def calc_error(self, data, follow_theta=False):
if self._style == LayerTypes.OUTPUT:
self._delta = self._nodes - data
else:
self._delta = np.multiply(
data * np.transpose(follow_theta)[:, 1:],
np.multiply(sigmoid(self._z_val), 1 - sigmoid(self._z_val)))
def indexed_weighted_logistic_regression(x_train,
x_test,
y_train,
y_test,
num_iter=10000,
Regulization_lamda=0.0001,
lr=0.001,
weight_Sigma=0.8):
weights = Calculate_logistic_weights(x_train, x_test, weight_Sigma)
weights = np.identity(weights.shape[0]) * weights
#weights=np.reshape(weights,(weights.shape[0],1))
theta = np.reshape(np.zeros(x_train.shape[1]), (x_train.shape[1], 1))
for i in range(num_iter):
z = np.dot(x_train, theta)
h = sigmoid(z)
h = np.reshape(h, (h.shape[0], 1))
#print(x_train.shape)
#print(weights.shape)
X = np.matmul(weights, x_train)
gradient = -np.dot(X.T, (h - y_train)) + (theta) * Regulization_lamda
#gradient = -np.dot(np.matmul(weights,x_train).T, (h - y_train)) + (theta) * Regulization_lamda
theta = theta + lr * gradient
y_pred = sigmoid(np.dot(x_test.T, theta))
#print(abs(y_test - y_pred),end='#############################\n')
return y_pred
def reconstruction_likelihood(net, t_image = 0.250):
spikes = np.zeros((len(x), n_outputs))
pi = ut.sigmoid(net._V)
likelihoods =[]
estimation_net = deepcopy(net)
estimation_net._current_time = 0
estimation_net._trace = deque([])
pbar = tqdm(total=len(x) * t_image, unit='Time [s]', position=1, desc="Reconstruction")
while estimation_net._current_time < len(x) * t_image:
pbar.n = int(estimation_net._current_time * 1000) / 1000
pbar.update(0)
z = estimation_net.step(lambda t: x[int(min(t, (len(x)-1) * t_image) / t_image)], update_weights=False)
sample = x[int(min(estimation_net._current_time, (len(x)-1) * t_image) / t_image)]
pi = ut.sigmoid(np.dot(z.reshape((1, -1)), net._V))
likelihoods.append(np.sum(np.log(sample * pi + (1 - sample) * (1 - pi)), axis=-1))
pbar.close()
return np.mean(likelihoods)
def forward_prop(self, inputs):
""" Calculate output from given inputs through the neural network """
layers = [inputs]
for i in range(len(self.h_layers)+1):
# Calculate the input * weights + bias
z = np.dot(layers[i], self.weights[i]) + self.biases[i]
# Apply activation function
out = []
if d.sensor_mode:
# Outputs are numbers
for j in range(len(z)):
o = sigmoid(clamp(-20, 20, z[j]))
out.append(o)
else:
# Output are vectors
for j in range(len(z)):
sigm_x = sigmoid(clamp(-20, 20, z[j].x))
sigm_y = sigmoid(clamp(-20, 20, z[j].y))
out.append(Vector2(sigm_x, sigm_y))
layers.append(out)
# Return the output
final_output = layers[len(layers)-1][0] # Last layer only has 1 output neuron
return final_output
def logistic_regression(x_train,
x_test,
y_train,
y_test,
num_iter=10000,
Regulization_lamda=0.0001,
lr=0.001):
theta = np.reshape(np.zeros(x_train.shape[1]), (x_train.shape[1], 1))
for i in range(num_iter):
z = np.dot(x_train, theta)
h = sigmoid(z)
h = np.reshape(h, (h.shape[0], 1))
gradient = -np.dot(x_train.T,
(h - y_train)) + (theta) * Regulization_lamda
#gradient = -np.dot(X.T, (h - y))
theta = theta + lr * gradient
z = np.dot(x_test, theta)
h = sigmoid(z)
h = np.reshape(h, (h.shape[0], 1))
acc = 0
for i in range(len(y_test)):
if ((h[i] > 0.5 and y_test[i] == 1)
or (h[i] < 0.5 and y_test[i] == 0)):
acc += 1
acc = acc / y_test.shape[0]
#print("simple Accurecy : "+str(acc))
return 100 * acc
def reconstruct_from_input(self, input):
output = numpy.dot(input, self.W.T) + self.hbias
hidden_possible = sigmoid(output)
input_after = numpy.dot(hidden_possible, self.W) + self.vbias
input_possible = sigmoid(input_after)
assert input.shape == input_possible.shape
return input_possible
def contrast_divergence_binomial(self, epoch):
dw = numpy.zeros((self.output_size, self.input_size))
# dvb = numpy.zeros(self.input_size)
# dhb = numpy.zeros(self.output_size)
output = numpy.dot(self.input, self.W.T) + self.hbias
hidden_possible = sigmoid(output)
hidden_state = numpy_rng.binomial(n=1, p=hidden_possible)
dw += self.learning_rate * numpy.dot(hidden_state.T, self.input)
# dvb += self.learning_rate * numpy.mean(self.input, axis=0)
# dhb += self.learning_rate * numpy.mean(hidden_possible, axis=0)
visible_output = numpy.dot(hidden_state, self.W) + self.vbias
visible_possible = sigmoid(visible_output)
visible_state = numpy_rng.binomial(n=1, p=visible_possible)
hidden_output = numpy.dot(visible_state, self.W.T) + self.hbias
hidden_possible_after = sigmoid(hidden_output)
dw -= self.learning_rate * numpy.dot(hidden_possible_after.T, visible_state)
# dvb -= self.learning_rate * numpy.mean(visible_state, axis=0)
# dhb -= self.learning_rate * numpy.mean(hidden_possible_after, axis=0)
####################
# parameter update
####################
self.W += dw / self.data_size
def get_reconstruction_cross_entropy(self):
pre_sigmoid_activation_h = numpy.dot(self.input, self.W.T) + self.hbias
sigmoid_activation_h = sigmoid(pre_sigmoid_activation_h)
pre_sigmoid_activation_v = numpy.dot(sigmoid_activation_h, self.W) + self.vbias
sigmoid_activation_v = sigmoid(pre_sigmoid_activation_v)
cross_entropy = - numpy.mean(
numpy.sum(self.input * numpy.log(sigmoid_activation_v) +
(1 - self.input) * numpy.log(1 - sigmoid_activation_v),
axis=1))
return cross_entropy
def predict(self, DAG, predictors):
if not isinstance(predictors, dict):
predictors = util.array2dict(predictors)
Z = DAG.get_latent_nodes()
Y = DAG.get_response_nodes()
X = DAG.get_input_nodes()
non_input_nodes = Z.union(Y)
non_input_nodes_by_rep = sorted(non_input_nodes, key=DAG.get_reputation, reverse=True)
x = next(iter(X))
n = len(predictors[x])
node_values = dict()
for node in non_input_nodes_by_rep:
if node in DAG.parents:
parents = DAG.parents[node]
weighted_parent_values = np.zeros((len(parents), n))
for ix, parent in enumerate(parents):
if parent in X:
parvals = predictors[parent]
else:
parvals = node_values[parent]
w = DAG.get_weight(parent, node)
weighted_parent_values[ix, :] += self.boolean_NOT(parvals, w)
node_values[node] = self.boolean_MEDIAN(weighted_parent_values)
else:
node_values[node] = np.zeros((n))
y = next(iter(Y))
parents_y = DAG.parents[y]
parent_y_values = np.zeros((len(parents_y), n))
for ix, i in enumerate(parents_y):
w = DAG.get_weight(i, y)
if i in X:
parent_y_values[ix, :] = self.boolean_NOT(predictors[i], w)
elif i in Z:
parent_y_values[ix, :] = self.boolean_NOT(node_values[i], w)
if parents_y == 1:
theta = util.sigmoid(2*parent_y_values-1, gain=self.gain)
else:
theta = util.sigmoid(2*np.mean(parent_y_values, axis=0)-1, gain=self.gain)
response = 1*(np.random.rand(n) < theta)
return response
def log_likelihood(self, DAG, predictors, response, epsilon=1e-3):
if not isinstance(predictors, dict):
predictors = util.array2dict(predictors)
X = DAG.get_input_nodes()
Z = DAG.get_latent_nodes()
Y = DAG.get_response_nodes()
zvalues = dict()
# We assume only 1 response node
y = next(iter(Y))
n = len(response)
# highest reputation first
Z_by_rep = sorted(Z, key=DAG.get_reputation, reverse=True)
for z in Z_by_rep:
if not z in DAG.parents:
zvalues[z] = np.zeros(shape=(n,))
else:
parents = DAG.parents[z]
parent_values = np.zeros((len(parents), n))
for ix, i in enumerate(parents):
w = DAG.get_weight(i, z)
if i in X:
parent_values[ix, :] = self.boolean_NOT(predictors[i], w)
else:
parent_values[ix, :] = self.boolean_NOT(zvalues[i], w)
parent_values = np.array(parent_values)
zvalue = self.boolean_MEDIAN(parent_values)
zvalues[z] = zvalue
parents_y = DAG.parents[y]
parent_y_values = np.zeros((len(parents_y), n))
for ix, i in enumerate(parents_y):
w = DAG.get_weight(i, y)
if i in X:
parent_y_values[ix, :] = self.boolean_NOT(predictors[i], w)
elif i in Z:
parent_y_values[ix, :] = self.boolean_NOT(zvalues[i], w)
if parents_y == 1:
theta = util.sigmoid(2*parent_y_values-1, gain=self.gain)
else:
theta = util.sigmoid(2*np.mean(parent_y_values, axis=0)-1, gain=self.gain)
L = np.sum(response*np.log(theta) + (1 - response)*np.log(1 - theta))
return L
def get_accuracy(x, y, theta):
length = len(x)
correct = 0
for i in range(length):
prediction = 1 if sigmoid(x[i].dot(theta)) >= 0.5 else 0
correct = correct + 1 if y[i] == prediction else correct
return (correct / length) * 100
def step(self, inputs):
inputs = inputs.reshape((-1, 1))
assert len(inputs) == self._n_inputs, "Input length does not match"
# u = V * input + b
u = np.dot(self._V, inputs) + self._b
z = np.zeros((self._n_outputs, 1))
# find out if network is spiking
if np.random.uniform(0, 1, 1) < self._delta_t * self._r_net:
# p = softmax(u)
p_z = np.exp(u) / np.sum(np.exp(u) + 1e-8)
# sample from softmax distribution
sum_p_z = np.cumsum(p_z)
diff = sum_p_z - np.random.uniform(0, 1, 1) > 0
k = np.argmax(diff)
z[k] = 1.0
self._b += self._delta_t * self._eta_b * (
self._delta_t * self._r_net * self._m_k - ut.dirac(z - 1))
self._V += self._delta_t * self._eta_v * ut.dirac(z - 1) * (
inputs.T - ut.sigmoid(self._V))
return z
def predict(self, X, Y):
m = X.shape[-1]
Z = np.dot(self.w, X) + self.b
A = sigmoid(Z)
self.Y_p = (A > 0.5)
correct = (self.Y_p == Y)
self.accuracy = np.sum(correct) / m
return self.accuracy
def linear_forward_block(a_prev, w, b, activation):
h = np.dot(w, a_prev) + b
if activation == 'sigmoid':
a = sigmoid(h)
else:
a = tanh(h)
cache = {'a_prev': a_prev, 'w': w, 'b': b}
return a, cache
def predict(self, x):
"""
Predicts value for a given datapoint x.
variables:
t - datapoint for which a value gets predicted by the model
"""
x = np.insert(x, 0, 1)
prob = sigmoid(x.dot(self.theta))
if self.verbose:
print(f'Prediction for {x[1:]} is: {100*prob:.1f}%')
return prob
def compute_cost(X, Y, w, b, lambd, regularized):
m = Y.shape[-1]
Z = np.dot(w, X) + b
A = sigmoid(Z)
cost = - np.sum(Y * np.log(A) + (1-Y) * np.log(1-A)) / m
if regularized == 1:
cost += lambd * np.linalg.norm(w, 1) / m
elif regularized == 2:
cost += lambd * np.linalg.norm(w, 2) / m
return cost
def forward_prop(self, inputs):
layers = [inputs]
for i in range(len(self.h_layers) + 1):
# Calculate the input * weights + bias
z = sum_matrix_float(np.dot(layers[i], self.weights[i]),
self.biases[i])
# Apply activation function
o = [sigmoid(clamp(-20, 20, z[j])) for j in range(len(z))]
layers.append(o)
# Return the output
final_output = layers[len(layers) - 1][0]
return final_output
def reconstruct(net, input, t_image=0.250):
estimation_net = deepcopy(net)
estimation_net._current_time = 0
estimation_net._trace = deque([])
reconstruction = np.zeros_like(input)
while estimation_net._current_time < t_image:
z = estimation_net.step(lambda t: input, update_weights=False)
reconstruction += z.dot(ut.sigmoid(net._V))
return reconstruction
def costFun(self, theta):
"""
Returns objective value for measuring fitness of model.
variables:
theta - current model parameters
"""
if self.verbose:
print("Iter: {} | theta: {}".format(self.iter, theta))
J = 0
m = len(self.y)
# Using np.finfo(float).eps to avoid dividing by zero errors/warnings
cost = -self.y * np.log(
sigmoid(self.X.dot(theta)) + np.finfo(float).eps) - (np.ones([
m,
]) - self.y) * np.log(1 - sigmoid(self.X.dot(theta)) +
np.finfo(float).eps)
J = 1 / m * sum(cost)
self.costHist.append(J)
self.theta = theta
self.iter += 1
return J
def compute_dd_loss(self, fst, snd, third):
x = self.predict_score(fst, snd, "dd") - \
self.predict_score(fst, third, "dd")
ranking_loss = -np.log(sigmoid(x))
complexity = 0.0
complexity += self.matrix_reg * np.dot(self.paper_latent_matrix[fst],
self.paper_latent_matrix[fst])
complexity += self.matrix_reg * np.dot(self.paper_latent_matrix[snd],
self.paper_latent_matrix[snd])
complexity += self.matrix_reg * np.dot(self.paper_latent_matrix[third],
self.paper_latent_matrix[third])
return ranking_loss + complexity
def generate_R_PR(self):
self.Pr = np.zeros((self.num_user, self.num_item))
self.R = np.zeros((self.num_user, self.num_item))
rel_list = []
exp_list = []
for m in range(self.C):
P, Q, c, d, a, b, e, f = self.sess.run([
self.P_list[m], self.Q_list[m], self.c_list[m], self.d_list[m],
self.a_list[m], self.b_list[m], self.e_list[m], self.f_list[m]
])
rel = np.matmul(P, Q.T)
rel = np.exp(rel)
rel /= np.sum(rel, axis=1, keepdims=True)
rel *= (self.num_item / 2.)
w = utility.sigmoid(np.matmul(Q, a) + b)
pop = np.power(w * utility.sigmoid(np.matmul(Q, c) + d) \
+ (1 - w) * self.item_pop, utility.sigmoid(np.matmul(Q, e) + f))
exp = np.zeros((self.num_user, self.num_item)) + pop.T
user_ids = self.df_list[m]['userId']
item_ids = self.df_list[m]['itemId']
rel_list.append(rel[user_ids, item_ids])
exp_list.append(exp[user_ids, item_ids])
self.R += rel
self.Pr += exp
self.R /= self.C
self.Pr /= self.C
for m in range(self.C):
user_ids = self.df_list[m]['userId']
item_ids = self.df_list[m]['itemId']
self.R[user_ids, item_ids] = rel_list[m]
self.Pr[user_ids, item_ids] = exp_list[m]
self.Pr[np.where(self.Pr < 0.01)] = 0.01
self.Pr[np.where(self.Pr > 0.99)] = 0.99
def costFunGrad(self, theta):
"""
Returns gradient of the objective.
variables:
theta - current model parameters
"""
Grad = np.zeros(theta.shape)
m = len(self.y)
for i in range(len(theta)):
Grad[i] = 1 / m * sum(
(sigmoid(self.X.dot(theta)) - self.y) * self.X[:, i])
return Grad
def step(self, data_generator_fn, update_weights=True):
# sample isi
isi = -np.log(np.random.uniform()) / self._r_net
new_time = self._current_time + isi
# now go back from T + isi - tau to T + isi, calculate input data
# calculate the activations
# update the weights
time_start = max(0, new_time - 2 * self._tau)
total_inputs = data_generator_fn(new_time).reshape((-1, 1))
for time in np.arange(time_start, new_time, self._delta_t):
inputs = data_generator_fn(time)
inputs = inputs.reshape((-1, 1))
inputs *= np.exp(-(new_time - time) / self._tau)
assert len(inputs) == self._n_inputs, "Input length does not match"
total_inputs += inputs
# u = V * input + b
u = np.dot(self._V, inputs) + self._b
z = np.zeros((self._n_outputs, 1))
# p = softmax(u)
p_z = np.exp(u) / np.sum(np.exp(u) + 1e-8)
# sample from softmax distribution
sum_p_z = np.cumsum(p_z)
diff = sum_p_z - np.random.uniform(0, 1, 1) > 0
k = np.argmax(diff)
z[k] = 1.0
if update_weights:
self._b += self._eta_b * (isi * self._r_net * self._m_k -
ut.dirac(z - 1))
self._V += self._eta_v * ut.dirac(z - 1) * (inputs.T -
ut.sigmoid(self._V))
self._current_time += isi
self._trace.append(
(self._current_time, inputs, z, u, self._V, self._b))
if len(self._trace) > self._max_trace_length:
self._trace.pop()
return z
def compute_output(self):
"""
Returns the output of this Neuron node, using a sigmoid as
the threshold function.
returns: number (float or int)
"""
inputs_, weights_ = np.array([
x.output() for x in self.my_inputs
]), np.array([x.my_value for x in self.my_weights])
# print(inputs_)
# print(weights_)
out = sigmoid(np.sum(inputs_ * weights_))
return out
raise NotImplementedError("Implement me!")
def compute_pp_loss(self, fst, snd, third):
"""
loss includes ranking loss and model complexity
"""
x = self.predict_score(fst, snd, "pp") - \
self.predict_score(fst, third, "pp")
ranking_loss = -np.log(sigmoid(x))
complexity = 0.0
complexity += self.matrix_reg * np.dot(self.author_latent_matrix[fst],
self.author_latent_matrix[fst])
complexity += self.matrix_reg * np.dot(self.author_latent_matrix[snd],
self.author_latent_matrix[snd])
complexity += self.matrix_reg * np.dot(
self.author_latent_matrix[third], self.author_latent_matrix[third])
return ranking_loss + complexity
def accuracy(self, threshold=0.5):
"""
Computes the accuracy of the trained model for the training set.
variables:
threshold - threshold for accepting a value as 1
"""
p = np.zeros([
len(self.y),
])
for i in range(len(self.y)):
if sigmoid(self.X[i].dot(self.theta)) >= threshold:
p[i] = 1
else:
p[i] = 0
self.acc = np.mean(p == self.y) * 100
def logistic_propagate(X, Y, w, b, lambd, regularized):
m = Y.shape[-1]
Z = np.dot(w, X) + b
A = sigmoid(Z)
cost = compute_cost(X, Y, w, b, lambd, regularized)
dw = np.dot((A - Y), X.T) / m
db = np.sum((A - Y), axis=1, keepdims=True) / m
if regularized == 2:
dw += 2 * lambd * w / m
elif regularized == 1:
dw += lambd * np.sign(w) / m
grad = {'dw':dw, 'db':db}
return grad, cost
def reconstruction_l2_loss(net, t_image=0.250):
estimation_net = deepcopy(net)
estimation_net._current_time = 0
estimation_net._trace = deque([])
spikes = np.zeros((len(x), n_outputs))
pbar = tqdm(total=len(x) * t_image, unit='Time [s]', position=1, desc="Reconstruction")
while estimation_net._current_time < len(x) * t_image:
pbar.n = int(estimation_net._current_time * 1000) / 1000
pbar.update(0)
z = estimation_net.step(lambda t: x[int(min(t, (len(x) - 1) * t_image) / t_image)], update_weights=False)
spikes[min(len(x)-1, int(estimation_net._current_time / t_image))] += z.flatten()
reconstructions = np.dot(spikes, ut.sigmoid(estimation_net._V)) / np.sum(spikes, axis=-1).reshape(-1, 1)
difference = np.mean((reconstructions - x) ** 2)
return difference
def update_dd_gradient(self, fst, snd, third):
x = self.predict_score(fst, snd, "dd") - \
self.predict_score(fst, third, "dd")
common_term = sigmoid(x) - 1
grad_fst = common_term * (self.paper_latent_matrix[snd] - \
self.paper_latent_matrix[third]) + \
2 * self.matrix_reg * self.paper_latent_matrix[fst]
self.paper_latent_matrix[fst] = self.paper_latent_matrix[fst] - \
self.alpha * grad_fst
grad_snd = common_term * self.paper_latent_matrix[fst] + \
2 * self.matrix_reg * self.paper_latent_matrix[snd]
self.paper_latent_matrix[snd]= self.paper_latent_matrix[snd] - \
self.alpha * grad_snd
grad_third = -common_term * self.paper_latent_matrix[fst] + \
2 * self.matrix_reg * self.paper_latent_matrix[third]
self.paper_latent_matrix[third] = self.paper_latent_matrix[third] - \
self.alpha * grad_third
def estimate_likelihood(estimation_duration=10.0):
log_likelihoods = deque([])
estimation_net = deepcopy(net)
estimation_net._current_time = 0
estimation_net._trace = deque([])
while estimation_net._current_time < estimation_duration:
estimation_net.step(lambda t: data_generator[t], update_weights=False)
pbar.n = int(net._current_time * 1000) / 1000
pbar.update(0)
# log likelihood
y = estimation_net._trace[-1][1].reshape((1, -1))
pi = ut.sigmoid(net._V)
log_likelihoods.append(
np.log(1.0 / n_outputs) + np.log(np.sum(np.prod(y * pi + (1 - y) * (1 - pi), axis=-1))))
return np.mean(log_likelihoods), np.std(log_likelihoods)
def __init__(self, weights):
self._fig = plt.figure(figsize=(3.5, 1.16), dpi=300)
i = 2
num_weights = len(weights)
while i < len(weights):
if num_weights % i == 0:
break
else:
i += 1
axes = add_axes_as_grid(self._fig,
i,
int(num_weights / i),
m_xc=0.01,
m_yc=0.01)
self._weight_shape = weights.shape[1:]
self._imshows = []
for i, ax in enumerate(list(axes.flatten())):
# disable legends
ax.set_yticks([])
ax.set_xticks([])
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.spines['left'].set_visible(False)
ax.spines['bottom'].set_visible(False)
if i >= len(weights):
self._imshows.append(
ax.imshow(np.zeros(self._weight_shape), vmin=0, vmax=1))
else:
self._imshows.append(
ax.imshow(ut.sigmoid(weights[i].reshape(
self._weight_shape)),
vmin=0,
vmax=1))
plt.show(block=False)
self._fig.canvas.draw()
self._fig.canvas.flush_events()
def reconstruction(nets, Xs, t_image=0.250):
reconstructions = np.zeros_like(x)
w = W // Xs.shape[0]
h = H // Xs.shape[0]
for n in range(Xs.shape[0]):
for m in range(Xs.shape[1]):
net = nets[n][m]
estimation_net = deepcopy(net)
estimation_net._current_time = 0
estimation_net._trace = deque([])
data = Xs[n][m]
spikes = np.zeros((len(data), n_outputs))
pbar = tqdm(total=len(data) * t_image,
unit='Time [s]',
position=1,
desc="Reconstruction")
while estimation_net._current_time < len(X) * t_image:
pbar.n = int(estimation_net._current_time * 1000) / 1000
pbar.update(0)
z = estimation_net.step(lambda t: data[int(
min(t, (len(x) - 1) * t_image) / t_image)],
update_weights=False)
spikes[min(
len(data) - 1, int(estimation_net._current_time /
t_image))] += z.flatten()
reconstructions[:, n*h:(n+1)*h, m*w:(m+1)*w] =\
(
np.dot(spikes, ut.sigmoid(estimation_net._V)) / np.sum(spikes, axis=-1).reshape(-1, 1)
).reshape(-1, h, w)
return reconstructions
def step(self, data_generator_fn, update_weights=True):
# sample isi
isi = -np.log(np.random.uniform()) / self._r_net
inputs = data_generator_fn(self._current_time + isi)
inputs = inputs.reshape((-1, 1))
assert len(inputs) == self._n_inputs, "Input length does not match"
# u = V * input + b
u = np.dot(self._V, inputs) + self._b
z = np.zeros((self._n_outputs, 1))
# p = softmax(u)
p_z = np.exp(u) / np.sum(np.exp(u))
# sample from softmax distribution
sum_p_z = np.cumsum(p_z)
diff = sum_p_z - np.random.uniform(0, 1, 1) > 0
k = np.argmax(diff)
z[k] = 1.0
if update_weights:
self._b += self._eta_b * (isi * self._r_net * self._m_k -
ut.dirac(z - 1))
self._V += self._eta_v * ut.dirac(z - 1) * (inputs.T -
ut.sigmoid(self._V))
self._current_time += isi
self._trace.append(
(self._current_time, inputs, z, u, self._V, self._b))
if len(self._trace) > self._max_trace_length:
self._trace.pop()
return z
def output(self):
output_rgblist = []
# print self.input.shape
if self.isRGB:
for i_rgb in xrange(3):
output_list = []
data_input = self.input[i_rgb]
for i in xrange(data_input.shape[0]):
if i % 100 == 0:
print 'output image:' + str(i), data_input.shape[0]
output_row = []
input_vector = data_input[i]
# print 'cnn output check!!'
# print input_vector.shape
# 入力データの１次元ベクトルを２次元に直す
input = []
for j in xrange(self.prev_shape[1]):
# j 0~51
input.append(input_vector[j*self.prev_shape[0]: (j+1)*self.prev_shape[0]])
input = numpy.array(input)
for y in xrange(self.post_shape[1]):
for x in xrange(self.post_shape[0]):
# x 0~73 74
# y 0~51 46
# input [52,80]
input_dot = input[y*self.filter_shift[0]:y*self.filter_shift[0]+self.filter_shape[1],
x*self.filter_shift[1]:x*self.filter_shift[1]+self.filter_shape[0]]
now_W = self.W
# if i_rgb == 0:
# now_W = self.WR
# if i_rgb == 1:
# now_W = self.WG
# if i_rgb == 2:
# now_W = self.WB
output = [a*b for (a, b) in zip(input_dot, now_W)]
output = numpy.array(output)
output = output.sum() + self.bias
output_possible = sigmoid(output)
output_row.append(output_possible)
output_list.append(output_row)
output_rgblist.append(output_list)
# return numpy.array(output_list)
return numpy.array(output_rgblist)
else:
for i in xrange(self.input.shape[0]):
print 'output image:' + str(i)
output_row = []
input_vector = self.input[i]
# 入力データの１次元ベクトルを２次元に直す
input = []
for j in xrange(self.prev_shape[1]):
# j 0~51
input.append(input_vector[j*self.prev_shape[0]: (j+1)*self.prev_shape[0]])
input = numpy.array(input)
for y in xrange(self.post_shape[1]):
for x in xrange(self.post_shape[0]):
# x 0~73 74
# y 0~51 46
# input [52,80]
# print input.shape
# print x,y
# print self.post_shape
input_dot = input[y*self.filter_shift[0]:y*self.filter_shift[0]+self.filter_shape[1],
x*self.filter_shift[1]:x*self.filter_shift[1]+self.filter_shape[0]]
output = [a*b for (a, b) in zip(input_dot, self.W)]
output = numpy.array(output)
output = output.sum() + self.bias
output_possible = sigmoid(output)
output_row.append(output_possible)
# [Height, Width] を [Width, Height]に変換する？？？不要？
# output_mat = []
# for x in xrange(self.post_shape[0]):
# # output_mat_vec = []
# for y in xrange(self.post_shape[1]):
# # print x,y
# index = y*self.post_shape[1] + x
# # print index
# output_mat.append(output_row[index])
output_list.append(output_row)
# output_list.append(output_mat)
return numpy.array(output_list)
def output_hr(self):
h_list = []
r_list = []
for i in xrange(self.data_size):
# for i in xrange(4):
# input = []
# if i is 0:
# input = self.input_v[0]
# elif i is 1:
# input = self.input_v[19]
# elif i is 2:
# input = self.input_v[39]
# elif i is 3:
# input = self.input_v[19]
input = self.input_v[i]
# r = 0
# if i == 0:
# r = numpy.dot(input, self.W.T) + self.hbias
# else:
# r = numpy.dot(input, self.W.T) + self.hbias + numpy.dot(r_list[i-1], self.U.T)
if i == 0:
r = numpy.dot(input, self.W.T) + self.hbias
h = numpy.dot(input, self.W.T) + self.hbias
r = (r - numpy.min(r)) / numpy.max(r - numpy.min(r)) * 6 - 3
h = (h - numpy.min(h)) / numpy.max(h - numpy.min(h)) * 6 - 3
# print numpy.max(h)
# print numpy.min(h)
else:
# print 'feature_check!!!!'
# print 'r_list[i-1]'
# print r_list[i-1]
# print 'self.U.T'
# print self.U.T
# print 'r_list[i-1]'
# print r_list[i-1]
# print 'self.U.T'
# print self.U.T
tmp_v = numpy.dot(input, self.W.T)
tmp_r = numpy.dot(r_list[i-1], self.U.T)
tmp_v = (tmp_v - numpy.min(tmp_v)) / numpy.max(tmp_v - numpy.min(tmp_v)) * 6 - 3
tmp_r = (tmp_r - numpy.min(tmp_r)) / numpy.max(tmp_r - numpy.min(tmp_r)) * 6 - 3
h = tmp_v + tmp_r
r = tmp_v + tmp_r
# h = tmp_v
# r = tmp_v
f = open('check_vrh.txt', 'a+')
# check_v = numpy.dot(input, self.W.T)
# check_r = numpy.dot(r_list[i-1], self.U.T)
# check_h = numpy.dot(input, self.W.T) + self.hbias + numpy.dot(r_list[i-1], self.U.T)
str_v = ''
str_r = ''
str_h = ''
for i_check in xrange(30):
str_v += str(tmp_v[i_check]) + ','
str_r += str(tmp_r[i_check]) + ','
str_h += str(h[i_check]) + ','
f.write(str_v + '\n')
f.write(str_r + '\n')
f.write(str_h + '\n')
f.write('\n')
f.close()
print numpy.max(h)
print numpy.min(h)
# r = numpy.dot(input, self.W.T) + self.hbias + numpy.dot(r_list[i-1], self.U.T)
# h = numpy.dot(input, self.W.T) + self.hbias + numpy.dot(r_list[i-1], self.U.T)
r = sigmoid(r)
h = sigmoid(h)
h_list.append(h)
r_list.append(r)
h_list = numpy.array(h_list)
r_list = numpy.array(r_list)
# print 'output_hr check'
# print numpy.max(self.input_v)
# print numpy.min(self.input_v)
# print numpy.average(self.input_v)
# print numpy.max(self.W)
# print numpy.min(self.W)
# print numpy.average(self.W)
# print numpy.max(h_list)
# print numpy.min(h_list)
# print numpy.average(h_list)
return h_list, r_list
def reconstruct_from_output(self, output):
input = numpy.dot(output, self.W) + self.vbias
input_possible = sigmoid(input)
# assert self.input.shape == input_possible.shape
return input_possible
def output_from_input(self, input):
output = numpy.dot(input, self.W.T) + self.hbias
hidden_possible = sigmoid(output)
return hidden_possible
def contrast_divergence(self, epoch):
v_list = self.input_v
h_list = []
r_list = []
d_list = []
v_iteration_list = []
h_iteration_list = []
# t=0,,,Tまで隠れ層の出力hとリカレントの出力rを計算する
for i in xrange(self.data_size):
input = v_list[i]
r = 0
h = 0
if i == 0:
r = numpy.dot(input, self.W.T) + self.hbias
h = numpy.dot(input, self.W.T) + self.hbias
else:
r = numpy.dot(input, self.W.T) + self.hbias + numpy.dot(r_list[i-1], self.U.T)
h = numpy.dot(input, self.W.T) + self.hbias + numpy.dot(r_list[i-1], self.U.T)
r = sigmoid(r)
# h = numpy.dot(input, self.W.T) + self.hbias
h = sigmoid(h)
h_list.append(h)
r_list.append(r)
# CD iterationはとりあえず１で試す
v_iteration = numpy.dot(h, self.W) + self.vbias
v_iteration = sigmoid(v_iteration)
h_iteration = 0
if i == 0:
h_iteration = numpy.dot(v_iteration, self.W.T) + self.hbias
else:
h_iteration = numpy.dot(v_iteration, self.W.T) + self.hbias + numpy.dot(r_list[i-1], self.U.T)
h_iteration = sigmoid(h_iteration)
v_iteration_list.append(v_iteration)
h_iteration_list.append(h_iteration)
# print numpy.array(v_iteration_list).shape
# print numpy.array(h_iteration_list).shape
v_list = numpy.array(v_list)
h_list = numpy.array(h_list)
r_list = numpy.array(r_list)
d_list = numpy.array(d_list)
v_iteration_list = numpy.array(v_iteration_list)
h_iteration_list = numpy.array(h_iteration_list)
d_reverse_list = []
for i in reversed(xrange(self.data_size)):
h_diff = h_list[i] - h_iteration_list[i]
d = 0
if i == self.data_size - 1:
d = numpy.dot(self.U, h_diff)
else:
d = numpy.dot(self.U, d_reverse_list[-1] * r_list[i] * (1 - r_list[i]) + h_diff)
d_reverse_list.append(d)
d_reverse_list = numpy.array(d_reverse_list)
d_list = []
for i in xrange(len(d_reverse_list)):
d_list.append(d_reverse_list[len(d_reverse_list) - i - 1])
d_list = numpy.array(d_list)
# calculate W H
delta_H_W = [numpy.dot(h[numpy.newaxis, :].T, v[numpy.newaxis, :]) \
- numpy.dot(ha[numpy.newaxis, :].T, va[numpy.newaxis, :]) \
for v,h,va,ha \
in zip(v_list, h_list, v_iteration_list, h_iteration_list)]
delta_H_W = numpy.array(delta_H_W)
delta_H_W = numpy.average(delta_H_W, axis=0)
# calculate W Q2
# delta_Q2_W = [numpy.dot(d*r*(1-r)[numpy.newaxis, :].T, v[numpy.newaxis, :]) \
# for d,r,v \
# in zip(d_list, r_list, v_list)]
_delta_Q2_W = [d*r*(1-r) \
for d,r,v \
in zip(d_list, r_list, v_list)]
_delta_Q2_W = numpy.array(_delta_Q2_W)
delta_Q2_W = [numpy.dot(dr[numpy.newaxis, :].T, v[numpy.newaxis, :]) \
for dr,v \
in zip(_delta_Q2_W, v_list)]
delta_Q2_W = numpy.average(delta_Q2_W, axis=0)
# calculate W delta
# delta_W = delta_H_W + delta_Q2_W
delta_W = delta_H_W
# calculate U delta
_delta_Q2_U = _delta_Q2_W + (h_list - h_iteration_list)
delta_Q2_U = numpy.dot(_delta_Q2_U.T, r_list)
# delta_Q2_U = numpy.average(delta_Q2_U, axis=0)
# 上の式，Uは(output_size, output_size)なので実行できても意味が通ってるか確認すべき
# !!! check d_list
# print numpy.max(d_list)
# print numpy.min(d_list)
# print numpy.average(d_list)
delta_U = delta_Q2_U
# calculate vbias
delta_vbias = v_list - v_iteration_list
delta_vbias = numpy.average(delta_vbias, axis=0)
# calculate hbias
delta_hbias = h_list - h_iteration_list + _delta_Q2_W
delta_hbias = numpy.average(delta_hbias, axis=0)
####################
# gradient check
####################
f = open('gradient_check_U_10000.txt', 'a+')
f.write(str(numpy.sum(numpy.fabs(delta_U))) + '\n')
print numpy.sum(numpy.fabs(delta_U))
f.close()
####################
# parameter update
####################
# print 'delta_U'
# print delta_U
# print 'U'
# print self.U
self.W += self.lr * delta_W
self.U += self.lr * delta_U - 0.01 * self.U
def contrast_divergence_eachdata(self, epoch):
# print 'input.shape[0] : '+ str(self.input.shape[0])
total_delta = numpy.zeros((self.output_size, self.input_size))
for i in xrange(self.input.shape[0]):
train_input = self.input[i]
dw = numpy.zeros((self.output_size, self.input_size))
dvb = numpy.zeros(self.input_size)
dhb = numpy.zeros(self.output_size)
output = numpy.dot(train_input, self.W.T) + self.hbias
hidden_possible = sigmoid(output)
# hidden_state = numpy_rng.binomial(n=1, p=hidden_possible)
delta_W = []
for j in xrange(hidden_possible.shape[0]):
# j: 0~340
delta_W_elem = train_input * hidden_possible[j]
delta_W.append(delta_W_elem)
delta_W = numpy.array(delta_W)
dw += delta_W
dvb += numpy.mean(train_input, axis=0)
dhb += numpy.mean(hidden_possible, axis=0)
visible_output = numpy.dot(hidden_possible, self.W) + self.vbias
visible_possible = sigmoid(visible_output)
hidden_output = numpy.dot(visible_possible, self.W.T) + self.hbias
hidden_possible_after = sigmoid(hidden_output)
delta_W = []
for j in xrange(hidden_possible_after.shape[0]):
# j: 0~340
delta_W_elem = visible_possible * hidden_possible_after[j]
delta_W.append(delta_W_elem)
delta_W = numpy.array(delta_W)
dw -= delta_W
dvb -= numpy.mean(visible_possible, axis=0)
dhb -= numpy.mean(hidden_possible_after, axis=0)
# data_sizeで割るのは結局必要なのか？
# self.W += self.learning_rate * dw
total_delta += dw
error = numpy.sum(numpy.abs(dw))
print error
os.chdir('result/rbm1_train' + str(epoch))
f = open('error.txt', 'a')
f.write(str(error) + ',')
f.close()
os.chdir('../../')
# total_delta: -700~700 / 7000
self.W += self.learning_rate * total_delta / self.data_size
os.chdir('result/rbm1_train' + str(epoch))
f = open('error.txt', 'a')
f.write('\n')
f.close()
os.chdir('../../')
def predict_sigmoid(self, input):
output = numpy.dot(input, self.W.T) + self.b
hidden_possible = sigmoid(output)
return hidden_possible
def pre_train(self):
for ep in xrange(self.epoch):
print 'pretrain epoch:' + str(ep+1)
loss = 0.0
if self.isRGB:
for i_rgb in xrange(3):
data_input = self.input[i_rgb]
for i in xrange(data_input.shape[0]):
input_vector = data_input[i]
# 入力データの１次元ベクトルを２次元に直す
input = []
for j in xrange(self.prev_shape[1]):
# j 0~51
input.append(input_vector[j*self.prev_shape[0]: (j+1)*self.prev_shape[0]])
input = numpy.array(input)
for y in xrange(self.post_shape[1]):
for x in xrange(self.post_shape[0]):
# print input.shape
# print x,y
# print x+self.filter_shape[0], y+self.filter_shape[1]
input_dot = input[y*self.filter_shift[0]:y*self.filter_shift[0]+self.filter_shape[1],
x*self.filter_shift[1]:x*self.filter_shift[1]+self.filter_shape[0]]
now_W = self.W
# if i_rgb == 0:
# now_W = self.WR
# if i_rgb == 1:
# now_W = self.WG
# if i_rgb == 2:
# now_W = self.WB
output = [a*b for (a, b) in zip(input_dot, now_W)]
output = numpy.array(output)
output = output.sum() + self.bias
output_possible = sigmoid(output)
# 0,1の２値にする必要があるのかは不明
# output_state = numpy_rng.binomial(n=1, p=output_possible)
# print output_state
visible = output_possible * now_W
visible_possible = sigmoid(visible)
hidden_output = [a*b for (a, b) in zip(visible_possible, now_W)]
hidden_output = numpy.array(hidden_output)
hidden_output = hidden_output.sum() + self.bias
hidden_possible = sigmoid(hidden_output)
dw = numpy.zeros(self.filter_shape)
# print x,y
# print input_dot.shape
# print visible_possible.shape
dw = input_dot*output_possible - visible_possible*hidden_possible
loss += numpy.average(dw * dw)
# if self.isRGB:
# if i_rgb == 0:
# self.WR += self.lr * dw
# elif i_rgb == 1:
# self.WG += self.lr * dw
# elif i_rgb == 2:
# self.WB += self.lr * dw
# else:
# self.W += self.lr * dw
self.W += self.lr * dw
print loss
else:
for i in xrange(self.input.shape[0]):
input_vector = self.input[i]
# 入力データの１次元ベクトルを２次元に直す
input = []
for j in xrange(self.prev_shape[1]):
# j 0~51
input.append(input_vector[j*self.prev_shape[0]: (j+1)*self.prev_shape[0]])
input = numpy.array(input)
for y in xrange(self.post_shape[1]):
for x in xrange(self.post_shape[0]):
# print input.shape
# print x,y
# print x+self.filter_shape[0], y+self.filter_shape[1]
input_dot = input[y*self.filter_shift[0]:y*self.filter_shift[0]+self.filter_shape[1],
x*self.filter_shift[1]:x*self.filter_shift[1]+self.filter_shape[0]]
output = [a*b for (a, b) in zip(input_dot, self.W)]
output = numpy.array(output)
output = output.sum() + self.bias
output_possible = sigmoid(output)
# 0,1の２値にする必要があるのかは不明
# output_state = numpy_rng.binomial(n=1, p=output_possible)
# print output_state
visible = output_possible * self.W
visible_possible = sigmoid(visible)
print 'check!!'
print visible_possible.shape
print self.W
hidden_output = [a*b for (a, b) in zip(visible_possible, self.W)]
hidden_output = numpy.array(hidden_output)
hidden_output = hidden_output.sum() + self.bias
hidden_possible = sigmoid(hidden_output)
dw = numpy.zeros(self.filter_shape)
# print x,y
# print input_dot.shape
# print visible_possible.shape
dw = input_dot*output_possible - visible_possible*hidden_possible
self.W += self.lr * dw
