def gan_loss(*args): p_X = dmodel.forward(X, 'train') random_X = gmodel.forward(noise, 'train') p_random_X = dmodel.forward(random_X, 'train') value = np.log(clip(p_X, lower, upper)) + np.log(clip(1 - p_random_X, lower, upper)) loss = np.sum(value) / float(N) return loss
def loss(self, ps, as_, vs, rs, advs):
ps = np.maximum(1.0e-5, np.minimum(1.0 - 1e-5, ps))
policy_grad_loss = -np.sum(np.log(ps) * as_ * advs)
vf_loss = 0.5*np.sum((vs - rs)**2)
entropy = -np.sum(ps*np.log(ps))
loss_ = policy_grad_loss + self.config.vf_wt*vf_loss - self.config.entropy_wt*entropy
return loss_
def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Either of the followings:
- One hot encoding of labels, of shape (N, C)
- Label index of shape (N, ), each y[i] is the label of i^th example
(0 <= y[i] < C)
Returns a tuple of:
- loss: Scalar giving the loss
"""
N = x.shape[0]
C = x.shape[1]
if len(y.shape) == 1:
#convert it to one hot encoding
onehot_y = np.zeros([N, C])
np.onehot_encode(y, onehot_y)
else:
onehot_y = y
probs = x - np.max(x, axis=1, keepdims=True)
loss = -np.sum(probs * onehot_y) / N
loss += np.sum(np.log(np.sum(np.exp(probs), axis=1, keepdims=True))) / N
return loss
def _softmax_loss(self, X, y, *args):
N = X.shape[0]
scores = self._forward(X, *args)
scores = np.exp(scores - np.max(scores, axis=1, keepdims=True))
prob = scores / np.sum(scores, axis=1, keepdims=True)
loss = np.sum(-np.log(prob[np.arange(N), y])) / float(N)
return loss
def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Vector of labels, of shape (N,) where y[i] is the label for x[i] and
0 <= y[i] < C
Returns a tuple of:
- loss: Scalar giving the loss
- dx: Gradient of the loss with respect to x
"""
#np.expand_dims(correct_class_scores, axis = 1)
#probs = np.exp(x - np.max(x, axis=1, keepdims=True))
#print "x.shape", x.shape
#Somehow Buggy. Max doesn't work.
probs = np.exp(x - np.max(x, axis=1))
#probs /= np.expand_dims(np.sum(probs, axis=1), axis = 1)
probs /= np.expand_dims(np.sum(probs, axis=1), axis = 1)
N = x.shape[0]
loss = -np.sum(np.log(probs[np.arange(N), y])) / N
dx = probs.copy()
dx[np.arange(N), y] -= 1
dx /= N
return loss, dx
def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Either of the followings:
- One hot encoding of labels, of shape (N, C)
- Label index of shape (N, ), each y[i] is the label of i^th example
(0 <= y[i] < C)
Returns a tuple of:
- loss: Scalar giving the loss
"""
N = x.shape[0]
C = x.shape[1]
if len(y.shape) == 1:
#convert it to one hot encoding
onehot_y = np.zeros([N, C])
np.onehot_encode(y, onehot_y)
else:
onehot_y = y
probs = x - np.max(x, axis=1, keepdims=True)
loss = -np.sum(probs * onehot_y) / N
loss += np.sum(np.log(np.sum(np.exp(probs), axis=1, keepdims=True))) / N
return loss
def softmax_cross_entropy(prob, label):
"""
Computes the cross entropy for softmax activation.
Inputs:
- prob: Probability, of shape (N, C) where x[i, j] is the probability for the jth class
for the ith input.
- label: Either of the followings:
- One hot encoding of labels, of shape (N, C)
- Label index of shape (N, ), each y[i] is the label of i^th example
(0 <= y[i] < C)
Returns a Value:
- cross_entropy
"""
N = prob.shape[0]
C = prob.shape[1]
if len(label.shape) == 1:
#convert it to one hot encoding
onehot_label = np.zeros([N, C])
np.onehot_encode(label, onehot_label)
else:
onehot_label = label
return -np.sum(np.log(prob) * onehot_label) / N
def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Vector of labels, of shape (N,) where y[i] is the label for x[i] and
0 <= y[i] < C
Returns a tuple of:
- loss: Scalar giving the loss
- dx: Gradient of the loss with respect to x
"""
#np.expand_dims(correct_class_scores, axis = 1)
#probs = np.exp(x - np.max(x, axis=1, keepdims=True))
#print "x.shape", x.shape
#Somehow Buggy. Max doesn't work.
probs = np.exp(x - np.max(x, axis=1))
#probs /= np.expand_dims(np.sum(probs, axis=1), axis = 1)
probs /= np.expand_dims(np.sum(probs, axis=1), axis=1)
N = x.shape[0]
loss = -np.sum(np.log(probs[np.arange(N), y])) / N
dx = probs.copy()
dx[np.arange(N), y] -= 1
dx /= N
return loss, dx
def softmax_cross_entropy(prob, label):
"""
Computes the cross entropy for softmax activation.
Inputs:
- prob: Probability, of shape (N, C) where x[i, j] is the probability for the jth class
for the ith input.
- label: Either of the followings:
- One hot encoding of labels, of shape (N, C)
- Label index of shape (N, ), each y[i] is the label of i^th example
(0 <= y[i] < C)
Returns a Value:
- cross_entropy
"""
N = prob.shape[0]
C = prob.shape[1]
if len(label.shape) == 1:
#convert it to one hot encoding
onehot_label = np.zeros([N, C])
np.onehot_encode(label, onehot_label)
else:
onehot_label = label
return -np.sum(np.log(prob) * onehot_label) / N
def softmax(x, y):
import numpy as np
y = y.astype(int)
probs = np.exp(x - np.max(x, axis=1, keepdims=True))
probs /= np.sum(probs, axis=1, keepdims=True)
N = x.shape[0]
loss = -np.sum(np.log(probs[np.arange(N), y])) / N
return loss
def softmax_crossentropy(x, y):
# x should be (batch, prob)
# y should be (batch, )
x_dev = x - np.max(x, axis=1, keepdims=True) # minpy doesn't support x.max()
sm = x_dev - np.log(np.sum(np.exp(x_dev), axis=1, keepdims=True))
ids = np.arange(0, y.shape[0])*x.shape[1] + y
ce = -np.sum(sm.reshape((sm.shape[0]*sm.shape[1],))[ids])/(1.0*y.shape[0]) # minpy doesn't support -1 in shape inference
return ce
def softmax_cross_entropy(prob, label):
N = prob.shape[0]
C = prob.shape[1]
if len(label.shape) == 1:
#convert it to one hot encoding
onehot_label = np.zeros([N, C])
np.onehot_encode(label, onehot_label)
else:
onehot_label = label
return -np.sum(np.log(prob) * onehot_label) / N
def softmax_cross_entropy(prob, label):
N = prob.shape[0]
C = prob.shape[1]
if len(label.shape) == 1:
#convert it to one hot encoding
onehot_label = np.zeros([N, C])
np.onehot_encode(label, onehot_label)
else:
onehot_label = label
return -np.sum(np.log(prob) * onehot_label) / N
def train_loss(*args):
inputs = args[0]
softmax_label = args[1]
probs = self.symbol_func(**self.make_mxnet_weight_dict(inputs, softmax_label, args[self.data_target_cnt:len(args)]))
if softmax_label is None:
return probs
samples_num = X.shape[0]
targets = np.zeros((samples_num, self.num_classes))
targets[np.arange(samples_num), softmax_label] = 1
loss = -np.sum(targets * np.log(probs)) / samples_num
for i in self.get_index_reg_weight():
loss = loss + np.sum(0.5*args[i]**2*self.reg)
return loss
def calculate_cross_entropy_loss(s, y, w, lambda_):
"""
Calculates the loss of a score matrix depending on the ground truth labels.
This method uses cross entropy loss (from Softmax).
:param s: The score matrix of form (N x K), where N is the number of observations and K is the number of classes.
:param y: The ground truth label vector of length N.
:param w: The weight matrix of the output layer of form (H x K), where H is the size of the previous layer.
:param lambda_: The regularization loss hyperparameter.
:return: The cross-entropy loss, where 0 indicates a perfect match between s and y
and +Inf indicates a perfect mismatch.
"""
probabilities = probs(s)
log_probabilities = -np.log(probabilities[np.array(range(y.shape[0])), y])
data_loss = np.sum(log_probabilities) / y.shape[0]
return data_loss + regularization_loss(w, lambda_)
def train_loss(*args):
inputs = args[0]
softmax_label = args[1]
probs = self.symbol_func(**self.make_mxnet_weight_dict(
inputs, softmax_label, args[self.data_target_cnt:len(args)]))
if softmax_label is None:
return probs
samples_num = X.shape[0]
targets = np.zeros((samples_num, self.num_classes))
targets[np.arange(samples_num), softmax_label] = 1
loss = -np.sum(targets * np.log(probs)) / samples_num
for i in self.get_index_reg_weight():
loss = loss + np.sum(0.5 * args[i]**2 * self.reg)
return loss
def softmax_crossentropy(x, y):
EPSI = 1e-6
batch_size, seq_len, prob_dim = x.shape
x = x.reshape((x.shape[0] * x.shape[1], x.shape[2]))
y = y.reshape((y.shape[0] * y.shape[1], ))
#print x.shape, y.shape
# x should be (batch, prob)
# y should be (batch, )
x_dev = x - np.max(x, axis=1,
keepdims=True) # minpy doesn't support x.max()
sm = x_dev - np.log(EPSI + np.sum(np.exp(x_dev), axis=1, keepdims=True))
ids = np.arange(0, y.shape[0]) * seq_len + y
ce = -np.sum(sm.reshape((sm.shape[0] * sm.shape[1], ))[ids]) / (
1.0 * y.shape[0]) # minpy doesn't support -1 in shape inference
return ce
def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Vector of labels, of shape (N,) where y[i] is the label for x[i] and
0 <= y[i] < C
Returns a tuple of:
- loss: Scalar giving the loss
"""
#TODO: Missing Max Operator
probs = np.exp(x - np.expand_dims(np.max(x, axis=1), axis = 1))
probs = probs / np.expand_dims(np.sum(probs, axis=1), axis = 1)
N = x.shape[0]
loss = -np.sum(np.log(probs[np.arange(N), y])) / N
return loss
def softmax_loss(x, y):
"""
Computes the loss and gradient for softmax classification.
Inputs:
- x: Input data, of shape (N, C) where x[i, j] is the score for the jth class
for the ith input.
- y: Vector of labels, of shape (N,) where y[i] is the label for x[i] and
0 <= y[i] < C
Returns a tuple of:
- loss: Scalar giving the loss
"""
#TODO: Missing Max Operator
probs = np.exp(x - np.expand_dims(np.max(x, axis=1), axis=1))
probs = probs / np.expand_dims(np.sum(probs, axis=1), axis=1)
N = x.shape[0]
loss = -np.sum(np.log(probs[np.arange(N), y])) / N
return loss
def temporal_softmax_loss(x, y, mask, verbose=False):
"""
A temporal version of softmax loss for use in RNNs. We assume that we are
making predictions over a vocabulary of size V for each timestep of a
timeseries of length T, over a minibatch of size N. The input x gives scores
for all vocabulary elements at all timesteps, and y gives the indices of the
ground-truth element at each timestep. We use a cross-entropy loss at each
timestep, summing the loss over all timesteps and averaging across the
minibatch.
As an additional complication, we may want to ignore the model output at some
timesteps, since sequences of different length may have been combined into a
minibatch and padded with NULL tokens. The optional mask argument tells us
which elements should contribute to the loss.
Inputs:
- x: Input scores, of shape (N, T, V)
- y: Ground-truth indices, of shape (N, T) where each element is in the range
0 <= y[i, t] < V
- mask: Boolean array of shape (N, T) where mask[i, t] tells whether or not
the scores at x[i, t] should contribute to the loss.
Returns a tuple of:
- loss: Scalar giving loss
- dx: Gradient of loss with respect to scores x.
"""
N, T, V = x.shape
x_flat = x.reshape(N * T, V)
y_flat = y.reshape(N * T)
mask_flat = mask.reshape(N * T)
probs = np.exp(x_flat - np.max(x_flat, axis=1, keepdims=True))
probs = probs / np.sum(probs, axis=1, keepdims=True)
loss = -np.sum(mask_flat * np.log(probs[np.arange(N * T), y_flat])) / N
return loss
def temporal_softmax_loss(x, y, mask, verbose=False):
"""
A temporal version of softmax loss for use in RNNs. We assume that we are
making predictions over a vocabulary of size V for each timestep of a
timeseries of length T, over a minibatch of size N. The input x gives scores
for all vocabulary elements at all timesteps, and y gives the indices of the
ground-truth element at each timestep. We use a cross-entropy loss at each
timestep, summing the loss over all timesteps and averaging across the
minibatch.
As an additional complication, we may want to ignore the model output at some
timesteps, since sequences of different length may have been combined into a
minibatch and padded with NULL tokens. The optional mask argument tells us
which elements should contribute to the loss.
Inputs:
- x: Input scores, of shape (N, T, V)
- y: Ground-truth indices, of shape (N, T) where each element is in the range
0 <= y[i, t] < V
- mask: Boolean array of shape (N, T) where mask[i, t] tells whether or not
the scores at x[i, t] should contribute to the loss.
Returns a tuple of:
- loss: Scalar giving loss
- dx: Gradient of loss with respect to scores x.
"""
N, T, V = x.shape
x_flat = x.reshape(N * T, V)
y_flat = y.reshape(N * T)
mask_flat = mask.reshape(N * T)
probs = np.exp(x_flat - np.max(x_flat, axis=1, keepdims=True))
probs = probs / np.sum(probs, axis=1, keepdims=True)
loss = -np.sum(mask_flat * np.log(probs[np.arange(N * T), y_flat])) / N
return loss
def train_loss(w, x):
prob = predict(w, x)
loss = -np.sum(label * np.log(prob)) / num_samples
return loss
def logsoftmax(x, valid_idx):
x[np.array(valid_idx)] += 1e6
x_max = np.max(x)
return x - x_max - np.log(np.sum(np.exp(x - x_max)))
def loss(self, ps, ys, rs):
# Prevent log of zero.
ps = np.maximum(1.0e-5, np.minimum(1.0 - 1e-5, ps))
step_losses = ys * np.log(ps) + (1.0 - ys) * np.log(1.0 - ps)
return -np.sum(step_losses * rs)
def training_loss(inputs, targets, fc_weight, fc_bias, conv_weight, conv_bias):
preds = predict(inputs, fc_weight, fc_bias, conv_weight, conv_bias)
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
return -np.sum(np.log(label_probabilities))
def loss(self, ps, ys, rs):
# Prevent log of zero.
ps = np.maximum(1.0e-5, np.minimum(1.0 - 1e-5, ps))
step_losses = ys * np.log(ps) + (1.0 - ys) * np.log(1.0 - ps)
return -np.sum(step_losses * rs)
def loss(w, x):
prob = predict(w, x)
return -np.sum(np.log(prob) * t) / 10000 + 0.5 * w * w
def logsumexp(X, axis=1):
max_X = np.max(X)
return max_X + np.log(np.sum(np.exp(X - max_X), axis=axis, keepdims=True))
def U(beta):
return mp.sum(mp.log(1 + mp.exp(mp.dot(X, beta))))-mp.dot(y.T,(mp.dot(X,beta)))+(0.5/alpha)*mp.sum(beta**2)
def loss(w, x):
prob = predict(w, x)
return np.reshape(-np.sum(np.log(prob) * t) / 10000, (1, 1)) + 0.5 * w * w
def loss_theta(layer):
pred = predict(layer, inputs)
return -np.sum(np.log(pred * targets)) # negative log likelihood
def cross_entropy(predictions, targets, epsilon=1e-12):
predictions = np.clip(predictions, epsilon, 1. - epsilon)
N = predictions.shape[0]
ce = -np.sum(targets * np.log(predictions + 1e-9)) / N
return float(ce[0])
def training_loss(weights, inputs):
preds = predict(weights, inputs)
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
l = -np.sum(np.log(label_probabilities))
return l
def test_ufunc():
x = np.array([-1.2, 1.2])
np.absolute(x)
np.absolute(1.2 + 1j)
x = np.linspace(start=-10, stop=10, num=101)
np.add(1.0, 4.0)
x1 = np.arange(9.0).reshape((3, 3))
x2 = np.arange(3.0)
np.add(x1, x2)
np.arccos([1, -1])
x = np.linspace(-1, 1, num=100)
np.arccosh([np.e, 10.0])
np.arccosh(1)
np.arcsin(0)
np.arcsinh(np.array([np.e, 10.0]))
np.arctan([0, 1])
np.pi/4
x = np.linspace(-10, 10)
x = np.array([-1, +1, +1, -1])
y = np.array([-1, -1, +1, +1])
np.arctan2(y, x) * 180 / np.pi
np.arctan2([1., -1.], [0., 0.])
np.arctan2([0., 0., np.inf], [+0., -0., np.inf])
np.arctanh([0, -0.5])
np.bitwise_and(13, 17)
np.bitwise_and(14, 13)
# np.binary_repr(12) return str
np.bitwise_and([14,3], 13)
np.bitwise_and([11,7], [4,25])
np.bitwise_and(np.array([2,5,255]), np.array([3,14,16]))
np.bitwise_and([True, True], [False, True])
np.bitwise_or(13, 16)
# np.binary_repr(29)
np.bitwise_or(32, 2)
np.bitwise_or([33, 4], 1)
np.bitwise_or([33, 4], [1, 2])
np.bitwise_or(np.array([2, 5, 255]), np.array([4, 4, 4]))
# np.array([2, 5, 255]) | np.array([4, 4, 4])
np.bitwise_or(np.array([2, 5, 255, 2147483647], dtype=np.int32),
np.array([4, 4, 4, 2147483647], dtype=np.int32))
np.bitwise_or([True, True], [False, True])
np.bitwise_xor(13, 17)
# np.binary_repr(28)
np.bitwise_xor(31, 5)
np.bitwise_xor([31,3], 5)
np.bitwise_xor([31,3], [5,6])
np.bitwise_xor([True, True], [False, True])
a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])
np.ceil(a)
a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])
np.trunc(a)
np.cos(np.array([0, np.pi/2, np.pi]))
np.cosh(0)
x = np.linspace(-4, 4, 1000)
rad = np.arange(12.)*np.pi/6
np.degrees(rad)
out = np.zeros((rad.shape))
r = np.degrees(rad, out)
# np.all(r == out) return bool
np.rad2deg(np.pi/2)
np.divide(2.0, 4.0)
x1 = np.arange(9.0).reshape((3, 3))
x2 = np.arange(3.0)
np.divide(2, 4)
np.divide(2, 4.)
np.equal([0, 1, 3], np.arange(3))
np.equal(1, np.ones(1))
x = np.linspace(-2*np.pi, 2*np.pi, 100)
np.exp2([2, 3])
np.expm1(1e-10)
np.exp(1e-10) - 1
np.fabs(-1)
np.fabs([-1.2, 1.2])
a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])
np.floor(a)
np.floor_divide(7,3)
np.floor_divide([1., 2., 3., 4.], 2.5)
np.fmod([-3, -2, -1, 1, 2, 3], 2)
np.remainder([-3, -2, -1, 1, 2, 3], 2)
np.fmod([5, 3], [2, 2.])
a = np.arange(-3, 3).reshape(3, 2)
np.fmod(a, [2,2])
np.greater([4,2],[2,2])
a = np.array([4,2])
b = np.array([2,2])
a > b
np.greater_equal([4, 2, 1], [2, 2, 2])
np.hypot(3*np.ones((3, 3)), 4*np.ones((3, 3)))
np.hypot(3*np.ones((3, 3)), [4])
np.bitwise_not is np.invert
np.invert(np.array([13], dtype=np.uint8))
# np.binary_repr(242, width=8)
np.invert(np.array([13], dtype=np.uint16))
np.invert(np.array([13], dtype=np.int8))
# np.binary_repr(-14, width=8)
np.invert(np.array([True, False]))
# np.isfinite(1)
# np.isfinite(0)
# np.isfinite(np.nan)
# np.isfinite(np.inf)
# np.isfinite(np.NINF)
x = np.array([-np.inf, 0., np.inf])
y = np.array([2, 2, 2])
np.isfinite(x, y)
# np.isinf(np.inf)
# np.isinf(np.nan)
# np.isinf(np.NINF)
# np.isinf([np.inf, -np.inf, 1.0, np.nan])
x = np.array([-np.inf, 0., np.inf])
y = np.array([2, 2, 2])
# np.isinf(x, y)
# np.isnan(np.nan)
# np.isnan(np.inf)
# np.binary_repr(5)
np.left_shift(5, 2)
# np.binary_repr(20)
np.left_shift(5, [1,2,3])
np.less([1, 2], [2, 2])
np.less_equal([4, 2, 1], [2, 2, 2])
x = np.array([0, 1, 2, 2**4])
xi = np.array([0+1.j, 1, 2+0.j, 4.j])
np.log2(xi)
prob1 = np.log(1e-50)
prob2 = np.log(2.5e-50)
prob12 = np.logaddexp(prob1, prob2)
prob12
np.exp(prob12)
prob1 = np.log2(1e-50)
prob2 = np.log2(2.5e-50)
prob12 = np.logaddexp2(prob1, prob2)
prob1, prob2, prob12
2**prob12
np.log1p(1e-99)
np.log(1 + 1e-99)
# np.logical_and(True, False)
# np.logical_and([True, False], [False, False])
x = np.arange(5)
# np.logical_and(x>1, x<4)
# np.logical_not(3)
# np.logical_not([True, False, 0, 1])
x = np.arange(5)
# np.logical_not(x<3)
# np.logical_or(True, False)
# np.logical_or([True, False], [False, False])
x = np.arange(5)
# np.logical_or(x < 1, x > 3)
# np.logical_xor(True, False)
# np.logical_xor([True, True, False, False], [True, False, True, False])
x = np.arange(5)
# np.logical_xor(x < 1, x > 3)
# np.logical_xor(0, np.eye(2))
np.maximum([2, 3, 4], [1, 5, 2])
# np.maximum([np.nan, 0, np.nan], [0, np.nan, np.nan])
# np.maximum(np.Inf, 1)
np.minimum([2, 3, 4], [1, 5, 2])
# np.minimum([np.nan, 0, np.nan],[0, np.nan, np.nan])
# np.minimum(-np.Inf, 1)
np.fmax([2, 3, 4], [1, 5, 2])
np.fmax(np.eye(2), [0.5, 2])
# np.fmax([np.nan, 0, np.nan],[0, np.nan, np.nan])
np.fmin([2, 3, 4], [1, 5, 2])
np.fmin(np.eye(2), [0.5, 2])
# np.fmin([np.nan, 0, np.nan],[0, np.nan, np.nan])
np.modf([0, 3.5])
np.modf(-0.5)
np.multiply(2.0, 4.0)
x1 = np.arange(9.0).reshape((3, 3))
x2 = np.arange(3.0)
np.multiply(x1, x2)
np.negative([1.,-1.])
np.not_equal([1.,2.], [1., 3.])
np.not_equal([1, 2], [[1, 3],[1, 4]])
x1 = range(6)
np.power(x1, 3)
x2 = [1.0, 2.0, 3.0, 3.0, 2.0, 1.0]
np.power(x1, x2)
x2 = np.array([[1, 2, 3, 3, 2, 1], [1, 2, 3, 3, 2, 1]])
np.power(x1, x2)
deg = np.arange(12.) * 30.
np.radians(deg)
out = np.zeros((deg.shape))
ret = np.radians(deg, out)
ret is out
np.deg2rad(180)
np.reciprocal(2.)
np.reciprocal([1, 2., 3.33])
np.remainder([4, 7], [2, 3])
np.remainder(np.arange(7), 5)
# np.binary_repr(10)
np.right_shift(10, 1)
# np.binary_repr(5)
np.right_shift(10, [1,2,3])
a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])
np.rint(a)
np.sign([-5., 4.5])
np.sign(0)
# np.sign(5-2j)
# np.signbit(-1.2)
np.signbit(np.array([1, -2.3, 2.1]))
np.copysign(1.3, -1)
np.copysign([-1, 0, 1], -1.1)
np.copysign([-1, 0, 1], np.arange(3)-1)
np.sin(np.pi/2.)
np.sin(np.array((0., 30., 45., 60., 90.)) * np.pi / 180. )
x = np.linspace(-np.pi, np.pi, 201)
np.sinh(0)
# np.sinh(np.pi*1j/2)
np.sqrt([1,4,9])
np.sqrt([4, -1, -3+4J])
np.cbrt([1,8,27])
np.square([-1j, 1])
np.subtract(1.0, 4.0)
x1 = np.arange(9.0).reshape((3, 3))
x2 = np.arange(3.0)
np.subtract(x1, x2)
np.tan(np.array([-pi,pi/2,pi]))
np.tanh((0, np.pi*1j, np.pi*1j/2))
x = np.arange(5)
np.true_divide(x, 4)
x = np.arange(9)
y1, y2 = np.frexp(x)
y1 * 2**y2
np.ldexp(5, np.arange(4))
x = np.arange(6)
np.ldexp(*np.frexp(x))
def training_loss(weights, inputs):
preds = predict(weights, inputs)
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
l = -np.sum(np.log(label_probabilities))
return l
def logsumexp(X, axis=1):
max_X = np.max(X)
return max_X + np.log(np.sum(np.exp(X - max_X), axis=axis, keepdims=True))
def train_loss(w, x):
prob = predict(w, x)
loss = -np.sum(label * np.log(prob)) / num_samples
return loss
def sigmoid(x):
return np.multiply(0.5, np.add(np.tanh(x), 1))
x = mx.sym.Variable(name='x')
fc = mx.sym.FullyConnected(name='fc', data=x)
#fc = mx.sym.FullyConnected(name='fc', data=x, num_hidden=inputs.shape[1])
act = mx.sym.Activation(data=fc, act_type='sigmoid'）
f = core.function(act)
def predict(weights, inputs):
return f(x=inputs, fc_weight=weights, ctx=mx.cpu())
def training_loss(weights, inputs):
preds = predict(weights, inputs)
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
return -np.sum(np.log(label_probabilities))
xshape = (256, 500)
wshape = (500, 250)
tshape = (256, 250)
inputs = np.random.rand(*xshape) - 0.5
targets = np.random.randint(0, 2, size=tshape)
weights = np.random.rand(*wshape) - 0.5
training_gradient_fun = core.grad(training_loss)
print('Initial loss: {}'.format(training_loss(weights, inputs)))
for i in range(100):
gr = training_gradient_fun(weights, inputs)
#print('Training gradient: {}'.format(gr))
weights -= gr * 0.1
def logsumexp(X, axis, keepdims=False):
max_X = np.max(X)
return max_X + np.log(np.sum(np.exp(X - max_X), axis=axis, keepdims=keepdims))
cumulative_update = np.zeros_like(net_params) # initialize update values
for ui, k_id in enumerate(kids_rank):
np.random.seed(noise_seed[k_id]) # reconstruct noise using seed
cumulative_update += utility[ui] * sign(k_id) * np.random.randn(
net_params.size)
gradients = optimizer.get_gradients(cumulative_update /
(2 * N_KID * SIGMA))
return net_params + gradients, rewards
if __name__ == "__main__":
# utility instead reward for update parameters (rank transformation)
base = N_KID * 2 # *2 for mirrored sampling
rank = np.arange(1, base + 1)
util_ = np.maximum(0, numpy.log(base / 2 + 1) - np.log(rank))
utility = util_ / sum(util_) - 1 / base
# training
net_shapes, net_params = build_net()
env = gym.make(CONFIG['game']).unwrapped
optimizer = SGD(net_params, LR)
pool = mp.Pool(processes=N_CORE)
mar = None # moving average reward
for g in range(N_GENERATION):
t0 = time.time()
net_params, kid_rewards = train(net_shapes, net_params, optimizer,
utility, pool)
# test trained net without noise
net_r = get_reward(
def test_ufunc():
x = np.array([-1.2, 1.2])
np.absolute(x)
np.absolute(1.2 + 1j)
x = np.linspace(start=-10, stop=10, num=101)
np.add(1.0, 4.0)
x1 = np.arange(9.0).reshape((3, 3))
x2 = np.arange(3.0)
np.add(x1, x2)
np.arccos([1, -1])
x = np.linspace(-1, 1, num=100)
np.arccosh([np.e, 10.0])
np.arccosh(1)
np.arcsin(0)
np.arcsinh(np.array([np.e, 10.0]))
np.arctan([0, 1])
np.pi / 4
x = np.linspace(-10, 10)
x = np.array([-1, +1, +1, -1])
y = np.array([-1, -1, +1, +1])
np.arctan2(y, x) * 180 / np.pi
np.arctan2([1., -1.], [0., 0.])
np.arctan2([0., 0., np.inf], [+0., -0., np.inf])
np.arctanh([0, -0.5])
np.bitwise_and(13, 17)
np.bitwise_and(14, 13)
# np.binary_repr(12) return str
np.bitwise_and([14, 3], 13)
np.bitwise_and([11, 7], [4, 25])
np.bitwise_and(np.array([2, 5, 255]), np.array([3, 14, 16]))
np.bitwise_and([True, True], [False, True])
np.bitwise_or(13, 16)
# np.binary_repr(29)
np.bitwise_or(32, 2)
np.bitwise_or([33, 4], 1)
np.bitwise_or([33, 4], [1, 2])
np.bitwise_or(np.array([2, 5, 255]), np.array([4, 4, 4]))
# np.array([2, 5, 255]) | np.array([4, 4, 4])
np.bitwise_or(np.array([2, 5, 255, 2147483647], dtype=np.int32),
np.array([4, 4, 4, 2147483647], dtype=np.int32))
np.bitwise_or([True, True], [False, True])
np.bitwise_xor(13, 17)
# np.binary_repr(28)
np.bitwise_xor(31, 5)
np.bitwise_xor([31, 3], 5)
np.bitwise_xor([31, 3], [5, 6])
np.bitwise_xor([True, True], [False, True])
a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])
np.ceil(a)
a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])
np.trunc(a)
np.cos(np.array([0, np.pi / 2, np.pi]))
np.cosh(0)
x = np.linspace(-4, 4, 1000)
rad = np.arange(12.) * np.pi / 6
np.degrees(rad)
out = np.zeros((rad.shape))
r = np.degrees(rad, out)
# np.all(r == out) return bool
np.rad2deg(np.pi / 2)
np.divide(2.0, 4.0)
x1 = np.arange(9.0).reshape((3, 3))
x2 = np.arange(3.0)
np.divide(2, 4)
np.divide(2, 4.)
np.equal([0, 1, 3], np.arange(3))
np.equal(1, np.ones(1))
x = np.linspace(-2 * np.pi, 2 * np.pi, 100)
np.exp2([2, 3])
np.expm1(1e-10)
np.exp(1e-10) - 1
np.fabs(-1)
np.fabs([-1.2, 1.2])
a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])
np.floor(a)
np.floor_divide(7, 3)
np.floor_divide([1., 2., 3., 4.], 2.5)
np.fmod([-3, -2, -1, 1, 2, 3], 2)
np.remainder([-3, -2, -1, 1, 2, 3], 2)
np.fmod([5, 3], [2, 2.])
a = np.arange(-3, 3).reshape(3, 2)
np.fmod(a, [2, 2])
np.greater([4, 2], [2, 2])
a = np.array([4, 2])
b = np.array([2, 2])
a > b
np.greater_equal([4, 2, 1], [2, 2, 2])
np.hypot(3 * np.ones((3, 3)), 4 * np.ones((3, 3)))
np.hypot(3 * np.ones((3, 3)), [4])
np.bitwise_not is np.invert
np.invert(np.array([13], dtype=np.uint8))
# np.binary_repr(242, width=8)
np.invert(np.array([13], dtype=np.uint16))
np.invert(np.array([13], dtype=np.int8))
# np.binary_repr(-14, width=8)
np.invert(np.array([True, False]))
# np.isfinite(1)
# np.isfinite(0)
# np.isfinite(np.nan)
# np.isfinite(np.inf)
# np.isfinite(np.NINF)
x = np.array([-np.inf, 0., np.inf])
y = np.array([2, 2, 2])
np.isfinite(x, y)
# np.isinf(np.inf)
# np.isinf(np.nan)
# np.isinf(np.NINF)
# np.isinf([np.inf, -np.inf, 1.0, np.nan])
x = np.array([-np.inf, 0., np.inf])
y = np.array([2, 2, 2])
# np.isinf(x, y)
# np.isnan(np.nan)
# np.isnan(np.inf)
# np.binary_repr(5)
np.left_shift(5, 2)
# np.binary_repr(20)
np.left_shift(5, [1, 2, 3])
np.less([1, 2], [2, 2])
np.less_equal([4, 2, 1], [2, 2, 2])
x = np.array([0, 1, 2, 2**4])
xi = np.array([0 + 1.j, 1, 2 + 0.j, 4.j])
np.log2(xi)
prob1 = np.log(1e-50)
prob2 = np.log(2.5e-50)
prob12 = np.logaddexp(prob1, prob2)
prob12
np.exp(prob12)
prob1 = np.log2(1e-50)
prob2 = np.log2(2.5e-50)
prob12 = np.logaddexp2(prob1, prob2)
prob1, prob2, prob12
2**prob12
np.log1p(1e-99)
np.log(1 + 1e-99)
# np.logical_and(True, False)
# np.logical_and([True, False], [False, False])
x = np.arange(5)
# np.logical_and(x>1, x<4)
# np.logical_not(3)
# np.logical_not([True, False, 0, 1])
x = np.arange(5)
# np.logical_not(x<3)
# np.logical_or(True, False)
# np.logical_or([True, False], [False, False])
x = np.arange(5)
# np.logical_or(x < 1, x > 3)
# np.logical_xor(True, False)
# np.logical_xor([True, True, False, False], [True, False, True, False])
x = np.arange(5)
# np.logical_xor(x < 1, x > 3)
# np.logical_xor(0, np.eye(2))
np.maximum([2, 3, 4], [1, 5, 2])
# np.maximum([np.nan, 0, np.nan], [0, np.nan, np.nan])
# np.maximum(np.Inf, 1)
np.minimum([2, 3, 4], [1, 5, 2])
# np.minimum([np.nan, 0, np.nan],[0, np.nan, np.nan])
# np.minimum(-np.Inf, 1)
np.fmax([2, 3, 4], [1, 5, 2])
np.fmax(np.eye(2), [0.5, 2])
# np.fmax([np.nan, 0, np.nan],[0, np.nan, np.nan])
np.fmin([2, 3, 4], [1, 5, 2])
np.fmin(np.eye(2), [0.5, 2])
# np.fmin([np.nan, 0, np.nan],[0, np.nan, np.nan])
np.modf([0, 3.5])
np.modf(-0.5)
np.multiply(2.0, 4.0)
x1 = np.arange(9.0).reshape((3, 3))
x2 = np.arange(3.0)
np.multiply(x1, x2)
np.negative([1., -1.])
np.not_equal([1., 2.], [1., 3.])
np.not_equal([1, 2], [[1, 3], [1, 4]])
x1 = range(6)
np.power(x1, 3)
x2 = [1.0, 2.0, 3.0, 3.0, 2.0, 1.0]
np.power(x1, x2)
x2 = np.array([[1, 2, 3, 3, 2, 1], [1, 2, 3, 3, 2, 1]])
np.power(x1, x2)
deg = np.arange(12.) * 30.
np.radians(deg)
out = np.zeros((deg.shape))
ret = np.radians(deg, out)
ret is out
np.deg2rad(180)
np.reciprocal(2.)
np.reciprocal([1, 2., 3.33])
np.remainder([4, 7], [2, 3])
np.remainder(np.arange(7), 5)
# np.binary_repr(10)
np.right_shift(10, 1)
# np.binary_repr(5)
np.right_shift(10, [1, 2, 3])
a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])
np.rint(a)
np.sign([-5., 4.5])
np.sign(0)
# np.sign(5-2j)
# np.signbit(-1.2)
np.signbit(np.array([1, -2.3, 2.1]))
np.copysign(1.3, -1)
np.copysign([-1, 0, 1], -1.1)
np.copysign([-1, 0, 1], np.arange(3) - 1)
np.sin(np.pi / 2.)
np.sin(np.array((0., 30., 45., 60., 90.)) * np.pi / 180.)
x = np.linspace(-np.pi, np.pi, 201)
np.sinh(0)
# np.sinh(np.pi*1j/2)
np.sqrt([1, 4, 9])
np.sqrt([4, -1, -3 + 4J])
np.cbrt([1, 8, 27])
np.square([-1j, 1])
np.subtract(1.0, 4.0)
x1 = np.arange(9.0).reshape((3, 3))
x2 = np.arange(3.0)
np.subtract(x1, x2)
np.tan(np.array([-pi, pi / 2, pi]))
np.tanh((0, np.pi * 1j, np.pi * 1j / 2))
x = np.arange(5)
np.true_divide(x, 4)
x = np.arange(9)
y1, y2 = np.frexp(x)
y1 * 2**y2
np.ldexp(5, np.arange(4))
x = np.arange(6)
np.ldexp(*np.frexp(x))
def garchLLH(y, par):
h = garchSim(np.square(y), par)
T = y.shape[0]
llh = -0.5 * (T - 1) * math.log(
2 * math.pi) - 0.5 * np.sum(np.log(h) + (y / np.sqrt(h))**2)
return llh[0]
def KL(P, Q):
return np.sum(P * np.log(P / Q), axis=1)
