def train(x1, rbm, learning_rate = 0.1):
rbm.units[rbm.visible] = x1
mask = rbm.hidden
Qh1 = sigmoid(rbm.gaps[mask])
h1 = Qh1 > numpy.random.random(len(Qh1))
rbm.units[mask] = h1.copy()
mask = rbm.visible
Px2 = sigmoid(rbm.gaps[mask])
x2 = Px2 > numpy.random.random(len(Px2))
rbm.units[mask] = x2.copy()
mask = rbm.hidden
Qh2 = sigmoid(rbm.gaps[mask])
h2 = Qh2 > numpy.random.random(len(Qh2))
rbm.units[mask] = h2.copy()
adj = (numpy.outer(h1,x1) - numpy.outer(Qh2,x2))
for i,wi in enumerate(rbm.visible):
for j,wj in enumerate(rbm.hidden):
if rbm.weights[wi,wj] == None: continue
rbm.weights[wi,wj] = rbm.weights[wi,wj] + learning_rate * adj[j,i]
rbm.biases[rbm.visible] += learning_rate * (x1 - x2)
rbm.biases[rbm.hidden] += learning_rate * (h1 - Qh2)
print(numpy.linalg.norm(x2-x1,1))
print(numpy.linalg.norm(h1-Qh2,1))
print(numpy.linalg.norm(adj,1))
def learn(self, data, num = 1, rate = 0.01):
"""
"""
if num <= 0: raise ValueError('Need at least one reconstruction')
nVisible = len(self._model.visible)
nHidden = len(self._model.hidden)
self._model.units[self._model.visible] = data
prob_hidden = sigmoid(self._model.gaps[self._model.hidden])
self._model.units[self._model.hidden] = prob_hidden > numpy.random.random(nHidden)
vh_data = self._model.units.copy()
for i in range(num):
prob_visible = sigmoid(self._model.gaps[self._model.visible])
self._model.units[self._model.visible] = prob_visible > numpy.random.random(nVisible)
prob_hidden = sigmoid(self._model.gaps[self._model.hidden])
self._model.units[self._model.hidden] = prob_hidden > numpy.random.random(nHidden)
vh_model = self._model.units.copy()
for v in self._model.visible:
for h in self._model.hidden:
dw = vh_data[v] * vh_data[h] - vh_model[v] * vh_model[h]
self._model.weights[v,h] = self._model.weights[v,h] + rate * dw
for v in self._model.visible:
da = vh_data[v] - vh_data[v]
self._model.biases[v] = self._model.biases[v] + rate * da
for h in self._model.hidden:
da = vh_data[h] - vh_data[h]
self._model.biases[h] = self._model.biases[h] + rate * da
def negSamplingCostAndGradient(predicted,
target,
outputVectors,
dataset,
K=10):
""" Negative sampling cost function for word2vec models
Implement the cost and gradients for one predicted word vector
and one target word vector as a building block for word2vec
models, using the negative sampling technique. K is the sample
size.
Note: See test_word2vec below for dataset's initialization.
Arguments/Return Specifications: same as softmaxCostAndGradient
"""
indices = [target]
indices.extend(getNegativeSamples(target, dataset, K))
grad = np.zeros(outputVectors.shape)
gradPred = np.zeros(predicted.shape)
cost = 0
z = sigmoid(np.dot(outputVectors[target], predicted))
cost -= np.log(z)
grad[target] += predicted * (z - 1.0)
gradPred += outputVectors[target] * (z - 1.0)
for k in range(K):
samp = indices[k + 1]
z = sigmoid(np.dot(outputVectors[samp], predicted))
cost -= np.log(1.0 - z)
grad[samp] += predicted * z
gradPred += outputVectors[samp] * z
return cost, gradPred, grad
def gradient(self, x, t):
# 请从参数字典获取网络参数
w1, b1 = self.params['W1'], self.params['b1']
w2, b2 = self.params['W2'], self.params['b2']
# 保存梯度结果
grads = {}
# forward
a1 = np.dot(x, w1) + b1
h1 = sigmoid(a1)
a2 = np.dot(h1, w2) + b2
output = softmax(a2)
# backward
dy = (output - t) / x.shape[0]
grads['W2'] = np.dot(h1.T, dy)
grads['b2'] = np.sum(dy, axis=0)
"""
grads['b2'] = np.sum(dy, axis=0),为什么求和？
- 首先输出为多少维度，那么b就是多少维的向量，和样本数量无关
因为正向传播过程中，偏置b向量会分别加到每一个样本数据上，因此只需把这些值加起来就好
也就是说：第一个样本产生由于b产生误差 dy1
第二个样本产生由于b产生误差 dy2
...
b产生的总误差为: dy1 + dy2 + ...
"""
da1 = np.dot(dy, w2.T)
ha1 = sigmoid(a1)
dz1 = (1.0 - ha1) * ha1 * da1
grads['W1'] = np.dot(x.T, dz1)
grads['b1'] = np.sum(dz1, axis=0)
return grads
def train(x1, rbm, learning_rate=0.1):
rbm.units[rbm.visible] = x1
mask = rbm.hidden
Qh1 = sigmoid(rbm.gaps[mask])
h1 = Qh1 > numpy.random.random(len(Qh1))
rbm.units[mask] = h1.copy()
mask = rbm.visible
Px2 = sigmoid(rbm.gaps[mask])
x2 = Px2 > numpy.random.random(len(Px2))
rbm.units[mask] = x2.copy()
mask = rbm.hidden
Qh2 = sigmoid(rbm.gaps[mask])
h2 = Qh2 > numpy.random.random(len(Qh2))
rbm.units[mask] = h2.copy()
adj = (numpy.outer(h1, x1) - numpy.outer(Qh2, x2))
for i, wi in enumerate(rbm.visible):
for j, wj in enumerate(rbm.hidden):
if rbm.weights[wi, wj] == None: continue
rbm.weights[wi,
wj] = rbm.weights[wi, wj] + learning_rate * adj[j, i]
rbm.biases[rbm.visible] += learning_rate * (x1 - x2)
rbm.biases[rbm.hidden] += learning_rate * (h1 - Qh2)
print(numpy.linalg.norm(x2 - x1, 1))
print(numpy.linalg.norm(h1 - Qh2, 1))
print(numpy.linalg.norm(adj, 1))
def forward_pass(self, inputs):
# decleare variables used forward pass
self.inputs = inputs
self.n_inp = len(inputs)
self.vr = []
self.vz = []
self.v_h = []
self.vo = []
self.r = []
self.z = []
self._h = []
self.h = {}
self.o = []
self.h[-1] = np.zeros((self.h_size, 1))
# performing recurrsion
for i in range(self.n_inp):
# calculating reset gate value
# self.vr.append(np.dot(self.w['ur'],inputs[i]) + np.dot(self.w['wr'], self.h[i-1]) + self.b['r'])
# self.r.append(sigmoid(self.vr[i]))
self.r.append(
sigmoid(
np.dot(self.w['ur'], inputs[i]) +
np.dot(self.w['wr'], self.h[i - 1]) + self.b['r']))
# calculation update gate value
# self.vz.append(np.dot(self.w['uz'],inputs[i]) + np.dot(self.w['wz'], self.h[i-1]) + self.b['z'])
# self.z.append(sigmoid(self.vz[i]))
self.z.append(
sigmoid(
np.dot(self.w['uz'], inputs[i]) +
np.dot(self.w['wz'], self.h[i - 1]) + self.b['z']))
# applying reset gate value
# self.v_h.append(np.dot(self.w['u_h'], inputs[i]) + np.dot(self.w['w_h'], np.multiply(self.h[i - 1], self.r[i])) + + self.b['_h'])
# self._h.append(tanh(self.v_h[i]))
self._h.append(
tanh(
np.dot(self.w['u_h'], inputs[i]) +
np.dot(self.w['w_h'], np.multiply(self.h[i -
1], self.r[i])) +
+self.b['_h']))
# applying update gate value
self.h[i] = np.multiply(self.z[i], self.h[i - 1]) + np.multiply(
1 - self.z[i], self._h[i])
# calculating output
# self.vo.append(np.dot(self.w['wo'], self.h[i]) + self.b['o'])
# self.o.append(softmax(self.vo[i]))
self.o.append(
softmax(np.dot(self.w['wo'], self.h[i]) + self.b['o']))
return self.o
def test(self, test_data, label):
n_pos = 0.0
for i in range(len(test_data)):
if sigmoid(self.W, test_data[i]) >= 0.5 and label[i] == 1:
n_pos += 1
elif sigmoid(self.W, test_data[i]) < 0.5 and label[i] == 0:
n_pos += 1
print n_pos, len(test_data), n_pos / len(test_data)
print self.W
def forward(self, inputs, return_activations=False):
# inputs specifies one float for each neuron in the first layer
# if return_activations is true, we return a list of activations at each layer
# otherwise, we only return the output layer
values = numpy.array([inputs], float).T # values in current layer
activations = [{'activations': values}]
for layer in self.layers[1:]:
if layer['type'] == 'sigmoid':
# compute the weighted sum of neuron inputs, plus bias term (for each neuron in current layer)
z_vector = numpy.dot(layer['weights'], values) + layer['bias']
# apply sigmoid activation function
values = util.sigmoid(z_vector)
if return_activations:
activations.append({'activations': values[:, 0]})
elif layer['type'] == 'softmax':
# compute the weighted sum of neuron inputs, plus bias term (for each neuron in current layer)
z_vector = numpy.dot(layer['weights'], values) + layer['bias']
# apply softmax
values = numpy.exp(z_vector - numpy.max(z_vector))
values = values / numpy.sum(values)
if return_activations:
activations.append({'activations': values[:, 0]})
elif layer['type'] == 'convsample':
# carry out convolution to get convolved values
convolved_out = numpy.zeros(
(layer['k'], layer['conv_count'], layer['conv_count']))
for k in xrange(len(layer['weights'])):
convolved_out[k] = scipy.signal.convolve2d(
values.reshape(layer['m'], layer['m']),
numpy.rot90(layer['weights'][k], 2), 'valid')
convolved_out[k] = util.sigmoid(convolved_out[k] +
layer['bias'][k])
# pool the convolved features
pooled = numpy.zeros((layer['k'], layer['o'], layer['o']))
for k in xrange(layer['k']):
for i in xrange(layer['o']):
for j in xrange(layer['o']):
pooled[k][i][j] = numpy.average(
convolved_out[k, (i * layer['p']):((i + 1) *
layer['p']),
(j * layer['p']):((j + 1) *
layer['p'])])
values = pooled.reshape(util.prod(pooled.shape), 1)
if return_activations:
activations.append({
'activations': values[:, 0],
'extra': convolved_out
})
if return_activations: return activations
else: return values[:, 0]
def feedForward(self):
for n in self.hiddenLayer:
sum = n.bias
for ni in self.inputLayer:
sum += self.weights[(ni, n)] * ni.value
n.value = util.sigmoid(sum)
for n in self.outputLayer:
sum = n.bias
for nh in self.hiddenLayer:
sum += self.weights[(nh, n)] * nh.value
n.value = util.sigmoid(sum)
def getOutput(self, inputs):
a_1 = [[
util.sigmoid(x)
for x in util.add(util.dot(inputs, self.ihWeights), self.ihBias)[0]
]]
#a_2 = [[util.sigmoid(x) for x in util.add(util.dot(a_1, self.hhWeights), self.hhBias)[0]]] # uncomment and fix a_3 for 2 hidden layers
a_3 = [[
util.sigmoid(x)
for x in util.add(util.dot(a_1, self.hoWeights), self.hoBias)[0]
]]
return a_3[0]
def tree_lstm(input, g_in_left, g_in_right, r_in_left, r_in_right):
g_in = 1 / 2 * (g_in_left + g_in_right)
r_in = 1 / 2 * (r_in_left + r_in_right)
f_t = sigmoid(concat_and_multiply(params['w_forget'], r_in, input))
k_t_1 = sigmoid(concat_and_multiply(params['w_k_1'], r_in, input))
r_t = np.tanh(concat_and_multiply(params['w_r'], r_in, input))
k_t_2 = sigmoid(concat_and_multiply(params['w_k_2'], r_in, input))
g_out = f_t * g_in + k_t_1 * r_t
r_out = k_t_2 * np.tanh(g_out)
return g_out, r_out
def update_backward_lstm(_input, hiddens, cells):
change = np.tanh(
concat_and_multiply(params['change_b'], _input, hiddens))
forget = sigmoid(
concat_and_multiply(params['forget_b'], _input, hiddens))
ingate = sigmoid(
concat_and_multiply(params['ingate_b'], _input, hiddens))
outgate = sigmoid(
concat_and_multiply(params['outgate_b'], _input, hiddens))
cells = cells * forget + ingate * change
hiddens = outgate * np.tanh(cells)
return hiddens, cells
def fitness_notequal_proba(self, x, x1):
feature_similarity_score = 1.0 - cdist(
x.reshape(1, -1), x1.reshape(1, -1), metric=self.metric).ravel()[0]
feature_similarity = sigmoid(feature_similarity_score)
y = self.bb_predict_proba(x.reshape(1, -1))[0]
y1 = self.bb_predict_proba(x1.reshape(1, -1))[0]
# target_similarity_score = np.sum(np.abs(y - y1))
target_similarity_score = 1.0 - cosine(y, y1)
target_similarity = 1.0 - sigmoid(target_similarity_score)
evaluation = self.alpha1 * feature_similarity + self.alpha2 * target_similarity
return evaluation,
def fitness_notequal(self, x, x1):
feature_similarity_score = 1.0 - cdist(
x.reshape(1, -1), x1.reshape(1, -1), metric=self.metric).ravel()[0]
# feature_similarity = feature_similarity_score if feature_similarity_score >= self.eta1 else 0.0
feature_similarity = sigmoid(feature_similarity_score)
y = self.bb_predict(x.reshape(1, -1))[0]
y1 = self.bb_predict(x1.reshape(1, -1))[0]
target_similarity_score = 1.0 - hamming(y, y1)
# target_similarity = target_similarity_score if target_similarity_score < self.eta2 else 0.0
target_similarity = 1.0 - sigmoid(target_similarity_score)
evaluation = self.alpha1 * feature_similarity + self.alpha2 * target_similarity
return evaluation,
def _backpropagation(self, x, y_true):
deltas = list()
activations = [x]
for w, b in zip(self._weights, self._biases):
a = sigmoid(w.dot(activations[-1]) + b)
activations.append(a)
output_delta = (activations[-1] -
y_true) * activations[-1] * (1 - activations[-1])
deltas.append(output_delta)
for index in range(2, len(activations)):
weight_index = 1 - index # = - (index - 1)
delta_index = index - 2
tmp1 = self._weights[weight_index].T.dot(deltas[delta_index])
tmp2 = activations[-index] * (1 - activations[-index])
new_delta = tmp1 * tmp2
deltas.append(new_delta)
deltas = deltas[::-1]
w_gradients = list()
b_gradients = list()
for index, delta in enumerate(deltas):
grad_b = delta
grad_w = np.outer(delta, activations[index])
w_gradients.append(grad_w)
b_gradients.append(grad_b)
return w_gradients, b_gradients
def forward(self):
temp = np.vstack((np.ones((1, self._input.shape[1])), self._input))
self._forward_cache_acted = [temp]
self._forward_cache_raw = [temp]
if not self._constructed:
print("use the build method before forwarding.")
assert 0
times = len(self._layers) - 1
for i in range(times):
# temp = np.vstack((np.ones((1, self._input.shape[1])), temp))
temp = np.dot(self._weights[i], temp)
self._forward_cache_raw.append(temp)
if not self._activations[i]:
pass
elif self._activations[i].lower() == 'sigmoid':
temp = util.sigmoid(temp)
elif self._activations[i].lower() == 'tanh':
temp = util.tanh(temp)
elif self._activations[i].lower() == 'relu':
temp = util.relu(temp)
else:
print(
"Activation function should be None, 'sigmoid', 'tanh' or 'relu'."
)
assert 0
self._forward_cache_acted.append(temp)
self._predictions = temp
return temp
def loss(A, b, x):
result = [A]
layer_in = A
for layer in x:
layer_out = util.sigmoid(np.dot(layer_in, layer))
layer_in = layer_out
return np.sum(np.linalg.norm(layer_out - b, axis=1))
def target(self, x, t, x0):
""" A directed sigmoid function target.
Parameters
----------
x: state
t: time
x0: initial state
Returns
-------
(x, u, a)
"""
pose0 = self.model.model_parameters(q=x0[:self.model.nd])
xt0 = self.task.x(pose0)
t0 = 1.0
A = 0.3
B = 4
o = np.asmatrix([[0], [0]])
xd, ud, ad = sigmoid(t, t0, A, B)
xd = rotate(o, np.asmatrix([[xd], [0]]), self.angle)
ud = rotate(o, np.asmatrix([[ud], [0]]), self.angle)
ad = rotate(o, np.asmatrix([[ad], [0]]), self.angle)
return (np.asmatrix(xt0 + xd),
np.asmatrix(ud),
np.asmatrix(ad))
def linear_activation_forward(A_prev, W, b, activation):
"""
Implement the forward propagation for the LINEAR->ACTIVATION layer
Arguments:
A_prev -- activations from previous layer (or input data): (size of previous layer, number of examples)
W -- weights matrix: numpy array of shape (size of current layer, size of previous layer)
b -- bias vector, numpy array of shape (size of the current layer, 1)
activation -- the activation to be used in this layer, stored as a text string: "sigmoid" or "relu"
Returns:
A -- the output of the activation function, also called the post-activation value
cache -- a python dictionary containing "linear_cache" and "activation_cache";
stored for computing the backward pass efficiently
"""
if activation == "sigmoid":
# Inputs: "A_prev, W, b". Outputs: "A, activation_cache".
Z, linear_cache = linear_forward(A_prev, W, b)
A, activation_cache = sigmoid(Z)
elif activation == "relu":
# Inputs: "A_prev, W, b". Outputs: "A, activation_cache".
Z, linear_cache = linear_forward(A_prev, W, b)
A, activation_cache = relu(Z)
assert (A.shape == (W.shape[0], A_prev.shape[1]))
cache = (linear_cache, activation_cache)
return A, cache
def backprop(self, x, y):
# Changes in weights and biases
deltaWeights = [np.zeros(w.shape) for w in self.weights]
deltaBiases = [np.zeros(b.shape) for b in self.biases]
# feed forward
params = zip(self.weights, self.biases)
activation = x
activations = [x]
zs = []
for w, b in params:
z = np.matmul(w, activation) + b
zs.append(z)
activation = util.sigmoid(z)
activations.append(activation)
# backprop
# delta is dc/da * da/dz = dc/dz for last layer
# because dz/db = 1 delta is also equal to dc/db
delta = (activations[-1] - y) * util.sigmoid_derivative(zs[-1])
deltaBiases[-1] = delta
deltaWeights[-1] = np.matmul(delta, activations[-2].transpose())
for layer in xrange(2, len(self.layerSizes)):
z = zs[-layer]
sp = util.sigmoid_derivative(z)
delta = np.matmul(self.weights[-layer + 1].transpose(), delta) * sp
deltaBiases[-layer] = delta
deltaWeights[-layer] = np.matmul(delta, activations[-layer - 1].transpose())
return (deltaWeights, deltaBiases, activations[-1])
def forwards(A, x):
result = [A]
layer_in = A
for layer in x:
layer_out = util.sigmoid(np.dot(layer_in, layer))
result += [layer_out]
layer_in = layer_out
return result
def forward(self, X):
Z = X.dot(self.W1) + self.b1
Z = relu(Z)
A = Z.dot(self.W2) + self.b2
A = sigmoid(A)
return A, Z
def train(self):
self.W = np.random.rand(self.n_feature + 1)
for iter in range(self.n_iter):
g = [0.0 for i in range(len(self.W))]
for index, d in enumerate(self.data):
mu = sigmoid(self.W, d)
g = [ x + y for x, y in zip([(mu - label[index]) * z for z in data[index]], g) ]
self.W = [x - self.eta * y for x, y in zip(self.W, g)]
def forward(self, x_t):
self.t += 1
t = self.t
h = self.h[t-1]
self.input_gate[t] = sigmoid(np.dot(self.W_hi, h) + np.dot(self.W_xi, x_t) + self.b_i)
self.forget_gate[t] = sigmoid(np.dot(self.W_hf, h) + np.dot(self.W_xf, x_t) + self.b_f)
self.output_gate[t] = sigmoid(np.dot(self.W_ho, h) + np.dot(self.W_xo, x_t) + self.b_o)
self.cell_update[t] = tanh(np.dot(self.W_hj, h) + np.dot(self.W_xj, x_t) + self.b_j)
self.c[t] = self.input_gate[t] * self.cell_update[t] + self.forget_gate[t] * self.c[t-1]
self.ct[t] = tanh(self.c[t])
self.h[t] = self.output_gate[t] * self.ct[t]
self.x[t] = x_t
return self.h[t]
def forward(self,X):
#Z = relu(X.dot(self.W1)+self.b1)
Z = tanh(X.dot(self.W1)+self.b1)
# print("Z.shape"+str(Z.shape))
# print("self.W2.shape"+str(self.W2.shape))
# print("self.b2.shape"+str(self.b2.shape))
ret = sigmoid(Z.dot(self.W2)+self.b2)
# print("ret.shape"+str(np.array(ret).shape))
return ret, Z
def sample_v_given_h(h, W, b):
"""
:param v: 2d np.ndarray, (N, n_visible)
:param W: 2d np.nadarray, (n_visible, n_hidden)
:param b: 1d np.ndarray, (n_visible, )
:return:
"""
proba = sigmoid(np.matmul(h, W.transpose()) + b)
return sample_binomial(proba)
def feedforward(self):
for i in xrange(self.num_layers + 1):
prev_layer = self.layers[i]
for neuron in self.layers[i + 1]:
sum = neuron.bias
for prev_neuron in prev_layer:
sum += self.weights[(prev_neuron,
neuron)] * prev_neuron.value
neuron.value = util.sigmoid(sum)
def sample_h_given_v(v, W, c):
"""
:param v: 2d np.ndarray, (N, n_visible)
:param W: 2d np.nadarray, (n_visible, n_hidden)
:param c: 1d np.ndarray, (n_hidden, )
:return:
"""
proba = sigmoid(np.matmul(v, W) + c)
return sample_binomial(proba)
def _forward_pass(self, x):
if len(x) != self._architecture[0]:
raise ValueError(
"Input dimensions don't correspond to input layer size")
values = x
for weights, bias in zip(self._weights, self._biases):
z = weights.dot(values) + bias
values = sigmoid(z)
return values
def logistic_graddesc(X, Y, T=1000, learning_rate=10e-2):
#Gradient descent for logistic regression ( analytic method )
#It garantees convergence even when a subset of columns of X are multiples
# https://deeplearningcourses.com/c/data-science-linear-regression-in-python
# https://www.udemy.com/data-science-linear-regression-in-python
w = np.random.randn(X.shape[1])
Yh = sigmoid(X.dot(w))
xe = []
for _ in xrange(T):
delta = Y - Yh
# gradient descent weight update
w += learning_rate * X.T.dot(delta)
# recalculate Y
Yh = sigmoid(X.dot(w))
xe.append(xentropy(Y, Yh))
return w, xe
def getFeatures(self, state, action):
features = util.Counter()
dir_vec = []
if action == util.Dirs.LEFT:
dir_vec = [0, -1]
elif action == util.Dirs.RIGHT:
dir_vec = [0, 1]
elif action == util.Dirs.UP:
dir_vec = [-1, 0]
elif action == util.Dirs.DOWN:
dir_vec = [1, 0]
head = util.vectorAdd(state.snake[0], dir_vec)
#performs BFS of 'search_size' number of positions to see if head is surrounded by walls
search_size = min(pow(max(len(state.snake)/4, 1), 2), int((state.walls[0] * state.walls[1] - len(state.snake)) * 0.75))
remaining_nodes = search_size
oldest_bar = (-1, -1)
oldest_bar_age = len(state.snake)
if head in state.snake or util.outOfBounds(head, state.walls):
remaining_nodes = 0
else:
head_t = (head[0], head[1])
visited_coords = set(head_t)
q = [head_t]
for i in range(0,search_size):
if q:
coord = q.pop(0)
else:
remaining_nodes = i
break
for neighbor in self.getNeighbors(coord):
if state.board[neighbor[0]][neighbor[1]] > 0 and state.board[neighbor[0]][neighbor[1]] < oldest_bar_age:
oldest_bar_age = state.board[neighbor[0]][neighbor[1]]
oldest_bar = neighbor
if util.outOfBounds(neighbor, state.walls) == False and state.board[neighbor[0]][neighbor[1]] < util.manhattanDist(neighbor, head) and neighbor not in visited_coords:
q.append(neighbor)
visited_coords.add(neighbor)
len_bin = int(util.sigmoid(len(state.snake)))
trapped = remaining_nodes < search_size
#Whether or not snake is touching a wall or itself
features["on-barrier"] = state.board[head[0]][head[1]] > 0 or util.outOfBounds(head, state.walls)
#Distance to oldest snake segment; only active if snake is trapped by itself
features["dist-oldest"] = 1.0 / pow(util.euclidDist(head, oldest_bar), 2) if trapped else 0
#Distance to food
features["dist-food"] = pow(util.manhattanDist(head, state.food) / float(state.walls[0] + state.walls[1]), .33)
#Indicated whether current "trapped" state is escapable under perfect circumstances
features["trapped"] = oldest_bar_age > remaining_nodes if trapped else 0
return features
def forward(self, inputs, targets, h_prev, c_prev):
# Forward pass of lstm
cache = defaultdict(lambda : defaultdict(np.float64))
cache['h'][-1] = np.copy(h_prev)
cache['c'][-1] = np.copy(c_prev)
loss = 0
for t in xrange(len(inputs)):
cache['i'][t] = sigmoid(np.dot(self.Wix, inputs[t]) + np.dot(self.Wih, cache['h'][t-1]) + self.bi) # i(t)
cache['o'][t] = sigmoid(np.dot(self.Wox, inputs[t]) + np.dot(self.Woh, cache['h'][t-1]) + self.bo) # o(t)
cache['f'][t] = sigmoid(np.dot(self.Wfx, inputs[t]) + np.dot(self.Wfh, cache['h'][t-1]) + self.bf) # f(t)
cache['g'][t] = np.tanh(np.dot(self.Wgx, inputs[t]) + np.dot(self.Wgh, cache['h'][t-1]) + self.bg) # g(t)
cache['c'][t] = cache['g'][t] * cache['i'][t] + cache['c'][t-1] * cache['f'][t] # c(t)
cache['h'][t] = cache['c'][t] * cache['o'][t] # h(t)
cache['y'][t] = np.dot(self.Why, cache['h'][t]) + self.by # unnormalized log probabilities
cache['p'][t] = np.exp(cache['y'][t]) / np.sum(np.exp(cache['y'][t])) # softmax for prediction
loss += -np.log(cache['p'][t][np.argmax(targets[t])])
return cache, loss
def predict_one_vs_all(all_theta, _X):
m = _X.shape[0]
X = np.hstack((np.ones((m, 1)), _X))
theta0 = all_theta[0]
p = sigmoid(np.dot(X, theta0))
print p.shape
print p
p = np.dot(X, all_theta.T)
return np.array(map(max_index, p))
def forward(self, x):
# 请从参数字典获取网络参数
w1, b1 = self.params['W1'], self.params['b1']
w2, b2 = self.params['W2'], self.params['b2']
# 实现第一层的运算
z1 = np.dot(x, w1) + b1
h1 = sigmoid(z1)
# 请实现第二层的运算
z2 = np.dot(h1, w2) + b2
return softmax(z2)
def predict(self, x):
"""Make a prediction given new inputs x.
Args:
x: Inputs of shape (m, n).
Returns:
Outputs of shape (m,).
"""
# *** START CODE HERE ***
return util.sigmoid(x.dot(self.theta))
def predict_one_vs_all(all_theta, _X):
m = _X.shape[0]
X = np.hstack((np.ones((m, 1)), _X))
theta0 = all_theta[0]
p = sigmoid(np.dot(X, theta0))
print p.shape
print p
p = np.dot(X, all_theta.T)
return np.array(map(max_index, p))
def forward(self, inputs, return_activations = False):
# inputs specifies one float for each neuron in the first layer
# if return_activations is true, we return a list of activations at each layer
# otherwise, we only return the output layer
values = numpy.array([inputs], float).T # values in current layer
activations = [{'activations': values}]
for layer in self.layers[1:]:
if layer['type'] == 'sigmoid':
# compute the weighted sum of neuron inputs, plus bias term (for each neuron in current layer)
z_vector = numpy.dot(layer['weights'], values) + layer['bias']
# apply sigmoid activation function
values = util.sigmoid(z_vector)
if return_activations: activations.append({'activations': values[:, 0]})
elif layer['type'] == 'softmax':
# compute the weighted sum of neuron inputs, plus bias term (for each neuron in current layer)
z_vector = numpy.dot(layer['weights'], values) + layer['bias']
# apply softmax
values = numpy.exp(z_vector - numpy.max(z_vector))
values = values / numpy.sum(values)
if return_activations: activations.append({'activations': values[:, 0]})
elif layer['type'] == 'convsample':
# carry out convolution to get convolved values
convolved_out = numpy.zeros((layer['k'], layer['conv_count'], layer['conv_count']))
for k in xrange(len(layer['weights'])):
convolved_out[k] = scipy.signal.convolve2d(values.reshape(layer['m'], layer['m']), numpy.rot90(layer['weights'][k], 2), 'valid')
convolved_out[k] = util.sigmoid(convolved_out[k] + layer['bias'][k])
# pool the convolved features
pooled = numpy.zeros((layer['k'], layer['o'], layer['o']))
for k in xrange(layer['k']):
for i in xrange(layer['o']):
for j in xrange(layer['o']):
pooled[k][i][j] = numpy.average(convolved_out[k, (i * layer['p']):((i + 1) * layer['p']), (j * layer['p']):((j + 1) * layer['p'])])
values = pooled.reshape(util.prod(pooled.shape), 1)
if return_activations: activations.append({'activations': values[:, 0], 'extra': convolved_out})
if return_activations: return activations
else: return values[:, 0]
def cost_function(theta, X, y, lambda_=0):
m = y.size
o = sigmoid(np.dot(X, theta))
J = (-1./m) * (np.dot(y.T, np.log(o)) + np.dot((1.-y).T, np.log(1.-o))) \
+ (lambda_/(2.*m)) * np.dot(theta[1:], theta[1:])
grad = (1. / m) * np.dot(X.T, o - y)
grad[1:] += (lambda_ / m) * theta[1:]
return J, grad
def forwardPass(self, x):
"""
Given x, computes the outputs of the Network
Parameters
----------
x : example
Returns
-------
nothing
"""
layer = self.input_layer
layer.xs[0] = 1 # bias term
layer.xs[1:] = x
for i in range(self.n_hidden_layers):
next_layer = self.hidden_layers[i]
next_layer.xs[0] = 1 # bias term
next_layer.xs[1:] = sigmoid(np.dot(layer.weights.transpose(), layer.xs))
layer = next_layer
self.output_layer.xs = sigmoid(np.dot(layer.weights.transpose(), layer.xs))
def predict_proba( self, X ): """ Returns the probability of class being 1 @params: X: observations to predict of second axis m """ if self.add_bias: # add the bias term bias_term = np.ones( ( X.shape[0], 1 ) ) X = np.append( bias_term, X, axis = 1 ) return sigmoid( np.dot( X, self.w_estimated ) )
def predict(self, x):
"""Make a prediction given new inputs x.
Args:
x: Inputs of shape (m, n).
Returns:
Outputs of shape (m,).
"""
# *** START CODE HERE ***
temp = util.sigmoid(np.dot(x, self.theta) + self.theta_0)
return temp
def predict_proba(self, X):
"""
Returns the probability of class being 1
@params:
X: observations to predict of second axis m
"""
if self.add_bias:
# add the bias term
bias_term = np.ones((X.shape[0], 1))
X = np.append(bias_term, X, axis=1)
return sigmoid(np.dot(X, self.w_estimated))
def sample_data(self, ss, hps):
"""
NOTE THIS ONLY SAMPLES FROM THE PARAM P and not from
suffstats.
"""
d = np.random.exponential(hps['mu_hp'])
p = util.sigmoid(d, ss['mu'], ss['lambda'])
p = p * (hps['p_max'] - hps['p_min']) + hps['p_min']
link = np.random.rand() < p
x = np.zeros(1, dtype=self.data_dtype())
x[0]['distance'] = d
x[0]['link'] = link
return x[0]
def mlp_decode(z, phi, tanh_scale=10., sigmoid_output=True):
nnet_params, ((W_mu, b_mu), (W_sigma, b_sigma)) = phi[:-2], phi[-2:]
z = z if z.ndim == 3 else z[:,None,:] # ensure z.shape == (T, K, n)
nnet = compose(tanh_layer(W, b) for W, b in nnet_params)
mu = linear_layer(W_mu, b_mu)
log_sigmasq = linear_layer(W_sigma, b_sigma)
nnet_outputs = nnet(np.reshape(z, (-1, z.shape[-1])))
mu = sigmoid(mu(nnet_outputs)) if sigmoid_output else mu(nnet_outputs)
log_sigmasq = tanh_scale * np.tanh(log_sigmasq(nnet_outputs) / tanh_scale)
shape = z.shape[:-1] + (-1,)
return mu.reshape(shape), log_sigmasq.reshape(shape)
def __init__(self, rng, data, W=None, b=None, filter_h=2, filter_num=50, k=300):
"""
:param data: a 3D tensor (sentence number, sentence length, word vector size).
:param W: a matrix (filter_num, word vector size)
:param filter_h: converlution operation window size.
:param filter_num: the feature map number of each converlution window size.
So the total feature maps are `filter_num`, which is
also the size of the new vector representation of the sentence.
"""
if W is None:
W = np.asarray(rng.uniform(size=(filter_num, k * filter_h)),
dtype=theano.config.floatX
)
self.W = theano.shared(value=W, name='W', borrow=True)
# initialize the biases b as a vector of n_out 0s
if b is None:
b = np.asarray(rng.uniform(
size=(filter_num,)),
dtype=theano.config.floatX
)
self.b = theano.shared(value=b, name='b', borrow=True)
X_h, X_w = data.shape[1], data.shape[2]
idx_range = T.arange(X_h - filter_h + 1)
self.window_results, updates = theano.scan(fn=lambda i, X, filter_h: T.flatten(data[:, i: i + filter_h], outdim=2),
sequences=idx_range,
outputs_info=None,
non_sequences=[data, filter_h]
)
self.window_results = T.transpose(self.window_results, axes=(1, 0, 2))
c = sigmoid(T.dot(self.window_results, self.W.T) + self.b)
# max pooling
c_max = T.max(c, axis=1)
self.c = c
# c_max (sentence number, filter_num)
self.c_max = c_max
self.params = [self.W, self.b]
def cross_validation(X_train, y_train, params, X_test=None, verbose_eval=False):
NUM_BOOST_ROUND = 1000
best_iterations = []
train_scores = []
valid_scores = []
y_preds = []
kf = KFold(y_train.shape[0], n_folds=5, shuffle=True, random_state=12345)
for train_index, valid_index in kf:
_X_train, _X_valid = X_train.ix[train_index], X_train.ix[valid_index]
_y_train, _y_valid = y_train[train_index], y_train[valid_index]
dtrain = xgb.DMatrix(_X_train, _y_train)
dvalid = xgb.DMatrix(_X_valid, _y_valid)
watchlist = [(dtrain, 'train'), (dvalid, 'eval')]
bst = xgb.train(params, dtrain, NUM_BOOST_ROUND, evals=watchlist,
early_stopping_rounds=200, verbose_eval=verbose_eval)
# best iterations and valid score
best_iterations.append(bst.best_iteration + 1)
valid_scores.append(bst.best_score)
if X_test is not None:
dtest = xgb.DMatrix(X_test)
y_pred = bst.predict(dtest, ntree_limit=bst.best_iteration)
y_preds.append(y_pred)
y_pred = util.sigmoid(np.mean(util.logit(np.array(y_preds)), axis=0))
result = {"best-iterations": best_iterations,
"best-iteration": np.mean(best_iterations),
"valid-score": np.mean(valid_scores),
"valid-scores": valid_scores,
"y_pred": y_pred,
"y_preds": y_preds}
return result
def gaussian_mean(inputs, sigmoid_mean=False):
mu_input, sigmasq_input = np.split(inputs, 2, axis=-1)
mu = sigmoid(mu_input) if sigmoid_mean else mu_input
sigmasq = log1pexp(sigmasq_input)
return make_tuple(mu, sigmasq)
def forward_propagation(self):
return util.sigmoid(np.dot(self.a, self.theta))
sentence = "the digit 1 is on the left of the digit 0 ."
y, words = sent2matrix(sentence, dictionary)
y = np.int32(np.repeat(y, 100, axis=0))
print y.shape
height = int(math.sqrt(dimX))
width = int(math.sqrt(dimX))
draw.height = height
draw.width = width
rvae = ReccurentAttentionVAE(dimY, dimLangRNN, dimAlign, dimX, dimReadAttent, dimWriteAttent, dimRNNEnc, dimRNNDec, dimZ, runSteps, batch_size, reduceLRAfter, data, labels, pathToWeights=args.weights)
ct_s, write_attent_params, alphas, _, _ = rvae.sample_from_prior(runSteps, y)
ct_s = sigmoid(ct_s)
dimension = int(math.sqrt(dimX))
print dimension, dimension
most_used = np.float32(np.zeros((y.shape[1])))
for i in xrange(runSteps):
total_image = np.zeros((dimension*10,dimension*10))
for j in xrange(100):
c = ct_s[i,j,:].reshape([dimension,dimension])
row = j/10
column = j%10
total_image[(row*dimension):((row+1)*dimension),(column*dimension):((column+1)*dimension)] = c[:][:]
def decode(X):
return sigmoid(mu(nnet(X)))
def forward(X, W, b):
return sigmoid(X.dot(W) + b)
def _sigmoid(self,gain):
for i in xrange(len(gain)):
gain[i]=sigmoid(gain[1],1,2,self._gain)
# Calculate cross-entropy error for random weights, and for closed-form solution to bayes classifier
import numpy as np
from util import sigmoid, cross_entropy
N = 100
D = 2
means = np.array(((-2,-2), (2,2)))
covar = np.eye(2)
# Artifically create 2 classes, center first 50 points at (-2,-2), last 50 at (2,2)
X = np.random.randn(N, D)
X[:N//2, :] = X[:N//2, :] + means[0] * np.ones((N//2,D))
X[N//2:, :] = X[N//2:, :] + means[1] * np.ones((N//2,D))
Xb = np.concatenate((np.ones((N, 1)), X), axis=1)
# Class labels, first 50 are 0, last 50 are 1
T = np.concatenate((np.zeros((N//2,)), np.ones((N//2,))))
# Random weights
w = np.random.randn(D+1)
Y = sigmoid(Xb @ w)
print('Random weights:', cross_entropy(T, Y))
# Closed form Bayes solution
w = ((means[1, None] - means[0, None]) @ np.linalg.inv(covar)).T
w = np.concatenate(((0,), w.reshape(D))) # Add weight for bias
Y = sigmoid(Xb @ w)
print('Closed form solution:', cross_entropy(T, Y))
def forward(self, X):
# Z = relu(X.dot(self.W1) + self.b1)
Z = np.tanh(X.dot(self.W1) + self.b1)
return sigmoid(Z.dot(self.W2) + self.b2), Z
def forward(self, X):
return sigmoid(X.dot(self.W) + self.b)
def value(self, params, ex):
return util.sigmoid(self.score_fxn.value(params, ex))
def _updateOnMask(self, mask):
self._network.units[mask] = sigmoid(self._network.gaps[mask]) > numpy.random.random(len(mask))
