def grad_loss(self, *args):
"""
Compute the gradient logistic loss function
Inputs:
- X: N x D array of data; each row is a data point.
- y: 1-dimensional array of length N with real values.
- reg: (float) regularization strength.
Returns: A tuple containing:
- loss as a single float
- gradient with respect to self.theta; an array of the same shape as theta
"""
theta,X,y,reg = args
m,dim = X.shape
grad = np.zeros((dim,))
##########################################################################
# Compute the gradient of the loss function for unregularized logistic #
# regression #
# TODO: 1 line of code expected #
##########################################################################
grad = X.T.dot((utils.sigmoid(X.dot(theta)) - y)) / m + reg * theta / m
grad[0] = X[:, :1].T.dot((utils.sigmoid(X.dot(theta)) - y)) / m
###########################################################################
# END OF YOUR CODE #
###########################################################################
return grad
def logistic_predict(weights, data):
"""
Compute the probabilities predicted by the logistic classifier.
Note: N is the number of examples and
M is the number of features per example.
Inputs:
weights: (M+1) x 1 vector of weights, where the last element
corresponds to the bias (intercepts).
data: N x M data matrix where each row corresponds
to one data point.
Outputs:
y: :N x 1 vector of probabilities. This is the output of the classifier.
"""
# In case of MNIST classification of 4 and 9, output will be integer values
# TODO: Finish this function
N, M = data.shape
y = [0]*N
for i in range(0, N):
z = 0
for j in range(0, M):
z = z + weights[j] * data[i,j] + weights[-1]
y[i] = sigmoid(z)
#iprint 'z y[i]', z, y[i]
augdata = np.ones((N, M+1))
augdata[:, :-1] = data
z = np.dot(augdata, weights) # z is N x 1
y = sigmoid(z)
return y
def cost(self, theta1, theta2):
z1 = np.dot(self.train, theta1)
a2 = utils.sigmoid(z1)
a2 = np.append(np.ones((a2.shape[0],1)), a2, 1)
z2 = np.dot(a2, theta2)
h = utils.sigmoid(z2)
return -sum(sum(self.goal*np.log(h) + (1-self.goal)*np.log(1-h)))/self.m
def forward(self, x_t, h_tm1, c_tm1):
i_t = sigmoid(T.dot(x_t, self.W_xi) + T.dot(h_tm1, self.W_hi) + c_tm1 * self.W_ci)
f_t = sigmoid(T.dot(x_t, self.W_xf) + T.dot(h_tm1, self.W_hf) + c_tm1 * self.W_cf)
c_t = f_t * c_tm1 + i_t * self.activation(T.dot(x_t, self.W_xc) + T.dot(h_tm1, self.W_hc))
o_t = sigmoid(T.dot(x_t, self.W_xo) + T.dot(h_tm1, self.W_ho) + c_t * self.W_co)
h_t = o_t * self.activation(c_t)
return h_t, c_t
def _step(x_t, ct_1, ht_1, Wi, Wf, Wo, Wc, Whi, Whf, Who, Whc, bi, bf, bo, bc): i = sigmoid(T.dot(x_t, Wi) + T.dot(ht_1, Whi) + bi) f = sigmoid(T.dot(x_t, Wf) + T.dot(ht_1, Whf) + bf) o = sigmoid(T.dot(x_t, Wo) + T.dot(ht_1, Who) + bo) c = tanh(T.dot(x_t, Wc) + T.dot(ht_1, Whc) + bc) c_new = i * c + f * ct_1 h_new = o * tanh(c_new) return c_new, h_new
def _step(x_t, ct_1, ht_1, W, Wh, b, dim): tmp = T.dot(x_t, W) + T.dot(ht_1, Wh) + b i = sigmoid(_slice(tmp, 0, dim)) f = sigmoid(_slice(tmp, 1, dim)) o = sigmoid(_slice(tmp, 2, dim)) c = tanh(_slice(tmp, 3, dim)) c_new = i * c + f * ct_1 h_new = o * tanh(c_new) return c_new, h_new
def predict(self, newData=None):
if newData is None:
newData = self.train
else:
newData = np.append(np.ones((newData.shape[0],1)), newData, 1)
z = utils.sigmoid(np.dot(newData, self.inputWeight))
z = np.append(np.ones((z.shape[0],1)), z, 1)
digitProb = utils.sigmoid(np.dot(z, self.hiddenWeight))
return np.argmax(digitProb,1)
def _step_index(x_t, ct_1, ht_1, Wi, Wf, Wo, Wc, Whi, Whf, Who, Whc, bi, bf, bo, bc): # x_t: array of type int32 # use indexing on Wi, Wf, Wo and Wc matrices instead of computing the product with the one-hot representation of the input for computational and memory efficiency i = sigmoid(Wi[x_t] + T.dot(ht_1, Whi) + bi) f = sigmoid(Wf[x_t] + T.dot(ht_1, Whf) + bf) o = sigmoid(Wo[x_t] + T.dot(ht_1, Who) + bo) c = tanh(Wc[x_t] + T.dot(ht_1, Whc) + bc) c_new = i * c + f * ct_1 h_new = o * tanh(c_new) return c_new, h_new
def _step_index(x_t, ct_1, ht_1, W, Wh, b, dim): # x_t: array of type int32 # use indexing on W matrix instead of computing dot product with the one-hot representation of the input for computational and memory efficiency tmp = W[x_t] + T.dot(ht_1, Wh) + b i = sigmoid(_slice(tmp, 0, dim)) f = sigmoid(_slice(tmp, 1, dim)) o = sigmoid(_slice(tmp, 2, dim)) c = tanh(_slice(tmp, 3, dim)) c_new = i * c + f * ct_1 h_new = o * tanh(c_new) return c_new, h_new
def forward(network, x): W1, W2, W3 = network['W1'], network['W2'], network['W3'] b1, b2, b3 = network['b1'], network['b2'], network['b3'] a1 = np.dot(x, W1) + b1 z1 = sigmoid(a1) a2 = np.dot(z1, W2) + b2 z2 = sigmoid(a2) a3 = np.dot(z2, W3) + b3 y = identity_function(a3) return y
def get_reconstruction_cross_entropy(self):
pre_sigmoid_activation_h = numpy.dot(self.input, self.W) + self.hbias
sigmoid_activation_h = sigmoid(pre_sigmoid_activation_h)
pre_sigmoid_activation_v = numpy.dot(sigmoid_activation_h, self.W.T) + 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 _CD1(self, visible_data, weights, visible_bias, hidden_bias): N = np.shape(visible_data)[0] # Positive phase visible_state = visible_data if self.visible_type == "SIGMOID" : visible_state = self._samplebinary(visible_state) elif self.visible_type == "LINEAR" : visible_state = self._add_gaussian_noise(visible_state); nw = np.dot(visible_state, weights) + np.tile(hidden_bias, (N, 1)) if self.hidden_type == "SIGMOID": hidden_probability = u.sigmoid(nw) hidden_state = self._samplebinary(hidden_probability) elif self.hidden_type == "LINEAR": hidden_state = self._add_gaussian_noise(nw) gradient1 = self._gradient_weights(visible_state, hidden_state, weights) visible_biases1 = self._gradient_biases(visible_state, visible_bias) hidden_biases1 = self._gradient_biases(hidden_state, hidden_bias) # Negative phase # Skip sampling as well... visible_state = np.dot(hidden_state, weights.T) + np.tile(visible_bias, (N, 1)) if self.visible_type == "SIGMOID": visible_state = u.sigmoid(visible_state) #visible_probability = u.sigmoid(visible_state) #visible_state = self._samplebinary(visible_probability) # skip sampling here nw = np.dot(visible_state, weights) + np.tile(hidden_bias, (N, 1)) if self.hidden_type == "SIGMOID": hidden_probability = hidden_probability = u.sigmoid(nw) hidden_state = hidden_probability elif self.hidden_type == "LINEAR" : hidden_state = nw gradient2 = self._gradient_weights(visible_state, hidden_state, weights) visible_biases2 = self._gradient_biases(visible_state, visible_bias) hidden_biases2 = self._gradient_biases(hidden_state, hidden_bias) # gradients weights = gradient1 - gradient2; visible_biases = visible_biases1 - visible_biases2; hidden_biases= hidden_biases1 - hidden_biases2; return weights, visible_biases, hidden_biases
def train(set_, dimension, lambda_):
temp_w = np.zeros(dimension)
w0 = 0.
w = temp_w
# print ("Lambda: " + str(lambda_))
prev_error = 0.
h = [0.5] * len(set_)
current_error = calc_error(set_, w, lambda_, h, dimension)
# num_iter = 0
while abs(current_error - prev_error) > 0.001:
delta_w0 = 0.
delta_w = np.zeros(dimension)
# print ("Current error: " + str(current_error))
for i in range(len(set_)):
h = utils.sigmoid(np.dot(set_[i][1], w) + w0)
y = set_[i][0]
delta_w0 += float(h) - y
temp_x = (float(h) - y) * set_[i][1]
delta_w = delta_w + temp_x
n, error, w, w0 = line_search(set_, dimension, lambda_, w, w0,
delta_w, delta_w0, current_error)
# print ("Line search result: ")
# print (str(w0) + " " + str(w))
if n == 0:
break
# num_iter += 1
prev_error = current_error
current_error = error
# print (current_error)
# print ("Num iter: " + str(num_iter))
# print ("params found: " + str(w0) + str(w))
# print ("-------------------------------")
# print()
h = [float(utils.sigmoid(np.dot(tup[1], w) + w0))
for tup in set_]
return w, w0, calc_error(set_, w, 0, h, dimension)
def minibatch_update(self,x,y,lr,regularization):
n_sample = x.shape[0]
info = x
hidden_cache = []
for i in xrange(self.n_hidden + 1):
if i == self.n_hidden:
probs = softmax(info.dot(self.W[i]) + self.b[i])
else:
info = sigmoid(info.dot(self.W[i]) + self.b[i])
hidden_cache.append(info)
loss = neg_log_likelihood(probs,y)
probs[np.arange(n_sample),y] -= 1.0
errors = probs
for i in range(self.n_hidden,-1,-1):
if i >= 1:
hidden_out = hidden_cache[i - 1]
grad_hidden_out = errors.dot(self.W[i].T)
self.W[i] -= (lr * (hidden_out.T).dot(errors) + regularization * self.W[i])
self.b[i] -= lr * np.sum(errors,axis = 0)
errors = hidden_out * (1 - hidden_out) * grad_hidden_out
else:
hidden_out = x
self.W[i] -= (lr * (hidden_out.T).dot(errors) + regularization * self.W[i])
self.b[i] -= lr * np.sum(errors,axis = 0)
return loss
def logistic(weights, data, targets, hyperparameters):
"""
Calculate negative log likelihood and its derivatives with respect to weights.
Also return the predictions.
Note: N is the number of examples and
M is the number of features per example.
Inputs:
weights: (M+1) x 1 vector of weights, where the last element
corresponds to bias (intercepts).
data: N x M data matrix where each row corresponds
to one data point.
targets: N x 1 vector of binary targets. Values should be either 0 or 1.
hyperparameters: The hyperparameters dictionary.
Outputs:
f: The sum of the loss over all data points. This is the objective that we want to minimize.
df: (M+1) x 1 vector of derivative of f w.r.t. weights.
y: N x 1 vector of probabilities.
"""
# TODO: Finish this function
N, M = data.shape
z = [0.0]*N
df = np.zeros((M+1, 1))
f = 0.0
augdata = np.ones((N, M+1))
augdata[:, :-1] = data
z = np.dot(augdata, weights) # z is N x 1
f = float(np.dot(np.transpose(1-targets), z)[0] + sum(np.log(np.exp(-z) + 1))) # f is scalar
df = (np.dot(np.transpose(augdata), np.subtract(sigmoid(z).reshape((N, 1)), targets.reshape((N, 1)))))
y = np.array(logistic_predict(weights, data))
return f, df, y
def loss(self, *args):
"""
Compute the logistic loss function
Inputs:
- X: N x D array of data; each row is a data point.
- y: 1-dimensional array of length N with real values.
- reg: (float) regularization strength.
Returns: A tuple containing:
- loss as a single float
- gradient with respect to self.theta; an array of the same shape as theta
"""
theta,X,y,reg = args
m,dim = X.shape
J = 0
##########################################################################
# Compute the loss function for regularized logistic regression #
# TODO: 1-2 lines of code expected #
##########################################################################
hx = utils.sigmoid(X.dot(theta))
J = -1*(np.log(hx).T.dot(y)+(np.log(1-hx)).T.dot(1-y)) / m + reg * theta[1:].T.dot(theta[1:]) / (2 * m)
###########################################################################
# END OF YOUR CODE #
###########################################################################
return J
def loss(self, *args):
"""
Compute the logistic loss function
Inputs:
- X: N x D array of data; each row is a data point.
- y: 1-dimensional array of length N with real values.
Returns: loss as a single float
"""
theta,X,y = args
m,dim = X.shape
J = 0
##########################################################################
# Compute the loss function for unregularized logistic regression #
# TODO: 1-2 lines of code expected #
##########################################################################
hx = utils.sigmoid(X.dot(theta))
J = -1*(np.log(hx).T.dot(y)+(np.log(1-hx)).T.dot(1-y)) / m
###########################################################################
# END OF YOUR CODE #
###########################################################################
return J
def propup(self, v):
# stacking 2d convolutions here along depth dimension
# https://github.com/lmjohns3/py-rbm/blob/master/lmj/rbm.py seems
# to use 1-d convolutions, and I'm not sure is that's ok
# not going to escape a couple of loops though
# using theano's conventions:
# h is 4d matrix (num_examples, num_feature_maps,
# feature_map_height, feature_map_width)
# one feature map kinda corresponds to one hidden unit
# by the same convention, v is 4d matrix too: (num_examples,
# num_images per example (1, or 3 for RGB), image_height,
# image_widht)
# the same format is for weights: (number of feature maps for visible
# layer (1 or 3), number of feature maps for hidden layer,
# filter height, filter width)
num_examples = v.shape[0]
activations = np.zeros(
(
num_examples,
self.num_fm,
self.img_height - self.fm_height + 1,
self.img_width - self.fm_width + 1
)
)
for i in xrange(num_examples):
for j in xrange(self.num_fm):
activations[i, j, :, :] = convolve2d(v[i, 0, :, :], self.w[0, j, ::-1, ::-1], mode='valid')
return sigmoid(activations + self.b_hid[None, :, None, None])
def predict(self, X):
"""
Use the trained weights of this linear classifier to predict labels for
data points.
Inputs:
- X: m x d array of training data.
Returns:
- y_pred: Predicted output for the data in X. y_pred is a 1-dimensional
array of length m, and each element is a class label from one of the
set of labels -- the one with the highest probability
"""
y_pred = np.zeros(X.shape[0])
###########################################################################
# Compute the predicted outputs for X #
# TODO: 2 lines of code expected #
###########################################################################
hx = utils.sigmoid(X.dot(self.theta.T))
y_pred = np.argmax(hx, axis=1)
###########################################################################
# END OF YOUR CODE #
###########################################################################
return y_pred
def get_prediction(feature_vector, w):
sum_val = w[0]
for feature in feature_vector:
feature_id = int(feature.split(':')[0])
# strength of each feature is 1.0 so we skip multiplying it.
sum_val += w[feature_id]
return sigmoid(sum_val)
def _get_p(self, xt, wt=None):
if wt is None:
wt = self._get_w(xt)
wTx = sum(wt)
# bounded sigmoid
wTx = max(min(wTx, 20.), -20.)
return sigmoid(wTx)
def feed_forward(self, train_input):
self.layers[0].input = train_input
self.layers[0].output = train_input
for i in xrange(len(self.layers) - 1):
self.layers[i + 1].input = (self.weights[i].transpose() * self.layers[i].output) + self.bias[i]
self.layers[i + 1].output = sigmoid(self.layers[i + 1].input)
return self.layers[-1].output
def loss(self, *args):
"""
Compute the logistic loss function
Inputs:
- X: N x D array of data; each row is a data point.
- y: 1-dimensional array of length N with real values.
- reg: (float) regularization strength.
Returns: A tuple containing:
- loss as a single float
- gradient with respect to self.theta; an array of the same shape as theta
"""
theta,X,y,reg = args
m,dim = X.shape
J = 0
##########################################################################
# Compute the loss function for regularized logistic regression #
# TODO: 1-2 lines of code expected #
##########################################################################
J = 1. / m * sum([-y[i] * np.log(utils.sigmoid(theta.dot(X[i]))) - (1 - y[i]) * np.log(1 - utils.sigmoid(theta.dot(X[i]))) for i in xrange(m)])
J += reg / (2. * m) * sum(theta[1:] ** 2)
###########################################################################
# END OF YOUR CODE #
###########################################################################
return J
def loss(self, *args):
"""
Compute the logistic loss function
Inputs:
- X: N x D array of data; each row is a data point.
- y: 1-dimensional array of length N with real values.
Returns: loss as a single float
"""
theta,X,y = args
m,dim = X.shape
J = 0
##########################################################################
# Compute the loss function for unregularized logistic regression #
# TODO: 1-2 lines of code expected #
##########################################################################
J = 1. / m * sum([-y[i] * np.log(utils.sigmoid(theta.dot(X[i]))) - (1 - y[i]) * np.log(1 - utils.sigmoid(theta.dot(X[i]))) for i in xrange(m)])
###########################################################################
# END OF YOUR CODE #
###########################################################################
return J
def grad_loss(self, *args):
"""
Compute the gradient logistic loss function
Inputs:
- X: N x D array of data; each row is a data point.
- y: 1-dimensional array of length N with real values.
Returns: gradient with respect to theta; an array of the same shape as theta
"""
theta,X,y = args
m,dim = X.shape
grad = np.zeros((dim,))
##########################################################################
# Compute the gradient of the loss function for unregularized logistic #
# regression #
# TODO: 1 line of code expected #
##########################################################################
for j in xrange(dim):
grad[j] = 1. / m * sum([(utils.sigmoid(theta.dot(X[i])) - y[i]) * X[i][j] for i in xrange(m)])
###########################################################################
# END OF YOUR CODE #
###########################################################################
return grad
def grad_loss(self, *args):
"""
Compute the gradient logistic loss function
Inputs:
- X: N x D array of data; each row is a data point.
- y: 1-dimensional array of length N with real values.
- reg: (float) regularization strength.
Returns: A tuple containing:
- loss as a single float
- gradient with respect to self.theta; an array of the same shape as theta
"""
theta,X,y,reg = args
m,dim = X.shape
grad = np.zeros((dim,))
##########################################################################
# Compute the gradient of the loss function for unregularized logistic #
# regression #
# TODO: 1 line of code expected #
##########################################################################
for j in xrange(dim):
grad[j] = 1. / m * sum([(utils.sigmoid(theta.dot(X[i])) - y[i]) * X[i][j] for i in xrange(m)])
grad[j] += 1. * reg / m * theta[j] if j >= 1 else 0
###########################################################################
# END OF YOUR CODE #
###########################################################################
return grad
def logistic(weights, data, targets, hyperparameters):
"""
Calculate negative log likelihood and its derivatives with respect to weights.
Also return the predictions.
Note: N is the number of examples and
M is the number of features per example.
Inputs:
weights: (M+1) x 1 vector of weights, where the last element
corresponds to bias (intercepts).
data: N x M data matrix where each row corresponds
to one data point.
targets: N x 1 vector of binary targets. Values should be either 0 or 1.
hyperparameters: The hyperparameters dictionary.
Outputs:
f: The sum of the loss over all data points. This is the objective that we want to minimize.
df: (M+1) x 1 vector of derivative of f w.r.t. weights.
y: N x 1 vector of probabilities.
"""
t = np.transpose(np.repeat(np.reshape(weights[:-1], (len(weights)-1, 1)), len(data), axis = 1))
f_e = data * t
z_sums = np.sum(f_e, axis=1)
y = sigmoid(z_sums +weights[-1])
f = np.sum(np.log(1 + np.exp(-z_sums - weights[-1])) + (1 - np.transpose(targets)) * (z_sums + weights[-1]))
df = np.sum(data * np.transpose(((-np.exp(-z_sums - weights[-1]) / (1 + np.exp(-z_sums - weights[-1]))) + (1 - np.transpose(targets)))), axis = 0)
df = np.append(df, np.sum(np.transpose(((-np.exp(-z_sums - weights[-1]) / (1 + np.exp(-z_sums - weights[-1]))) + (1 - np.transpose(targets)))), axis = 0))
df = np.reshape(df, ((len(df), 1)))
return f, df, np.reshape(y, (len(y), 1))
def predict(self, X):
"""
Use the trained weights of this linear classifier to predict labels for
data points.
Inputs:
- X: N x D array of training data. Each row is a D-dimensional point.
Returns:
- y_pred: Predicted output for the data in X. y_pred is a 1-dimensional
array of length N, and each element is a real number.
"""
y_pred = np.zeros(X.shape[0])
###########################################################################
# Compute the predicted outputs for X #
# TODO: 1 line of code expected #
# #
###########################################################################
y_pred = np.round(utils.sigmoid(self.theta.T.dot(X.T)))
###########################################################################
# END OF YOUR CODE #
###########################################################################
return y_pred
def think(self, inputs): cur = inputs states = [cur] for syn in self.synapses: cur = utils.sigmoid(np.dot(cur, syn)) states.append(cur) return states
def logistic_predict(weights, data):
"""
Compute the probabilities predicted by the logistic classifier.
Note: N is the number of examples and
M is the number of features per example.
Inputs:
weights: (M+1) x 1 vector of weights, where the last element
corresponds to the bias (intercepts).
data: N x M data matrix where each row corresponds
to one data point.
Outputs:
y: :N x 1 vector of probabilities. This is the output of the classifier.
"""
# TODO: Finish this function
y = []
b = weights[-1]
w = weights[:-1]
sig_list = []
#print ("len of data",len(data))
for i in xrange(len(data)):
sig_val = sigmoid(np.dot(w.T, data[i]) + b)
sig_list.append(sig_val)
for i in sig_list:
y.append(i)
return y
def gradient(self, beta):
"""Summary
Args:
beta (TYPE): (D+1)-dim parameter (including bias) parametrising conditional probilities in logistic regression.
Returns:
TYPE: (D+1)-dim gradient of total loglikelihood with respect to beta
"""
sigmas = sigmoid(np.dot(self.X_ext, beta.T))
vect = (self.y - sigmas) * self.X_ext
return sum(vect)
def policy_network_forward_pass(self, x):
"""Computes forward pass of the policy network.
Policy network spits out a softmax over possible actions (i.e. P(going left) )
Network also uses relu activations in hidden layer
"""
h = np.dot(
self.model['W1'], x.T
) # W1 is a row but x is also a row. we want a column so transpose
h[h < 0] = 0 # relu on hidden state
p_left = np.dot(self.model['W2'], h)
p_left = utils.sigmoid(p_left)
return p_left, h
def predict(
X, weights
): # Used for predicting the output after the updation of weights to check the results.
n = len(weights) // 2
for i in range(0, n - 1):
Z = np.dot(X,
weights["W" + str(i + 1)].T) + weights["b" + str(i + 1)].T
A = utils.relu(Z)
X = A
Z = np.dot(X, weights["W" + str(n)].T) + weights["b" + str(n)].T
A = utils.sigmoid(Z)
return A
def excite(self, inputs, bias, filter_sigmoid=False):
'''excites this layer with the inputs and the bias and returns the
outputs as a tuple. if filter sigmoid is True, output from each neuron
is filtered through the sigmoid response curve.'''
outputs = []
for n in self:
res = n.excite(inputs, bias)
if filter_sigmoid:
res = sigmoid(res)
outputs.append(res)
return outputs
def tensor_factorization_loss(self):
# function to compute and monitor the tensor factorization loss (not directly minimized by the model)
samples = [
np.sort(sample_without_replacement(n_population=vsize, n_samples=32))
for vsize in self.vsizes
]
ijk_ = self.all_tuples_indices[self.all_tuples.isin(it.product(*samples[:-1]))]
num = self.tensor[ijk_][:, samples[-1]]
if self.sp:
num = num.toarray()
den = self.marginals[0][samples[0]]
for i in range(1, self.order):
den = np.tensordot(den, self.marginals[i][samples[i]], axes=0)
PMI_np = np.log(num.ravel()/den.ravel() + 1e-30)
batch = cartesian(samples)
return np.sum(np.abs(sigmoid(PMI_np - np.log(self.k_neg))\
- sigmoid(self.model(tf.tuple([batch[:,i] for i in range(self.order)])))))
def _predict(self, Xi):
"""Auxiliary function of predict.
Arguments:
Xi {list} -- 1d list object with int or float
Returns:
float -- prediction of yi
"""
z = self._linear(Xi)
return sigmoid(z)
def mean_pseudo_likelihood(self, x, idx): xfe = self.free_energy(x) # swaps the ith bit of each image # flip the i-th visible unit x[:, idx] = 1 - x[:, idx] nxfe = self.free_energy(x) # flip back the i-th visible unit x[:, idx] = 1 - x[:, idx] #logsig = np.log(sigmoid(nxfe - xfe)) cost = np.mean(self.nvisible * np.log(sigmoid(nxfe - xfe))) return cost
def __predict_nodes(self, nodes_idx: list, relations_idx: list,
known_triples: set):
scores = np.zeros((len(nodes_idx), self._gp.num_vertices))
doubler_score_node, doubler_score_doc = self._doubler.predict_nodes(
nodes_idx, relations_idx)
if self._normalize_score:
doubler_score_node = utils.sigmoid(doubler_score_node)
if doubler_score_doc is not None:
doubler_score_doc = utils.sigmoid(doubler_score_doc)
scores = np.sum([scores, doubler_score_node], axis=0)
if doubler_score_doc is not None:
scores = np.sum([scores, doubler_score_doc], axis=0)
ranks = []
num_scores = len(scores)
for idx, row in tqdm.tqdm(enumerate(scores),
total=num_scores,
desc='> Compute Ranking'):
is_head = idx % 2 == 0
node_idx_given = nodes_idx[idx]
relation_idx_given = relations_idx[idx]
node_idx_wanted = nodes_idx[idx +
1] if is_head else nodes_idx[idx - 1]
threshold_lower = row[node_idx_wanted]
rank_cleaned = 1
for node_idx, score in enumerate(row):
if score <= threshold_lower:
continue
elif is_head and not (node_idx_given, relation_idx_given,
node_idx) in known_triples:
rank_cleaned += 1
elif not is_head and not (node_idx, relation_idx_given,
node_idx_given) in known_triples:
rank_cleaned += 1
ranks.append(rank_cleaned)
return ranks, scores
def sample(self, h_prev, c_prev, num_char):
hs = np.copy(h_prev)
cs = np.copy(c_prev)
x = np.zeros((self.len_of_vocab, 1))
x[np.random.randint(0, self.len_of_vocab), 0] = 1
idxs = []
for _ in range(num_char):
I = np.dot(self.Wi, x) + np.dot(self.Ri,
hs) + self.Pi * cs + self.bi
i_gate = sigmoid(I)
F = np.dot(self.Wf, x) + np.dot(self.Rf,
hs) + self.Pf * cs + self.bo
f_gate = sigmoid(F)
Z = np.dot(self.Wz, x) + np.dot(self.Rz, hs) + self.bz
z = np.tanh(Z)
cs = i_gate * z + f_gate * cs
O = np.dot(self.Wo, x) + np.dot(self.Ro,
hs) + self.Po * cs + self.bo
o_gate = sigmoid(O)
hs = o_gate * np.tanh(cs)
if self.mode == 'lstm':
out = np.dot(self.Wout, hs) + self.bout
if self.mode == 'blstm':
out = np.dot(self.Wout, hs)
p = softmax(out)
idx = np.random.choice(self.len_of_vocab, 1, p=p.ravel())[0]
x = np.zeros((self.len_of_vocab, 1))
x[idx, 0] = 1
idxs.append(idx)
print(''.join(idx_to_char[c] for c in idxs))
def loadGameData(self, filePath):
numWeights = len(self.weights)
ret = []
with open(filePath, 'r') as f:
data = f.readlines()
# cleanup
data = [i.rstrip(' \n') for i in data]
data = [i.rstrip('\n') for i in data]
for i in range(self.numberOfGames):
newGame = SingleGame()
counter = 0
while data[counter] != 'END':
newGame.movesList.append(data[counter])
counter += 1
try:
featureValues = [
float(i) for i in data[counter].split(' ')
]
except:
print([i for i in data[counter].split(' ')])
sys.exit(0)
if (counter % 4 == 3):
#player 2's move, so features need to be interchanged
t = len(featureValues)
featureValues = featureValues[t /
2:] + featureValues[:t /
2]
newGame.featureValuesList.append(featureValues)
newGame.stateValues.append(
sigmoid(dotProduct(featureValues, self.weights)))
if (data[counter + 1] == 'END'):
#Victory & loss score
newGame.stateValues[
-1] = 1.0 #+ 0.01*newGame.featureValuesList[ -1][params.features.index("OppRingsCount")]
newGame.stateValues[-2] = 0.0
try:
assert (len(
newGame.featureValuesList[-1]) == numWeights)
except AssertionError:
print(len(newGame.featureValuesList[-1]))
print(numWeights)
sys.exit(0)
counter += 1
ret.append(newGame)
return ret
def _forward_propagate(self, x, l):
ai, i = np.concatenate(([1], x)), 0
acts = [ai]
#Forward propagation
while i < l:
ai = sigmoid(np.dot(self.W[i], ai))
ai = np.concatenate(([1], ai))
acts.append(ai)
i = i + 1
return acts
def sample(self, seed, number_of_characters_to_generate):
h = self.h
C = self.C
x = char_to_one_hot(seed)
ixes = []
for t in range(number_of_characters_to_generate):
x = np.matmul(self.W_1, x) + self.b_1
f = sigmoid(np.matmul(self.W_f, np.concatenate((h, x))) + self.b_f)
i = sigmoid(np.matmul(self.W_i, np.concatenate((h, x))) + self.b_i)
C_hat = np.tanh(
np.matmul(self.W_c, np.concatenate((h, x))) + self.b_c)
C = f * C + i * C_hat
o = sigmoid(np.matmul(self.W_o, np.concatenate((h, x))) + self.b_o)
h = o * np.tanh(C)
y = np.matmul(self.W_2, h) + self.b_2
p = np.exp(y) / np.sum(np.exp(y))
ix = np.random.choice(range(alphabet_size), p=p)
x = np.zeros(alphabet_size)
x[ix] = 1
ixes.append(ix)
return indices_to_string(ixes)
def forward(self, h_prev, c_prev):
self.hs = {}
self.cs = {}
self.i_gate = {}
self.f_gate = {}
self.o_gate = {}
self.z = {}
self.hs[-1] = np.copy(h_prev)
self.cs[-1] = np.copy(c_prev)
self.p = {}
self.loss = 0
for t in range(self.time_steps):
x = np.zeros((self.len_of_vocab, 1))
x[self.input[t], 0] = 1
I = np.dot(self.Wi, x) + np.dot(
self.Ri, self.hs[t - 1]) + self.Pi * self.cs[t - 1] + self.bi
self.i_gate[t] = sigmoid(I)
F = np.dot(self.Wf, x) + np.dot(
self.Rf, self.hs[t - 1]) + self.Pf * self.cs[t - 1] + self.bf
self.f_gate[t] = sigmoid(F)
Z = np.dot(self.Wz, x) + np.dot(self.Rz, self.hs[t - 1]) + self.bz
self.z[t] = np.tanh(Z)
self.cs[t] = self.i_gate[t] * self.z[t] + self.f_gate[t] * self.cs[
t - 1]
O = np.dot(self.Wo, x) + np.dot(
self.Ro, self.hs[t - 1]) + self.Po * self.cs[t] + self.bo
self.o_gate[t] = sigmoid(O)
self.hs[t] = self.o_gate[t] * np.tanh(self.cs[t])
if self.mode == 'lstm':
out = np.dot(self.Wout, self.hs[t]) + self.bout
self.p[t] = softmax(out)
self.loss += -np.log(self.p[t][self.output[t], 0])
def forwardProp(self, input_vectors):
"""
input_vector can be tuple, list or ndarray
"""
for i in np.arange(1, len(self.structure)):
layer = self.structure[i]
prev_layer = self.structure[i - 1]
if prev_layer['bias_node']:
input_vectors = np.concatenate(
(input_vectors, np.ones((input_vectors.shape[0], 1))),
axis=1) # change 1. to self.bias?
output_vectors = np.dot(layer['weight_matrix'], input_vectors.T).T
# activation functions
if layer['dropout_prob'] != 0:
if layer['activation'] == 'relu':
layer['output_vectors'] = layer['dropout_vector'] * relu(
output_vectors)
elif layer['activation'] == 'softmax':
layer['output_vectors'] = layer[
'dropout_vector'] * softmax(output_vectors.T).T
elif layer['activation'] == 'sigmoid':
layer['output_vectors'] = layer[
'dropout_vector'] * sigmoid(output_vectors)
else:
raise Exception('Activation function not recognised')
else:
if layer['activation'] == 'relu':
layer['output_vectors'] = relu(output_vectors)
elif layer['activation'] == 'softmax':
layer['output_vectors'] = softmax(output_vectors.T).T
elif layer['activation'] == 'sigmoid':
layer['output_vectors'] = sigmoid(output_vectors)
else:
raise Exception('Activation function not recognised')
input_vectors = layer['output_vectors']
def predictions(self, beta, tau=0.5):
"""Summary
Args:
beta (TYPE): (D+1)-dim parameter (including bias) parametrising conditional probilities in logistic regression.
tau (float, optional): threshold for predictions
Returns:
TYPE: 1-D array with predicted labels assigning 1 if conditional prob of 1 (given by logistic regression) > tau
"""
sigmas = sigmoid(np.dot(self.X_ext, beta.T))
self.y_pred = np.array(sigmas > tau * np.ones((self.N, 1)), dtype=int)
return self.y_pred
def predict(self, X):
Z = np.dot(self.w.T, X) + self.b
A = sigmoid(Z)
m = X.shape[1]
Y_pred = np.zeros((1, m))
for i in range(A.shape[1]):
if A[0, i] >= 0.5:
Y_pred[0, i] = 1
else:
Y_pred[0, i] = 0
return Y_pred
def fit(self, x, y, learningRate=0.001, noEpochs=1000):
self.coef_ = [0.0 for _ in range(1 + len(x[0]))] # y = w0 + w1 * x1 + w2 * x2 + ...
for epoch in range(noEpochs):
for i in range(len(x)): # for each sample from the training data
ycomputed = sigmoid(self.eval(x[i], self.coef_)) # estimate the output
crtError = float(ycomputed) - float(y[i]) # compute the error for the current sample
for j in range(0, len(x[0])): # update the coefficients
self.coef_[j + 1] = self.coef_[j + 1] - learningRate * crtError * float(x[i][j])
self.coef_[0] = self.coef_[0] - learningRate * crtError * 1
self.intercept_ = self.coef_[0]
self.coef_ = self.coef_[1:]
def forward_propagation(X, params):
"""
Args:
X -- Input data of shape (n_x, m) n_x -> no of input nodes; m -> no of training examples
params -- Dictionary with the initialized weights and the biases
Returns:
A2 -- Sigmoid output of 2nd or output layer of shape (n_y, m)
neuron_functions -- Dictionary containing the computed linear function Z and activation function A for each neuron
"""
W1 = params["W1"]
b1 = params["b1"]
W2 = params["W2"]
b2 = params["b2"]
# Compute the linear function Z and sigmoid Activation function for each neuron
Z1 = np.dot(W1, X) + b1
A1 = utils.sigmoid(Z1)
Z2 = np.dot(W2, A1) + b2
A2 = utils.sigmoid(Z2)
neuron_functions = {"Z1": Z1, "A1": A1, "Z2": Z2, "A2": A2}
return A2, neuron_functions
def lstm_numpy(x, W, U, b):
z = numpy.dot(x, W) + b
n_hidden = b.shape[0] / 4
h = numpy.zeros((x.shape[0], n_hidden), dtype=x.dtype)
prev_h = numpy.zeros(n_hidden, dtype=x.dtype)
prev_c = numpy.zeros(n_hidden, dtype=x.dtype)
def _slice(_x, n, dim):
return _x[n * dim:(n + 1) * dim]
for n in range(len(h)):
preact = numpy.dot(prev_h, U) + z[n]
i = utils.sigmoid(_slice(preact, 0, n_hidden))
f = utils.sigmoid(_slice(preact, 1, n_hidden))
o = utils.sigmoid(_slice(preact, 2, n_hidden))
c = utils.tanh(_slice(preact, 3, n_hidden))
c = f * prev_c + i * c
h[n] = o * utils.tanh(c)
prev_c = c
prev_h = h[n]
return h
def _linear_activation_forward(self, A_prev, W, b, activation):
"""
Get Activation A for a layer with given parameters.
Returns cache to use in backward propagation.
"""
if activation == "sigmoid":
Z, linear_cache = self._linear_forward(A_prev, W, b)
A, activation_cache = sigmoid(Z)
elif activation == "relu":
Z, linear_cache = self._linear_forward(A_prev, W, b)
A, activation_cache = relu(Z)
cache = (linear_cache, activation_cache)
return A, cache
def forward_propagation(self, input):
"""
Executes forward propagation.
Notice that the z and a of the first layer (l = 0) are equal to the NN's input.
:param input: input to the network.
:type input: (num_inputs, 1) numpy matrix.
:return z: values computed by applying weights and biases at each layer of the NN.
:rtype z: 3-dimensional list of (num_neurons[l], 1) numpy matrices.
:return a: activations computed by applying the activation function to z at each layer.
:rtype a: 3-dimensional list of (num_neurons[l], 1) numpy matrices.
"""
z = [None] * 3
a = [None] * 3
z[0] = input
a[0] = input
# Add logic for neural network inference
z[1] = self.weights[1] * a[0] + self.biases[1]
a[1] = sigmoid(z[1])
z[2] = self.weights[2] * a[1] + self.biases[2]
a[2] = sigmoid(z[2])
return z, a
def _loss_and_gradient_sigmoid(self, X, y):
"""
二分类 sigmoid
使用GD计算梯度与损失
:param X: NxD m_samples, n_features np.ndarray
:param y: N
:return:
"""
m_samples, n_features = X.shape
loss = 0.0
for m in range(m_samples):
interaction = 0.0
for i in range(n_features):
for j in range(i + 1, n_features):
interaction += np.dot(self.V[i, self.featureToField[j]],
self.V[j, self.featureToField[i]])
y_pred = self.wo + np.dot(self.W, X[m]) + interaction
delta = (sigmoid(y[m] * y_pred) - 1) * y[m]
# 更新参数
self.wo = self.wo - self.learning_rate * delta
for i in range(n_features):
if X[m, i] != 0:
self.W[i] = self.W[i] - self.learning_rate * delta * X[m,
i]
for j in range(i + 1, n_features):
self.V[i, self.featureToField[
j]] -= self.learning_rate * delta * self.V[
j, self.featureToField[i]] * X[m, i] * X[m, j]
self.V[j, self.featureToField[
i]] -= self.learning_rate * delta * self.V[
i, self.featureToField[j]] * X[m, i] * X[m, j]
# 计算loss
loss += -np.log(sigmoid(y[m] * y_pred))
print(loss / m_samples)
def predict(X, user, parameters, layers_dim):
A_prev = X
L = len(layers_dim)
for l in range(L-2):
Z = np.dot(parameters["W" + str(l+1)], A_prev) + parameters["b" + str(l+1)]
A = sigmoid(Z)
A_prev = A
ZL = np.dot(parameters["W" + str(L-1)], A_prev) + parameters["b" + str(L-1)]
AL = softmax(ZL)
print (AL)
Y = np.argmax(AL)
return Y
def lrCostFunction(theta, X, y, lambda_):
m = y.size
if y.dtype == bool:
y = y.astype(int)
J = 0
n=np.shape(theta)[0]
theta=np.reshape(theta,(n,1))
grad = np.zeros(theta.shape)
y=np.reshape(y,(np.shape(y)[0],1))
J=-(1/m)*(np.transpose(y).dot(np.log(utils.sigmoid(X.dot(theta))))+(np.transpose(1-y).dot(np.log(1-utils.sigmoid(X.dot(theta))))))+(lambda_/(2*m))*(np.transpose(theta[1:n]).dot(theta[1:n]))
for i in range(0, n):
a = 0
for j in range(0, m):
a = a + ((utils.sigmoid(np.dot(X[j, :], np.reshape(theta, (n, 1)))) - y[j]) * X[j, i])
if i == 0:
grad[i] = a / m
else:
grad[i] = a / m + (lambda_ / m) * theta[i]
grad = np.reshape(grad,(1,n))
return J, grad
def _negative_phase(self, neg_data):
"""Evaluate the negative phase.
Args:
neg_data: shape = [`_batch_size`, `_num_hidden`]
"""
# Hidden states, using probability itself. Shape = [_batch_size, _num_hidden]
neg_h_probs = utils.sigmoid(np.matmul(neg_data, self.weight) +
np.tile(self.hidden_bias, (self._batch_size, 1)))
self._neg_products = np.matmul(neg_data.transpose(), neg_h_probs)
# Shape = [1, _num_visible] / [1, _num_hidden]
self._neg_visible_act = np.mean(neg_data, axis=0)
self._neg_hidden_act = np.mean(neg_h_probs, axis=0)
def get_batch(self, batch_size, delayed_reward, discount_factor):
memory = zip(reversed(self.memory_sample[-batch_size:]),
reversed(self.memory_action[-batch_size:]),
reversed(self.memory_value[-batch_size:]),
reversed(self.memory_policy[-batch_size:]),
reversed(self.memory_reward[-batch_size:]),
reversed(self.memory_target_policy[-batch_size:]),
reversed(self.memory_target_action[-batch_size:]),
reversed(self.memory_value2[-batch_size:]))
sample_batch = np.zeros(
(batch_size, self.num_steps, self.num_features))
x_sample_next = np.zeros(
(batch_size, self.num_steps, self.num_features))
y_value = np.zeros((batch_size, self.agent.NUM_ACTIONS))
y_value2 = np.zeros((batch_size, self.agent.NUM_ACTIONS))
y_policy = np.full((batch_size, self.agent.NUM_ACTIONS), .5)
y_target_value = np.zeros((batch_size, self.agent.NUM_ACTIONS))
y_target_policy = np.full((batch_size, self.agent.NUM_ACTIONS), .5)
rewards = np.zeros(batch_size)
action_batch = np.asarray([data[1] for data in memory])
value_max_next = 0
value_max_next2 = 0
target_max_next = 0
x_sample_next[0] = self.memory_sample[-1]
reward_next = self.memory_reward[-1]
for i, (sample, action, value, policy, reward,target_policy,target_action, value2) \
in enumerate(memory):
sample_batch[i] = sample
y_value[i] = value
y_value2[i] = value2
y_policy[i] = policy
q1_vals = self.critic.target_model1_predict(sample)
q2_vals = self.critic.target_model2.predict(sample)
y_target_value[i] = np.min(np.vstack(
[q1_vals.transpose(), q2_vals.transpose()]),
axis=0)
y_target_policy[i] = target_policy
rewards[i] = (delayed_reward + reward_next - reward * 2) * 100
y_target_value[i, target_action] = rewards[
i] + discount_factor * target_max_next # q_value
# # y_target_policy[i, target_action] = sigmoid(target_value[target_action])
y_value[i, action] = value_max_next # q_value
y_value2[i, action] = value_max_next2 # q_value
y_policy[i, action] = sigmoid(y_value[action])
target_max_next = y_target_value.max()
value_max_next = value.max()
value_max_next2 = value2.max()
reward_next = reward
return sample_batch, y_policy, y_value, y_value2, y_target_value
def nnCostFunction(nn_params,
input_layer_size,
hidden_layer_size,
output_layer_size,
X,
y,
lambda_=0.0):
t1 = nn_params[:(input_layer_size + 1) * hidden_layer_size].reshape(
hidden_layer_size, input_layer_size + 1)
t2 = nn_params[(input_layer_size + 1) * hidden_layer_size:].reshape(
output_layer_size, hidden_layer_size + 1)
m = X.shape[0]
X = np.concatenate([np.ones((m, 1)), X], axis=1)
z2 = (np.dot(X, t1.T))
a2 = utils.sigmoid(z2) ### 5000*25
z3 = (np.dot(np.concatenate([np.ones((m, 1)), a2], axis=1), t2.T))
h = utils.sigmoid(z3)
y_matrix = np.zeros((m, output_layer_size))
for i in range(m):
y_matrix[i, y[i]] = 1.
####regularized costFunction
J = 1.*(1/m)*(-y_matrix*np.log(h) - (1-y_matrix)*np.log(1-h)).sum(axis=1).sum() \
+ lambda_*(0.5/m)*((t1**2)[:,1:].sum().sum()+(t2**2)[:,1:].sum().sum())
###back propagation
delta3 = h - y_matrix ##5000*10 t2: 10*26 t1:25*401
delta2 = np.dot(delta3, t2) * np.concatenate(
[np.zeros((m, 1)), a2 * (1 - a2)], axis=1) ## 5000*26
grad2 = (1/m) * np.dot(delta3.T,np.concatenate([np.ones((m,1)),a2], axis=1)) \
+ (lambda_/m)*np.concatenate([np.zeros((t2.shape[0],1)),t2[:,1:]],axis=1)
grad1 = (1/m) * np.dot(delta2.T[1:],X) \
+ (lambda_/m)*np.concatenate([np.zeros((t1.shape[0],1)),t1[:,1:]],axis=1)
grad = np.concatenate((grad1.ravel(), grad2.ravel()))
return J, grad
def predictive_novelty(state):
global Q, prev_state, prev_action, novelty_coeff, novelty_thresh
'''
This function will implement predictive novelty motivation as described in P.Y.Oudeye
Variables needed:
previous estimate
CURRENT ESTIMATE: THIS WILL BE PULLED FROM THE q VALUE
this value will be then passed to the sigmoid function to get
'''
# Find the next best action based om the current state
best_option = np.argwhere(Q[state[0],state[1],:] == np.amax(Q[state[0],state[1],:]))
num_options = len(best_option)
best_option = (random.choice(best_option))
action = best_option[0]
diff_array = np.zeros([4])
# this finds the new values for all the states.
for a in range(len(Q[state[0], state[1], :])):
diff_array[a] = utils.sigmoid(Q[state[0], state[1], a])
Q[state[0],state[1], a] =+ (diff_array[a] - novelty_thresh)*novelty_coeff
#Feed the Q values to the sigmoid function.
prev_estimate = utils.sigmoid(Q[prev_state[0], prev_state[1], prev_action])
this_estimate = utils.sigmoid(Q[state[0], state[1], action])
# make an array of deltas
# print('Prev estimate ', prev_estimate)
# print('This estimate ', this_estimate)
# delta = this_estimate - prev_estimate
diff_array = diff_array - 0.5
#print('Predictive novelty:', diff_array)
delta = diff_array.max()
# lets see what happens if I add this to the Q values?
return delta
def getAvgLoss(self, W, V, usr2NonzeroCols, usr2negsNonzeroCols,
usr2itemsIndx, pooler):
loss = 0.0
cnt = 0
for usrid in usr2itemsIndx:
try:
usrloss = 0.0
usr_rep = pooler.pool_all(usr2itemsIndx[usrid], V)
# 0. -log( sigmoid( usr_rep * sumedW_y) )
y_nonzeroCols = usr2NonzeroCols[usrid]
sumedW_y = sumOverW(W, y_nonzeroCols)
usrloss += (-1) * math.log(
sigmoid(usr_rep.transpose().dot(sumedW_y)))
# 1. summation log( sigmoid( usr_rep * sumedW_neg ) )
y_negsNonzeroCols = usr2negsNonzeroCols[usrid]
sumedW_negs = map(
lambda y_negNonzeroCols: sumOverW(W, y_negNonzeroCols).
reshape(self.ITEM_FIELDS_NUM, 1), y_negsNonzeroCols)
usrloss += (-1) * sum(
map(
lambda sumedW_neg: math.log(
sigmoid(
(-1) * usr_rep.transpose().dot(sumedW_neg))),
sumedW_negs))
# 2. l2 norm
l2norm = np.linalg.norm(W) + sum(
map(lambda v: np.linalg.norm(v), V))
usrloss += l2norm
loss += usrloss
cnt += 1
except:
loss += 0.0
cnt += 0
return loss / cnt
def __init__(self, link='softplus'):
"""
:param link: link function, either 'exp' or 'softplus' (note that the link is modified with an offset)
"""
super().__init__(hyp=None)
if link == 'exp':
self.link_fn = lambda mu: np.exp(mu - 0.5)
self.dlink_fn = lambda mu: np.exp(mu - 0.5)
elif link == 'softplus':
self.link_fn = lambda mu: softplus(mu - 0.5) + 1e-10
self.dlink_fn = lambda mu: sigmoid(mu - 0.5)
else:
raise NotImplementedError('link function not implemented')
self.name = 'Heteroscedastic Noise'