def test_sigmoid_dx_gradient():
epsilon = 1e-4
for x in np.arange(-4, 4.5, 0.5):
sl = sigmoid(x - epsilon)
sr = sigmoid(x + epsilon)
s_dx = (sr - sl) / (2 * epsilon)
assert_almost_equal(sigmoid_dx(x), s_dx)
def test_sigmoid_dx_gradient():
epsilon = 1e-4
for x in np.arange(-4, 4.5, 0.5):
sl = sigmoid(x-epsilon)
sr = sigmoid(x+epsilon)
s_dx = (sr - sl) / (2 * epsilon)
assert_almost_equal(sigmoid_dx(x), s_dx)
def forward_pass(self, theta, X, carry=None):
W_in, W_r = self.unpackTheta(theta)
X = np.atleast_2d(X)
Y = np.zeros((X.shape[0], self.output_dim))
if carry is None:
carry = np.zeros((1, self.output_dim))
for i, x in enumerate(X):
carry = sigmoid(x.dot(W_in) + carry.dot(W_r)) # sigmoid
Y[i] = carry
return Y
def forward_pass(self, theta, X, carry=None):
W_in, W_r = self.unpackTheta(theta)
X = np.atleast_2d(X)
Y = np.zeros((X.shape[0], self.output_dim))
if carry is None:
carry = np.zeros((1, self.output_dim))
for i, x in enumerate(X):
carry = sigmoid(x.dot(W_in) + carry.dot(W_r)) # sigmoid
Y[i] = carry
return Y
def test_sigmoid_np_arrays():
X = np.array(known_values.keys())
Y = np.array(known_values.values())
assert_almost_equal(sigmoid(X), Y)
def test_sigmoid_some_numbers():
for x, y in known_values.items():
assert_almost_equal(sigmoid(x), y)
def forward_pass(self, _, X):
return sigmoid(X)
def test_sigmoid_np_arrays():
X = np.array(known_values.keys())
Y = np.array(known_values.values())
assert_almost_equal(sigmoid(X), Y)
def test_sigmoid_some_numbers():
for x, y in known_values.items():
assert_almost_equal(sigmoid(x), y)
#!/usr/bin/python
# coding: utf-8
from __future__ import division, unicode_literals, print_function
import numpy as np
from scipy.optimize import approx_fprime
from helpers import *
from neural_nets.connections import FullConnection, FullConnectionWithBias, SigmoidLayer, RecurrentConnection, ForwardAndRecurrentConnection, ForwardAndRecurrentSigmoidConnection
from neural_nets.fann import FANN
from neural_nets.functions import sigmoid
theta = np.array([[-1, 1, 0, 1]]).reshape(-1)
X = np.array([[0, 0, 0, 1], [1, 0, 0, 1], [0, 1, 0, 1], [0, 0, 1, 1],
[1, 1, 0, 1]])
X_nb = X[:, :-1] # no bias included
T = sigmoid(np.array([[1, 0, 2, 1, 1]]).T)
E = 0.5 * T**2
def test_FANN_dimensions():
fc = FullConnection(5, 1)
nn = FANN([fc])
assert_equal(nn.input_size, 5)
assert_equal(nn.output_size, 1)
def test_FANN_with_bias_dimensions():
fc = FullConnectionWithBias(5, 1)
nn = FANN([fc])
assert_equal(nn.input_size, 5)
assert_equal(nn.output_size, 1)
#!/usr/bin/python
# coding: utf-8
from __future__ import division, unicode_literals, print_function
import numpy as np
from scipy.optimize import approx_fprime
from helpers import *
from neural_nets.connections import FullConnection, FullConnectionWithBias, SigmoidLayer, RecurrentConnection, ForwardAndRecurrentConnection, ForwardAndRecurrentSigmoidConnection
from neural_nets.fann import FANN
from neural_nets.functions import sigmoid
theta = np.array([[-1, 1, 0, 1]]).reshape(-1)
X = np.array([[0, 0, 0, 1], [1, 0, 0, 1],[0, 1, 0, 1],[0, 0, 1, 1],[1, 1, 0, 1]])
X_nb = X[:,:-1] # no bias included
T = sigmoid(np.array([[1, 0, 2, 1, 1]]).T)
E = 0.5 * T**2
def test_FANN_dimensions():
fc = FullConnection(5, 1)
nn = FANN([fc])
assert_equal(nn.input_size, 5)
assert_equal(nn.output_size, 1)
def test_FANN_with_bias_dimensions():
fc = FullConnectionWithBias(5, 1)
nn = FANN([fc])
assert_equal(nn.input_size, 5)
assert_equal(nn.output_size, 1)
def test_FANN_feed_forward_single_sample():
fc = FullConnection(4, 1)
def forward_pass(self, _, X):
return sigmoid(X)
