def reconstruct(rbm_struct, test_image):
pre_sigmoid_activation_h = np.dot(rbm_struct.weights,
test_image) #+ rbm_struct.ci
sigmoid_activation_h = function_logistic(pre_sigmoid_activation_h)
pre_sigmoid_activation_v = np.dot(rbm_struct.weights[:, 785:795].T,
sigmoid_activation_h) #+ rbm_struct.bj
sigmoid_activation_v = function_logistic(pre_sigmoid_activation_v)
return sigmoid_activation_v
def function_feed_forward(inputData, backPropStruct):
#Create activation function for inputNodes
backPropStruct.Y_i[1:] = inputData
#Compute hidden activation function from input layer
backPropStruct.V_j = np.dot(backPropStruct.W_ji, backPropStruct.Y_i)
backPropStruct.Y_j[1:] = function_logistic(backPropStruct.V_j[1:])
#Compute output activation function from hidden layer
backPropStruct.V_k = np.dot(backPropStruct.W_kj, backPropStruct.Y_j)
backPropStruct.Y_k = function_logistic(backPropStruct.V_k)
return backPropStruct
def get_reconstruction_cross_entropy(rbm_struct):
pre_sigmoid_activation_h = np.tensordot(rbm_struct.visible_states.T,
rbm_struct.weights[:, :, :], 1).T
#pre_sigmoid_activation_h = np.dot(rbm_struct.weights , rbm_struct.visible_states)
sigmoid_activation_h = function_logistic(pre_sigmoid_activation_h)
pre_sigmoid_activation_v = np.dot(rbm_struct.weights.T,
sigmoid_activation_h) + rbm_struct.bj
sigmoid_activation_v = function_logistic(pre_sigmoid_activation_v)
cross_entropy = -np.mean(
np.sum(
rbm_struct.visible_states * np.log(sigmoid_activation_v) +
(1 - rbm_struct.visible_states) * np.log(1 - sigmoid_activation_v),
axis=1))
return cross_entropy
def get_y_given_h(rbm_struct, x_i_input_image, h0_sample):
a = np.tensordot(x_i_input_image.T, rbm_struct.weights[:, :, :], 1).T
b = np.tensordot(h0_sample, a, 1)
p_v_c_h_k = function_logistic(b + rbm_struct.bj)
sampled_p_v_c_h_k = rbm_struct.rng.binomial(size=p_v_c_h_k.shape,
n=1,
p=p_v_c_h_k)
return [p_v_c_h_k, sampled_p_v_c_h_k]
def get_h_given_v(rbm_struct, x_i_input_image, y_j_output_image):
a = np.tensordot(
np.tensordot(x_i_input_image.T, rbm_struct.weights[:, :, :], 1).T,
y_j_output_image)
a = function_reshape_arr(a, rbm_struct.h_val, 1, 0)
p_h_c_v_k = function_logistic(a + rbm_struct.ci)
p_h_c_v_k = function_reshape_arr(p_h_c_v_k, rbm_struct.h_val, 0, 0)
sampled_p_h_c_v_k = rbm_struct.rng.binomial(size=p_h_c_v_k.shape,
n=1,
p=p_h_c_v_k)
return [p_h_c_v_k, sampled_p_h_c_v_k]
def get_h_given_v(rbm_struct, x_i_input_image, y_j_output_image):
a = np.tensordot(
np.tensordot(x_i_input_image.T, rbm_struct.weights[:, :, :], 1).T,
y_j_output_image)
p_h_c_v_k = function_logistic(a)
sampled_p_h_c_v_k = rbm_struct.rng.binomial(
size=p_h_c_v_k.shape, # discrete: binomial
n=1,
p=p_h_c_v_k)
return [p_h_c_v_k, sampled_p_h_c_v_k]