, sig0, fc1, sig1])
theta = np.random.randn(nn.get_param_dim())
for x, t in zip(X_nb, T):
grad_c = nn.calculate_gradient(theta, x, t)
grad_e = approx_fprime(theta, nn.calculate_error, 1e-8, x, t)
assert_almost_equal(grad_c, grad_e)
def test_FANN_converges_on_xor_problem():
fc0 = FullConnectionWithBias(2, 2)
fc1 = FullConnectionWithBias(2, 1)
sig0 = SigmoidLayer(2)
sig1 = SigmoidLayer(1)
nn = FANN([fc0, sig0, fc1, sig1])
xor = load_xor()
theta = np.random.randn(nn.get_param_dim())
for i in range(2000):
g = nn.calculate_gradient(theta, xor.data, xor.target)
theta -= g * 1
error = nn.calculate_error(theta, xor.data, xor.target)
assert_less(error, 0.4)
def test_RANN_converges_on_ropot_problem():
frc = ForwardAndRecurrentSigmoidConnection(5, 5)
nn = FANN([frc])
rpot = generate_remember_pattern_over_time()
theta = np.random.randn(nn.get_param_dim())
for i in range(100):
grad = np.zeros_like(theta)
for X, T in seqEnum(rpot):
grad += nn.calculate_gradient(theta, X, T)
theta -= grad * 1
error = np.sum(nn.calculate_error(theta, X, T) for X, T in seqEnum(rpot))
assert_less(error, 10.)
def test_FANN_converges_on_xor_problem():
fc0 = FullConnectionWithBias(2, 2)
fc1 = FullConnectionWithBias(2, 1)
sig0 = SigmoidLayer(2)
sig1 = SigmoidLayer(1)
nn = FANN([fc0, sig0, fc1, sig1])
xor = load_xor()
theta = np.random.randn(nn.get_param_dim())
for i in range(2000):
g = nn.calculate_gradient(theta, xor.data, xor.target)
theta -= g * 1
error = nn.calculate_error(theta, xor.data, xor.target)
assert_less(error, 0.4)
def test_RANN_converges_on_ropot_problem():
frc = ForwardAndRecurrentSigmoidConnection(5, 5)
nn = FANN([frc])
rpot = generate_remember_pattern_over_time()
theta = np.random.randn(nn.get_param_dim())
for i in range(100):
grad = np.zeros_like(theta)
for X, T in seqEnum(rpot):
grad += nn.calculate_gradient(theta, X, T)
theta -= grad * 1
error = np.sum(nn.calculate_error(theta, X, T) for X, T in seqEnum(rpot))
assert_less(error, 10.)
