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 svm_loss(x, y):
"""
Computes the loss and gradient using for multiclass SVM 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
"""
N = x.shape[0]
correct_class_scores = x[np.arange(N), y]
#TODO: Support broadcast case: (X,) (X, Y)
#shape(x) is (d0, d1)
#shape(correct_class_scores) is (d0,)
#margins = np.maximum(0, x - correct_class_scores + 1.0)
margins = np.transpose(
np.maximum(0,
np.transpose(x) - np.transpose(correct_class_scores) + 1.0))
loss = (np.sum(margins) - np.sum(margins[np.arange(N), y])) / N
return loss
def svm_loss(x, y, mode):
"""
Computes the loss and gradient using for multiclass SVM 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
"""
if mode == 'cpu':
np.set_policy(policy.OnlyNumpyPolicy())
else:
np.set_policy(policy.PreferMXNetPolicy())
N = x.shape[0]
correct_class_scores = x[np.arange(N), y]
#margins = np.maximum(0, x - correct_class_scores[:, np.newaxis] + 1.0)
margins = np.maximum(0, x - np.expand_dims(correct_class_scores, axis = 1) + 1.0)
margins[np.arange(N), y] = 0
loss = np.sum(margins) / N
num_pos = np.sum(margins > 0, axis=1)
dx = np.zeros_like(x)
dx[margins > 0] = 1
dx[np.arange(N), y] -= num_pos
dx /= N
return loss, dx
def svm_loss(x, y):
"""
Computes the loss and gradient using for multiclass SVM 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
"""
N = x.shape[0]
correct_class_scores = x[np.arange(N), y]
#TODO: Support broadcast case: (X,) (X, Y)
#shape(x) is (d0, d1)
#shape(correct_class_scores) is (d0,)
#margins = np.maximum(0, x - correct_class_scores + 1.0)
margins = np.transpose(np.maximum(0, np.transpose(x) - np.transpose(correct_class_scores) + 1.0))
loss = (np.sum(margins) - np.sum(margins[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
- 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(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 test_lr_grad():
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def predict(weights, inputs):
return sigmoid(np.dot(inputs, weights))
def training_loss(inputs):
preds = predict(weights, inputs)
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
l = -np.sum(np.log(label_probabilities))
return l
def training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(np.argmax(preds, axis=1) - np.argmax(targets, axis=1))
return (256 - error) * 100 / 256.0
wshape = (500, 250)
weights = random.rand(*wshape) - 0.5
xshape = (256, 500)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
gradient_checker.quick_grad_check(training_loss, inputs)
def _select_polygon_points(self, num):
x = self.np_random.randint(self.screen_width / 2 - 6,
self.screen_width / 2 + 20, 1)
y = self.np_random.randint(self.screen_width / 2 - 3,
self.screen_width / 2 + 5, 1)
start = (x[0], y[0])
base = 180 / config['rotate_degree']
base_angle = math.pi / base
angle_options = (np.arange(base - 2) + 1) * base_angle
angle = 0
rotate_degree = self.np_random.choice(angle_options, 3)
size = np.array(np.random.choice(range(3, 5), 3)) * 5
#size = [3,3,4]
points = []
points.append(start)
for i in range(num - 1):
angle += rotate_degree[i]
next_x = int(start[0] + size[i] * math.cos(angle))
next_y = int(start[1] + size[i] * math.sin(angle))
next_point = (next_x, next_y)
points.append(next_point)
start = next_point
dist = points[0][0] - points[-1][0]
if dist < 0:
next_y = math.tan(angle_options[0]) * dist
else:
next_y = math.tan(-angle_options[0]) * dist
next_point = (points[0][0], int(points[-1][1] + next_y))
points.append(next_point)
return points
def test_lr_grad():
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def predict(weights, inputs):
return sigmoid(np.dot(inputs, weights))
def training_loss(inputs):
preds = predict(weights, inputs)
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
l = -np.sum(np.log(label_probabilities))
return l
def training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(
np.argmax(preds, axis=1) - np.argmax(targets, axis=1))
return (256 - error) * 100 / 256.0
wshape = (500, 250)
weights = random.rand(*wshape) - 0.5
xshape = (256, 500)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
gradient_checker.quick_grad_check(training_loss, inputs)
def test_context():
set_context(gpu(1)) # set the global context as gpu(1)
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def predict(weights, inputs):
return sigmoid(np.dot(inputs, weights))
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 training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(np.argmax(preds, axis=1) - np.argmax(targets, axis=1))
return (256 - error) * 100 / 256.0
with gpu(0):
xshape = (256, 500)
wshape = (500, 250)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = random.rand(*wshape) - 0.5
training_gradient_fun = grad(training_loss)
for i in range(20):
print('Trained loss accuracy #{}: {}%'.format(i, training_accuracy(weights, inputs)))
gr = training_gradient_fun(weights, inputs)
weights -= gr * 0.01
print("\nff and bp on {0}".format(weights.context))
print("\nexecute on cpu")
with cpu():
x_cpu = random.rand(32, 64) - 0.5
y_cpu = random.rand(64, 32) - 0.5
z_cpu = np.dot(x_cpu, y_cpu)
print('z_cpu.context = {0}'.format(z_cpu.context))
print("\nexecute on gpu(0)")
with gpu(0):
x_gpu0 = random.rand(32, 64) - 0.5
y_gpu0 = random.rand(64, 32) - 0.5
z_gpu0 = np.dot(x_gpu0, y_gpu0)
z_gpu0.asnumpy()
print('z_gpu0.context = {0}'.format(z_gpu0.context))
print("\n[use global context] execute on gpu(1)")
x_gpu1 = random.rand(32, 64) - 0.5
y_gpu1 = random.rand(64, 32) - 0.5
z_gpu1 = np.dot(x_gpu1, y_gpu1)
z_gpu1.asnumpy()
print('z_gpu1.context = {0}'.format(z_gpu1.context))
def test_index_slice():
a = np.arange(6)
a = np.reshape(a, (3, 2))
print(a[:])
a[1:2] = -1
print(a)
d = np.slice_axis(a, axis=1, begin=1, end=2)
print(d)
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 grad(g):
import numpy as np
y = label.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]
dx = probs.copy()
dx[np.arange(N), y] -= 1
dx /= N
return dx
def attack_mnist(model, alpha=0.2, beta=0.001, isTarget=False, num_attacks=10):
imgs = np.array(
nnp.load('data/mnist_data.npy')[60000:].transpose(
(0, 2, 3, 1)).astype(np.float32) / 255.)
labs = nnp.load('data/mnist_labels.npy')[60000:]
nb_labs = nnp.max(labs)
print(
"\n\n Running {} attack on {} random MNIST test images for alpha= {} beta= {}\n\n"
.format("targetted" if isTarget else "untargetted", num_attacks, alpha,
beta))
total_distortion = []
samples = []
for i in range(nb_labs + 1):
samples.append(
np.random.permutation(np.arange(len(labs))[labs == i])[0])
# samples = [6312, 6891, 4243, 8377, 7962, 6635, 4970, 7809, 5867, 9559, 3579, 8269, 2282, 4618, 2290, 1554, 4105, 9862, 2408, 5082, 1619, 1209, 5410, 7736, 9172, 1650, 5181, 3351, 9053, 7816, 7254, 8542, 4268, 1021, 8990, 231, 1529, 6535, 19, 8087, 5459, 3997, 5329, 1032, 3131, 9299, 3910, 2335, 8897, 7340, 1495, 5244,8323, 8017, 1787, 4939, 9032, 4770, 2045, 8970, 5452, 8853, 3330, 9883, 8966, 9628, 4713, 7291, 9770, 6307, 5195, 9432, 3967, 4757, 3013, 3103, 3060, 541, 4261, 7808, 1132, 1472, 2134, 634, 1315, 8858, 6411, 8595, 4516, 8550, 3859, 3526]
#true_labels = [3, 1, 6, 6, 9, 2, 7, 5, 5, 3, 3, 4, 5, 6, 7, 9, 1, 6, 3, 4, 0, 6, 5, 9, 7, 0, 3, 1, 6, 6, 9, 6, 4, 7, 6, 3, 4, 3, 4, 3, 0, 7, 3, 5, 3, 9, 3, 1, 9, 1, 3, 0, 2, 9, 9, 2, 2, 3, 3, 3, 0, 5, 2, 5, 2, 7, 2, 2, 5, 7, 4, 9, 9, 0, 0, 7, 9, 4, 5, 5, 2, 3, 5, 9, 3, 0, 9, 0, 1, 2, 9, 9]
for idx in samples:
#idx = random.randint(100, len(test_dataset)-1)
image, label = imgs[idx], labs[idx]
print("\n\n\n\n======== Image %d =========" % idx)
print("Original label: ", label)
lab = predict(model, image)
print("Predicted label: ", lab)
if lab != label:
print(
'CHANGE IMAGES#{}: prediction of original image is not the same with true label'
.format(i))
continue
#target = None if not isTarget else random.choice(list(range(label)) + list(range(label+1, 10)))
advs = [image]
for i in range(nb_labs):
target = (label + i) % (nb_labs + 1)
adv = attack_targeted(model,
imgs[labs == target],
image,
label,
target,
alpha=alpha,
beta=beta,
iterations=1000)
print(i, "Predicted label for adversarial example: ",
predict(model, adversarial))
advs.append(np.clip(adv, 0, 1))
total_distortion.append(
np.linalg.norm(adv.reshape(-1) - image.reshape(-1)))
np.save('advs/opt_attacks_{}_show.npy'.format(idx), advs)
print("Average distortion on random {} images is {}".format(
len(total_distortion), np.mean(total_distortion)))
def svm_loss(x, y, mode):
"""
Computes the loss and gradient using for multiclass SVM 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
"""
if mode == 'cpu':
np.set_policy(policy.OnlyNumpyPolicy())
else:
np.set_policy(policy.PreferMXNetPolicy())
N = x.shape[0]
correct_class_scores = x[np.arange(N), y]
#TODO: Support broadcast case: (X,) (X, Y)
#margins = np.maximum(0, x - correct_class_scores + 1.0)
margins = np.transpose(
np.maximum(0,
np.transpose(x) - np.transpose(correct_class_scores) + 1.0))
#margins[np.arange(N), y] = 0
#loss = np.sum(margins) / N
loss = (np.sum(margins) - np.sum(margins[np.arange(N), y])) / N
margins[np.arange(N), y] = 0
num_pos = np.sum(margins > 0, axis=1)
dx = np.zeros_like(x)
dx[margins > 0] = 1
dx[np.arange(N), y] -= num_pos
dx /= N
return loss, dx
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 predict(models, img, t=0):
img = np.clip(img, 0, 1) * 255
img = extend_data(config['permutation'], np.array([img]))
scores = np.hstack([m.predict(img) for m in models])[0]
#print(scores.shape)
nat_labels = np.zeros(scores.shape).astype(np.float32)
nat_labels[scores >= 0.5] = 1.
rep = rep_labels[:len(scores)].T
tmp = np.repeat([nat_labels], rep.shape[0], axis=0)
dists = np.sum(np.absolute(tmp - rep), axis=-1)
min_dist = np.min(dists)
pred_labels = np.arange(len(dists))[dists == min_dist]
pred_scores = [
np.sum([
scores[k] if rep[j][k] == 1 else 1 - scores[k]
for k in np.arange(len(scores))
]) for j in pred_labels
]
pred_label = pred_labels[np.argmax(pred_scores)]
if min_dist <= 0:
return pred_label
else:
return -1
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 test_op_statistics():
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def predict(weights, inputs):
return sigmoid(np.dot(inputs, weights))
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 training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(
np.argmax(
preds, axis=1) - np.argmax(
targets, axis=1))
return (256 - error) * 100 / 256.0
np.record_op_stat()
xshape = (256, 500)
wshape = (500, 250)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = random.rand(*wshape) - 0.5
training_gradient_fun = grad(training_loss)
for i in range(30):
print('Trained accuracy #{}: {}%'.format(i, training_accuracy(weights,
inputs)))
gr = training_gradient_fun(weights, inputs)
weights -= gr * 0.01
# Print Op Statistics Info
np.show_op_stat()
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 test_lr_grad():
inputs = rng.rand(32, 64) * 0.1
targets = np.zeros((32, 10))
truth = rng.randint(0, 10, 32)
targets[np.arange(32), truth] = 1
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def training_loss(weights):
preds = sigmoid(np.dot(inputs, weights))
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
l = -np.sum(np.log(label_probabilities))
return l
weights = rng.rand(64, 10) * 0.01
return gradient_checker.quick_grad_check(training_loss, weights, rs=rng)
def test_lr_grad():
inputs = rng.rand(32, 64) * 0.1
targets = np.zeros((32, 10))
truth = rng.randint(0, 10, 32)
targets[np.arange(32), truth] = 1
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def training_loss(weights):
preds = sigmoid(np.dot(inputs, weights))
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
l = -np.sum(np.log(label_probabilities))
return l
weights = rng.rand(64, 10) * 0.01
return gradient_checker.quick_grad_check(training_loss, weights, rs=rng)
def test_mxnet_logistic():
def sigmoid(x):
return np.multiply(0.5, np.add(np.tanh(x), 1))
xshape = (256, 500)
#needs to reverse. because of mxnet's setting
wshape = (250, 500)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = np.random.rand(*wshape) - 0.5
x = mx.sym.Variable(name='x')
fc = mx.sym.FullyConnected(name='fc', data=x, num_hidden=250)
act = mx.sym.Activation(data=fc, act_type='sigmoid')
f = core.Function(act, {'x': xshape})
def predict(weights, inputs):
#return f( data=[('x', inputs)], weight=[('fc_weight', weights)], ctx=mx.cpu())
return f(x=inputs, fc_weight=weights)
def training_loss(weights, inputs):
preds = predict(weights, inputs)
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
return -np.sum(np.log(label_probabilities))
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
if i % 10 == 0:
print('Trained loss: {}'.format(training_loss(weights, inputs)))
# The training loss should be around 300 in a bug-free Minpy
if (training_loss(weights, inputs)[0] > 600):
assert (False)
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 test_logistic():
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def predict(weights, inputs):
return sigmoid(np.dot(inputs, weights))
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 training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(
np.argmax(
preds, axis=1) - np.argmax(
targets, axis=1))
return (256 - error) * 100 / 256.0
xshape = (256, 500)
wshape = (500, 250)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = random.rand(*wshape) - 0.5
training_gradient_fun = grad(training_loss)
for i in range(200):
print('Trained accuracy #{}: {}%'.format(i, training_accuracy(weights,
inputs)))
gr = training_gradient_fun(weights, inputs)
weights -= gr * 0.01
def test_slice():
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def predict(weights, inputs):
# Test Slice
sliced_weights = weights[:, ::2]
y = sigmoid(np.dot(inputs, sliced_weights))
return y
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 training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(
np.argmax(preds, axis=1) - np.argmax(targets, axis=1))
return (256 - error) * 100 / 256.0
xshape = (256, 500)
# wshape = (500, 250)
wshape = (500, 500)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = random.rand(*wshape) - 0.5
training_gradient_fun = grad(training_loss)
for i in range(20):
print('Trained loss accuracy #{}: {}%'.format(
i, training_accuracy(weights, inputs)))
gr = training_gradient_fun(weights, inputs)
print('Gradient Size', gr.shape)
print('Gradient example', gr[0, :10].asnumpy())
weights -= gr * 0.01
def test_logistic():
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def predict(weights, inputs):
return sigmoid(np.dot(inputs, weights))
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 training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(
np.argmax(
preds, axis=1) - np.argmax(
targets, axis=1))
return (256 - error) * 100 / 256.0
xshape = (256, 500)
wshape = (500, 250)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = random.rand(*wshape) - 0.5
training_gradient_fun = grad(training_loss)
for i in range(200):
print('Trained accuracy #{}: {}%'.format(
i, training_accuracy(weights, inputs)))
gr = training_gradient_fun(weights, inputs)
weights -= gr * 0.01
# The accuracy should be 100 in bug-free MinPy
if (training_accuracy(weights, inputs) < 95):
assert (False)
def forward(X,y,*p): N, C, H, W = X.shape X = X.reshape((N,C*H*W)) print '>>',X.shape print '>>',p[0].shape first = np.dot( X, p[0] ) + p[1] second = np.dot( first, p[2] ) + p[3] exp = np.exp(second) pred = exp / np.sum(exp) N = X.shape[0] loss = -np.sum( pred[np.arange(N),y] ) return loss
def test_slice():
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def predict(weights, inputs):
# Test Slice
sliced_weights = weights[:, ::2]
y = sigmoid(np.dot(inputs, sliced_weights))
return y
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 training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(np.argmax(preds, axis=1) - np.argmax(targets, axis=1))
return (256 - error) * 100 / 256.0
xshape = (256, 500)
# wshape = (500, 250)
wshape = (500, 500)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = random.rand(*wshape) - 0.5
training_gradient_fun = grad(training_loss)
for i in range(20):
print('Trained loss accuracy #{}: {}%'.format(i, training_accuracy(weights, inputs)))
gr = training_gradient_fun(weights, inputs)
print('Gradient Size', gr.shape)
print('Gradient example', gr[0,:10].asnumpy())
weights -= gr * 0.01
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
preds = predict(weights, bias, inputs)
label_probabilities = preds * targets + (1 - preds) * (1 - targets)
l = -np.sum(np.log(label_probabilities))
return l
def training_accuracy(weights, bias, inputs):
preds = predict(weights, bias, inputs)
error = np.count_nonzero(np.argmax(preds, axis=1) - np.argmax(targets, axis=1))
return (256 - error) * 100 / 256.0
xshape = (256, 500)
wshape = (500, 250)
bshape = (250)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = random.rand(*wshape) - 0.5
#bias = random.rand(bshape) - 0.5
#print bias.shape
bias = np.zeros(bshape)
print bias.shape
training_gradient_fun = grad(training_loss)
for i in range(20):
print('Trained loss accuracy #{}: {}%'.format(i, training_accuracy(weights, bias, inputs)))
gr = training_gradient_fun(weights, bias, inputs)
weights -= gr * 0.01
def identity(n):
array = np.zeros((n, n))
array[np.arange(n), np.arange(n)] = 1
return array
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 test_numeric():
# 'newaxis', 'ndarray', 'flatiter', 'nditer', 'nested_iters', 'ufunc',
# 'arange', 'array', 'zeros', 'count_nonzero', 'empty', 'broadcast',
# 'dtype', 'fromstring', 'fromfile', 'frombuffer', 'int_asbuffer',
# 'where', 'argwhere', 'copyto', 'concatenate', 'fastCopyAndTranspose',
# 'lexsort', 'set_numeric_ops', 'can_cast', 'promote_types',
# 'min_scalar_type', 'result_type', 'asarray', 'asanyarray',
# 'ascontiguousarray', 'asfortranarray', 'isfortran', 'empty_like',
# 'zeros_like', 'ones_like', 'correlate', 'convolve', 'inner', 'dot',
# 'einsum', 'outer', 'vdot', 'alterdot', 'restoredot', 'roll',
# 'rollaxis', 'moveaxis', 'cross', 'tensordot', 'array2string',
# 'get_printoptions', 'set_printoptions', 'array_repr', 'array_str',
# 'set_string_function', 'little_endian', 'require', 'fromiter',
# 'array_equal', 'array_equiv', 'indices', 'fromfunction', 'isclose', 'load',
# 'loads', 'isscalar', 'binary_repr', 'base_repr', 'ones', 'identity',
# 'allclose', 'compare_chararrays', 'putmask', 'seterr', 'geterr',
# 'setbufsize', 'getbufsize', 'seterrcall', 'geterrcall', 'errstate',
# 'flatnonzero', 'Inf', 'inf', 'infty', 'Infinity', 'nan', 'NaN', 'False_',
# 'True_', 'bitwise_not', 'full', 'full_like', 'matmul'
x = np.arange(6)
x = x.reshape((2, 3))
np.zeros_like(x)
y = np.arange(3, dtype=np.float)
np.zeros_like(y)
np.ones(5)
np.ones((5,), dtype=np.int)
np.ones((2, 1))
s = (2,2)
np.ones(s)
x = np.arange(6)
x = x.reshape((2, 3))
np.ones_like(x)
y = np.arange(3, dtype=np.float)
np.ones_like(y)
np.full((2, 2), np.inf)
x = np.arange(6, dtype=np.int)
np.full_like(x, 1)
np.full_like(x, 0.1)
np.full_like(y, 0.1)
np.count_nonzero(np.eye(4))
np.count_nonzero([[0,1,7,0,0],[3,0,0,2,19]])
np.count_nonzero([[0,1,7,0,0],[3,0,0,2,19]], axis=0)
np.count_nonzero([[0,1,7,0,0],[3,0,0,2,19]], axis=1)
a = [1, 2]
np.asarray(a)
a = np.array([1, 2])
np.asarray(a) is a
a = np.array([1, 2], dtype=np.float32)
np.asarray(a, dtype=np.float32) is a
np.asarray(a, dtype=np.float64) is a
np.asarray(a) is a
np.asanyarray(a) is a
a = [1, 2]
np.asanyarray(a)
np.asanyarray(a) is a
x = np.arange(6).reshape(2,3)
np.ascontiguousarray(x, dtype=np.float32)
x = np.arange(6).reshape(2,3)
y = np.asfortranarray(x)
x = np.arange(6).reshape(2,3)
y = np.require(x, dtype=np.float32, requirements=['A', 'O', 'W', 'F'])
a = np.array([[1, 2, 3], [4, 5, 6]], order='C')
np.isfortran(a)
b = np.array([[1, 2, 3], [4, 5, 6]], order='FORTRAN')
np.isfortran(b)
a = np.array([[1, 2, 3], [4, 5, 6]], order='C')
np.isfortran(a)
b = a.T
np.isfortran(b)
np.isfortran(np.array([1, 2], order='FORTRAN'))
x = np.arange(6).reshape(2,3)
np.argwhere(x>1)
x = np.arange(-2, 3)
np.flatnonzero(x)
np.correlate([1, 2, 3], [0, 1, 0.5])
np.correlate([1, 2, 3], [0, 1, 0.5], "same")
np.correlate([1, 2, 3], [0, 1, 0.5], "full")
np.correlate([1+1j, 2, 3-1j], [0, 1, 0.5j], 'full')
np.correlate([0, 1, 0.5j], [1+1j, 2, 3-1j], 'full')
np.convolve([1, 2, 3], [0, 1, 0.5])
np.convolve([1,2,3],[0,1,0.5], 'same')
np.convolve([1,2,3],[0,1,0.5], 'valid')
rl = np.outer(np.ones((5,)), np.linspace(-2, 2, 5))
# im = np.outer(1j*np.linspace(2, -2, 5), np.ones((5,)))
# grid = rl + im
x = np.array(['a', 'b', 'c'], dtype=object)
np.outer(x, [1, 2, 3])
a = np.arange(60.).reshape(3,4,5)
b = np.arange(24.).reshape(4,3,2)
c = np.tensordot(a,b, axes=([1,0],[0,1]))
c.shape
# A slower but equivalent way of computing the same...
d = np.zeros((5,2))
a = np.array(range(1, 9))
A = np.array(('a', 'b', 'c', 'd'), dtype=object)
x = np.arange(10)
np.roll(x, 2)
x2 = np.reshape(x, (2,5))
np.roll(x2, 1)
np.roll(x2, 1, axis=0)
np.roll(x2, 1, axis=1)
a = np.ones((3,4,5,6))
np.rollaxis(a, 3, 1).shape
np.rollaxis(a, 2).shape
np.rollaxis(a, 1, 4).shape
x = np.zeros((3, 4, 5))
np.moveaxis(x, 0, -1).shape
np.moveaxis(x, -1, 0).shape
np.transpose(x).shape
np.moveaxis(x, [0, 1], [-1, -2]).shape
np.moveaxis(x, [0, 1, 2], [-1, -2, -3]).shape
x = [1, 2, 3]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2, 0]
y = [4, 5, 6]
np.cross(x, y)
x = [1,2]
y = [4,5]
np.cross(x, y)
x = np.array([[1,2,3], [4,5,6]])
y = np.array([[4,5,6], [1,2,3]])
np.cross(x, y)
np.cross(x, y, axisc=0)
x = np.array([[1,2,3], [4,5,6], [7, 8, 9]])
y = np.array([[7, 8, 9], [4,5,6], [1,2,3]])
np.cross(x, y)
np.cross(x, y, axisa=0, axisb=0)
# np.array_repr(np.array([1,2]))
# np.array_repr(np.ma.array([0.]))
# np.array_repr(np.array([], np.int32))
x = np.array([1e-6, 4e-7, 2, 3])
# np.array_repr(x, precision=6, suppress_small=True)
# np.array_str(np.arange(3))
a = np.arange(10)
x = np.arange(4)
np.set_string_function(lambda x:'random', repr=False)
grid = np.indices((2, 3))
grid.shape
grid[0] # row indices
grid[1] # column indices
x = np.arange(20).reshape(5, 4)
row, col = np.indices((2, 3))
x[row, col]
np.fromfunction(lambda i, j: i == j, (3, 3), dtype=int)
np.fromfunction(lambda i, j: i + j, (3, 3), dtype=int)
np.isscalar(3.1)
np.isscalar([3.1])
np.isscalar(False)
# np.binary_repr(3)
# np.binary_repr(-3)
# np.binary_repr(3, width=4)
# np.binary_repr(-3, width=3)
# np.binary_repr(-3, width=5)
# np.base_repr(5)
# np.base_repr(6, 5)
# np.base_repr(7, base=5, padding=3)
# np.base_repr(10, base=16)
# np.base_repr(32, base=16)
np.identity(3)
np.allclose([1e10,1e-7], [1.00001e10,1e-8])
np.allclose([1e10,1e-8], [1.00001e10,1e-9])
np.allclose([1e10,1e-8], [1.0001e10,1e-9])
# np.allclose([1.0, np.nan], [1.0, np.nan])
# np.allclose([1.0, np.nan], [1.0, np.nan], equal_nan=True)
np.isclose([1e10,1e-7], [1.00001e10,1e-8])
np.isclose([1e10,1e-8], [1.00001e10,1e-9])
np.isclose([1e10,1e-8], [1.0001e10,1e-9])
# np.isclose([1.0, np.nan], [1.0, np.nan])
# np.isclose([1.0, np.nan], [1.0, np.nan], equal_nan=True)
np.array_equal([1, 2], [1, 2])
np.array_equal(np.array([1, 2]), np.array([1, 2]))
np.array_equal([1, 2], [1, 2, 3])
np.array_equal([1, 2], [1, 4])
np.array_equiv([1, 2], [1, 2])
np.array_equiv([1, 2], [1, 3])
np.array_equiv([1, 2], [[1, 2], [1, 2]])
np.array_equiv([1, 2], [[1, 2, 1, 2], [1, 2, 1, 2]])
np.array_equiv([1, 2], [[1, 2], [1, 3]])
def test_fromnumeric():
# Functions
# 'alen', 'all', 'alltrue', 'amax', 'amin', 'any', 'argmax',
# 'argmin', 'argpartition', 'argsort', 'around', 'choose', 'clip',
# 'compress', 'cumprod', 'cumproduct', 'cumsum', 'diagonal', 'mean',
# 'ndim', 'nonzero', 'partition', 'prod', 'product', 'ptp', 'put',
# 'rank', 'ravel', 'repeat', 'reshape', 'resize', 'round_',
# 'searchsorted', 'shape', 'size', 'sometrue', 'sort', 'squeeze',
# 'std', 'sum', 'swapaxes', 'take', 'trace', 'transpose', 'var',
a = [4, 3, 5, 7, 6, 8]
indices = [0, 1, 4]
np.take(a, indices)
a = np.array(a)
# a[indices]
np.take(a, [[0, 1], [2, 3]])
a = np.zeros((10, 2))
b = a.T
a = np.arange(6).reshape((3, 2))
np.reshape(a, (2, 3)) # C-like index ordering
np.reshape(np.ravel(a), (2, 3)) # equivalent to C ravel then C reshape
np.reshape(a, (2, 3), order='F') # Fortran-like index ordering
np.reshape(np.ravel(a, order='F'), (2, 3), order='F')
a = np.array([[1,2,3], [4,5,6]])
np.reshape(a, 6)
np.reshape(a, 6, order='F')
np.reshape(a, (3,-1)) # the unspecified value is inferred to be 2
choices = [[0, 1, 2, 3], [10, 11, 12, 13],
[20, 21, 22, 23], [30, 31, 32, 33]]
np.choose([2, 3, 1, 0], choices)
np.choose([2, 4, 1, 0], choices, mode='clip') # 4 goes to 3 (4-1)
np.choose([2, 4, 1, 0], choices, mode='wrap') # 4 goes to (4 mod 4)
a = [[1, 0, 1], [0, 1, 0], [1, 0, 1]]
choices = [-10, 10]
np.choose(a, choices)
a = np.array([0, 1]).reshape((2,1,1))
c1 = np.array([1, 2, 3]).reshape((1,3,1))
c2 = np.array([-1, -2, -3, -4, -5]).reshape((1,1,5))
np.choose(a, (c1, c2)) # result is 2x3x5, res[0,:,:]=c1, res[1,:,:]=c2
np.repeat(3, 4)
x = np.array([[1,2],[3,4]])
np.repeat(x, 2)
np.repeat(x, 3, axis=1)
np.repeat(x, [1, 2], axis=0)
a = np.arange(5)
np.put(a, [0, 2], [-44, -55])
a = np.arange(5)
np.put(a, 22, -5, mode='clip')
x = np.array([[1,2,3]])
np.swapaxes(x,0,1)
x = np.array([[[0,1],[2,3]],[[4,5],[6,7]]])
np.swapaxes(x,0,2)
x = np.arange(4).reshape((2,2))
np.transpose(x)
x = np.ones((1, 2, 3))
np.transpose(x, (1, 0, 2)).shape
a = np.array([3, 4, 2, 1])
np.partition(a, 3)
np.partition(a, (1, 3))
x = np.array([3, 4, 2, 1])
x[np.argpartition(x, 3)]
x[np.argpartition(x, (1, 3))]
x = [3, 4, 2, 1]
np.array(x)[np.argpartition(x, 3)]
a = np.array([[1,4],[3,1]])
np.sort(a) # sort along the last axis
np.sort(a, axis=None) # sort the flattened array
np.sort(a, axis=0) # sort along the first axis
dtype = [('name', 'S10'), ('height', float), ('age', int)]
values = [('Arthur', 1.8, 41), ('Lancelot', 1.9, 38),
('Galahad', 1.7, 38)]
a = np.array(values, dtype=dtype) # create a structured array
np.sort(a, order='height') # doctest: +SKIP
np.sort(a, order=['age', 'height']) # doctest: +SKIP
x = np.array([3, 1, 2])
np.argsort(x)
x = np.array([[0, 3], [2, 2]])
np.argsort(x, axis=0)
np.argsort(x, axis=1)
x = np.array([(1, 0), (0, 1)], dtype=[('x', '<i4'), ('y', '<i4')])
np.argsort(x, order=('x','y'))
np.argsort(x, order=('y','x'))
a = np.arange(6).reshape(2,3)
np.argmax(a)
np.argmax(a, axis=0)
np.argmax(a, axis=1)
b = np.arange(6)
b[1] = 5
np.argmax(b) # Only the first occurrence is returned.
a = np.arange(6).reshape(2,3)
np.argmin(a)
np.argmin(a, axis=0)
np.argmin(a, axis=1)
b = np.arange(6)
b[4] = 0
np.argmin(b) # Only the first occurrence is returned.
np.searchsorted([1,2,3,4,5], 3)
np.searchsorted([1,2,3,4,5], 3, side='right')
np.searchsorted([1,2,3,4,5], [-10, 10, 2, 3])
a=np.array([[0,1],[2,3]])
np.resize(a,(2,3))
np.resize(a,(1,4))
np.resize(a,(2,4))
x = np.array([[[0], [1], [2]]])
x.shape
np.squeeze(x).shape
np.squeeze(x, axis=(2,)).shape
a = np.arange(4).reshape(2,2)
a = np.arange(8).reshape(2,2,2); a
a[:,:,0] # main diagonal is [0 6]
a[:,:,1] # main diagonal is [1 7]
np.trace(np.eye(3))
a = np.arange(8).reshape((2,2,2))
np.trace(a)
a = np.arange(24).reshape((2,2,2,3))
np.trace(a).shape
x = np.array([[1, 2, 3], [4, 5, 6]])
np.ravel(x)
x.reshape(-1)
np.ravel(x, order='F')
np.ravel(x.T)
np.ravel(x.T, order='A')
a = np.arange(3)[::-1]; a
# a = np.arange(12).reshape(2,3,2).swapaxes(1,2); a
x = np.eye(3)
np.nonzero(x)
x[np.nonzero(x)]
np.transpose(np.nonzero(x))
a = np.array([[1,2,3],[4,5,6],[7,8,9]])
a > 3
np.nonzero(a > 3)
np.shape(np.eye(3))
np.shape([[1, 2]])
np.shape([0])
np.shape(0)
a = np.array([(1, 2), (3, 4)], dtype=[('x', 'i4'), ('y', 'i4')])
np.shape(a)
a.shape
a = np.array([[1, 2], [3, 4], [5, 6]])
np.compress([0, 1], a, axis=0)
np.compress([False, True, True], a, axis=0)
np.compress([False, True], a, axis=1)
np.compress([False, True], a)
a = np.arange(10)
np.clip(a, 1, 8)
np.clip(a, 3, 6, out=a)
a = np.arange(10)
np.clip(a, [3,4,1,1,1,4,4,4,4,4], 8)
np.sum([])
np.sum([0.5, 1.5])
np.sum([0.5, 0.7, 0.2, 1.5], dtype=np.int32)
np.sum([[0, 1], [0, 5]])
np.sum([[0, 1], [0, 5]], axis=0)
np.sum([[0, 1], [0, 5]], axis=1)
# np.ones(128, dtype=np.int8).sum(dtype=np.int8)
# np.any([[True, False], [True, True]])
# np.any([[True, False], [False, False]], axis=0)
# np.any([-1, 0, 5])
# np.any(np.nan)
# np.all([[True,False],[True,True]])
# np.all([[True,False],[True,True]], axis=0)
# np.all([-1, 4, 5])
# np.all([1.0, np.nan])
a = np.array([[1,2,3], [4,5,6]])
np.cumsum(a)
np.cumsum(a, dtype=float) # specifies type of output value(s)
np.cumsum(a,axis=0) # sum over rows for each of the 3 columns
np.cumsum(a,axis=1) # sum over columns for each of the 2 rows
x = np.arange(4).reshape((2,2))
np.ptp(x, axis=0)
np.ptp(x, axis=1)
a = np.arange(4).reshape((2,2))
np.amax(a) # Maximum of the flattened array
np.amax(a, axis=0) # Maxima along the first axis
np.amax(a, axis=1) # Maxima along the second axis
b = np.arange(5, dtype=np.float)
# b[2] = np.NaN
np.amax(b)
np.nanmax(b)
a = np.arange(4).reshape((2,2))
np.amin(a) # Minimum of the flattened array
np.amin(a, axis=0) # Minima along the first axis
np.amin(a, axis=1) # Minima along the second axis
b = np.arange(5, dtype=np.float)
# b[2] = np.NaN
np.amin(b)
np.nanmin(b)
a = np.zeros((7,4,5))
a.shape[0]
np.alen(a)
x = np.array([536870910, 536870910, 536870910, 536870910])
np.prod(x) #random
np.prod([])
np.prod([1.,2.])
np.prod([[1.,2.],[3.,4.]])
np.prod([[1.,2.],[3.,4.]], axis=1)
x = np.array([1, 2, 3], dtype=np.uint8)
# np.prod(x).dtype == np.uint
x = np.array([1, 2, 3], dtype=np.int8)
# np.prod(x).dtype == np.int
a = np.array([1,2,3])
np.cumprod(a) # intermediate results 1, 1*2
a = np.array([[1, 2, 3], [4, 5, 6]])
np.cumprod(a, dtype=float) # specify type of output
np.cumprod(a, axis=0)
np.cumprod(a,axis=1)
np.ndim([[1,2,3],[4,5,6]])
np.ndim(np.array([[1,2,3],[4,5,6]]))
np.ndim(1)
a = np.array([[1,2,3],[4,5,6]])
np.size(a)
np.size(a,1)
np.size(a,0)
np.around([0.37, 1.64])
np.around([0.37, 1.64], decimals=1)
np.around([.5, 1.5, 2.5, 3.5, 4.5]) # rounds to nearest even value
np.around([1,2,3,11], decimals=1) # ndarray of ints is returned
np.around([1,2,3,11], decimals=-1)
a = np.array([[1, 2], [3, 4]])
np.mean(a)
np.mean(a, axis=0)
np.mean(a, axis=1)
a = np.zeros((2, 512*512), dtype=np.float32)
a[0, :] = 1.0
a[1, :] = 0.1
np.mean(a)
np.mean(a, dtype=np.float64)
a = np.array([[1, 2], [3, 4]])
np.std(a)
np.std(a, axis=0)
np.std(a, axis=1)
a = np.zeros((2, 512*512), dtype=np.float32)
a[0, :] = 1.0
a[1, :] = 0.1
np.std(a)
np.std(a, dtype=np.float64)
a = np.array([[1, 2], [3, 4]])
np.var(a)
np.var(a, axis=0)
np.var(a, axis=1)
a = np.zeros((2, 512*512), dtype=np.float32)
a[0, :] = 1.0
a[1, :] = 0.1
np.var(a)
np.var(a, dtype=np.float64)
def string_to_one_hot(string, maxchar):
"""Converts an ASCII string to a one-of-k encoding."""
ascii = np.array([ord(c) for c in string]).T
return np.array(ascii[:, None] == np.arange(maxchar)[None, :], dtype=int)
l = -np.sum(np.log(label_probabilities))
return l
def training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(np.argmax(preds, axis=1) - np.argmax(targets, axis=1))
return (256 - error) * 100 / 256.0
xshape = (256, 500)
# wshape = (500, 250)
wshape = (500, 500)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = random.rand(*wshape) - 0.5
training_gradient_fun = grad(training_loss)
def NumpyVarToMinpy(var):
return minpy.array.Value.wrap(var)
def MinpyVarToNumpy(var):
return minpy.array.Value.wrap(var).get_data(ArrayType.NUMPY)
for i in range(20):
print('Trained loss accuracy #{}: {}%'.format(i, training_accuracy(weights, inputs)))
gr = training_gradient_fun(weights, inputs)
print('Gradient Size', gr.shape)
print('Gradient example', MinpyVarToNumpy(gr[0,:10]))
caffe.TanhLayer(),
caffe.FullyConnectedLayer(10),
caffe.SoftMaxLayer()]
# Training parameters
param_scale = 0.1
learning_rate = 1e-3
momentum = 0.9
batch_size = 256
num_epochs = 50
# Load and process MNIST data (borrowing from Kayak)
print("Loading training data...")
import imp, urllib
add_color_channel = lambda x : x.reshape((x.shape[0], 1, x.shape[1], x.shape[2]))
one_hot = lambda x, K : np.array(x[:,None] == np.arange(K)[None, :], dtype=int)
source, _ = urllib.urlretrieve(
'https://raw.githubusercontent.com/HIPS/Kayak/master/examples/data.py')
data = imp.load_source('data', source).mnist()
train_images, train_labels, test_images, test_labels = data
train_images = add_color_channel(train_images) / 255.0
test_images = add_color_channel(test_images) / 255.0
train_labels = one_hot(train_labels, 10)
test_labels = one_hot(test_labels, 10)
N_data = train_images.shape[0]
# Make neural net functions
N_weights, pred_fun, loss_fun, frac_err = make_nn_funs(input_shape, caffe_layer_specs, L2_reg)
loss_grad = grad(loss_fun)
# Initialize weights
def test_numeric():
# 'newaxis', 'ndarray', 'flatiter', 'nditer', 'nested_iters', 'ufunc',
# 'arange', 'array', 'zeros', 'count_nonzero', 'empty', 'broadcast',
# 'dtype', 'fromstring', 'fromfile', 'frombuffer', 'int_asbuffer',
# 'where', 'argwhere', 'copyto', 'concatenate', 'fastCopyAndTranspose',
# 'lexsort', 'set_numeric_ops', 'can_cast', 'promote_types',
# 'min_scalar_type', 'result_type', 'asarray', 'asanyarray',
# 'ascontiguousarray', 'asfortranarray', 'isfortran', 'empty_like',
# 'zeros_like', 'ones_like', 'correlate', 'convolve', 'inner', 'dot',
# 'einsum', 'outer', 'vdot', 'alterdot', 'restoredot', 'roll',
# 'rollaxis', 'moveaxis', 'cross', 'tensordot', 'array2string',
# 'get_printoptions', 'set_printoptions', 'array_repr', 'array_str',
# 'set_string_function', 'little_endian', 'require', 'fromiter',
# 'array_equal', 'array_equiv', 'indices', 'fromfunction', 'isclose', 'load',
# 'loads', 'isscalar', 'binary_repr', 'base_repr', 'ones', 'identity',
# 'allclose', 'compare_chararrays', 'putmask', 'seterr', 'geterr',
# 'setbufsize', 'getbufsize', 'seterrcall', 'geterrcall', 'errstate',
# 'flatnonzero', 'Inf', 'inf', 'infty', 'Infinity', 'nan', 'NaN', 'False_',
# 'True_', 'bitwise_not', 'full', 'full_like', 'matmul'
x = np.arange(6)
x = x.reshape((2, 3))
np.zeros_like(x)
y = np.arange(3, dtype=np.float)
np.zeros_like(y)
np.ones(5)
np.ones((5, ), dtype=np.int)
np.ones((2, 1))
s = (2, 2)
np.ones(s)
x = np.arange(6)
x = x.reshape((2, 3))
np.ones_like(x)
y = np.arange(3, dtype=np.float)
np.ones_like(y)
np.full((2, 2), np.inf)
x = np.arange(6, dtype=np.int)
np.full_like(x, 1)
np.full_like(x, 0.1)
np.full_like(y, 0.1)
np.count_nonzero(np.eye(4))
np.count_nonzero([[0, 1, 7, 0, 0], [3, 0, 0, 2, 19]])
np.count_nonzero([[0, 1, 7, 0, 0], [3, 0, 0, 2, 19]], axis=0)
np.count_nonzero([[0, 1, 7, 0, 0], [3, 0, 0, 2, 19]], axis=1)
a = [1, 2]
np.asarray(a)
a = np.array([1, 2])
np.asarray(a) is a
a = np.array([1, 2], dtype=np.float32)
np.asarray(a, dtype=np.float32) is a
np.asarray(a, dtype=np.float64) is a
np.asarray(a) is a
np.asanyarray(a) is a
a = [1, 2]
np.asanyarray(a)
np.asanyarray(a) is a
x = np.arange(6).reshape(2, 3)
np.ascontiguousarray(x, dtype=np.float32)
x = np.arange(6).reshape(2, 3)
y = np.asfortranarray(x)
x = np.arange(6).reshape(2, 3)
y = np.require(x, dtype=np.float32, requirements=['A', 'O', 'W', 'F'])
a = np.array([[1, 2, 3], [4, 5, 6]], order='C')
np.isfortran(a)
b = np.array([[1, 2, 3], [4, 5, 6]], order='FORTRAN')
np.isfortran(b)
a = np.array([[1, 2, 3], [4, 5, 6]], order='C')
np.isfortran(a)
b = a.T
np.isfortran(b)
np.isfortran(np.array([1, 2], order='FORTRAN'))
x = np.arange(6).reshape(2, 3)
np.argwhere(x > 1)
x = np.arange(-2, 3)
np.flatnonzero(x)
np.correlate([1, 2, 3], [0, 1, 0.5])
np.correlate([1, 2, 3], [0, 1, 0.5], "same")
np.correlate([1, 2, 3], [0, 1, 0.5], "full")
np.correlate([1 + 1j, 2, 3 - 1j], [0, 1, 0.5j], 'full')
np.correlate([0, 1, 0.5j], [1 + 1j, 2, 3 - 1j], 'full')
np.convolve([1, 2, 3], [0, 1, 0.5])
np.convolve([1, 2, 3], [0, 1, 0.5], 'same')
np.convolve([1, 2, 3], [0, 1, 0.5], 'valid')
rl = np.outer(np.ones((5, )), np.linspace(-2, 2, 5))
# im = np.outer(1j*np.linspace(2, -2, 5), np.ones((5,)))
# grid = rl + im
x = np.array(['a', 'b', 'c'], dtype=object)
np.outer(x, [1, 2, 3])
a = np.arange(60.).reshape(3, 4, 5)
b = np.arange(24.).reshape(4, 3, 2)
c = np.tensordot(a, b, axes=([1, 0], [0, 1]))
c.shape
# A slower but equivalent way of computing the same...
d = np.zeros((5, 2))
a = np.array(range(1, 9))
A = np.array(('a', 'b', 'c', 'd'), dtype=object)
x = np.arange(10)
np.roll(x, 2)
x2 = np.reshape(x, (2, 5))
np.roll(x2, 1)
np.roll(x2, 1, axis=0)
np.roll(x2, 1, axis=1)
a = np.ones((3, 4, 5, 6))
np.rollaxis(a, 3, 1).shape
np.rollaxis(a, 2).shape
np.rollaxis(a, 1, 4).shape
x = np.zeros((3, 4, 5))
np.moveaxis(x, 0, -1).shape
np.moveaxis(x, -1, 0).shape
np.transpose(x).shape
np.moveaxis(x, [0, 1], [-1, -2]).shape
np.moveaxis(x, [0, 1, 2], [-1, -2, -3]).shape
x = [1, 2, 3]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2, 0]
y = [4, 5, 6]
np.cross(x, y)
x = [1, 2]
y = [4, 5]
np.cross(x, y)
x = np.array([[1, 2, 3], [4, 5, 6]])
y = np.array([[4, 5, 6], [1, 2, 3]])
np.cross(x, y)
np.cross(x, y, axisc=0)
x = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
y = np.array([[7, 8, 9], [4, 5, 6], [1, 2, 3]])
np.cross(x, y)
np.cross(x, y, axisa=0, axisb=0)
# np.array_repr(np.array([1,2]))
# np.array_repr(np.ma.array([0.]))
# np.array_repr(np.array([], np.int32))
x = np.array([1e-6, 4e-7, 2, 3])
# np.array_repr(x, precision=6, suppress_small=True)
# np.array_str(np.arange(3))
a = np.arange(10)
x = np.arange(4)
np.set_string_function(lambda x: 'random', repr=False)
grid = np.indices((2, 3))
grid.shape
grid[0] # row indices
grid[1] # column indices
x = np.arange(20).reshape(5, 4)
row, col = np.indices((2, 3))
x[row, col]
np.fromfunction(lambda i, j: i == j, (3, 3), dtype=int)
np.fromfunction(lambda i, j: i + j, (3, 3), dtype=int)
np.isscalar(3.1)
np.isscalar([3.1])
np.isscalar(False)
# np.binary_repr(3)
# np.binary_repr(-3)
# np.binary_repr(3, width=4)
# np.binary_repr(-3, width=3)
# np.binary_repr(-3, width=5)
# np.base_repr(5)
# np.base_repr(6, 5)
# np.base_repr(7, base=5, padding=3)
# np.base_repr(10, base=16)
# np.base_repr(32, base=16)
np.identity(3)
np.allclose([1e10, 1e-7], [1.00001e10, 1e-8])
np.allclose([1e10, 1e-8], [1.00001e10, 1e-9])
np.allclose([1e10, 1e-8], [1.0001e10, 1e-9])
# np.allclose([1.0, np.nan], [1.0, np.nan])
# np.allclose([1.0, np.nan], [1.0, np.nan], equal_nan=True)
np.isclose([1e10, 1e-7], [1.00001e10, 1e-8])
np.isclose([1e10, 1e-8], [1.00001e10, 1e-9])
np.isclose([1e10, 1e-8], [1.0001e10, 1e-9])
# np.isclose([1.0, np.nan], [1.0, np.nan])
# np.isclose([1.0, np.nan], [1.0, np.nan], equal_nan=True)
np.array_equal([1, 2], [1, 2])
np.array_equal(np.array([1, 2]), np.array([1, 2]))
np.array_equal([1, 2], [1, 2, 3])
np.array_equal([1, 2], [1, 4])
np.array_equiv([1, 2], [1, 2])
np.array_equiv([1, 2], [1, 3])
np.array_equiv([1, 2], [[1, 2], [1, 2]])
np.array_equiv([1, 2], [[1, 2, 1, 2], [1, 2, 1, 2]])
np.array_equiv([1, 2], [[1, 2], [1, 3]])
def string_to_one_hot(string, maxchar):
"""Converts an ASCII string to a one-of-k encoding."""
ascii = np.array([ord(c) for c in string]).T
return np.array(ascii[:,None] == np.arange(maxchar)[None, :], dtype=int)
def test_fromnumeric():
# Functions
# 'alen', 'all', 'alltrue', 'amax', 'amin', 'any', 'argmax',
# 'argmin', 'argpartition', 'argsort', 'around', 'choose', 'clip',
# 'compress', 'cumprod', 'cumproduct', 'cumsum', 'diagonal', 'mean',
# 'ndim', 'nonzero', 'partition', 'prod', 'product', 'ptp', 'put',
# 'rank', 'ravel', 'repeat', 'reshape', 'resize', 'round_',
# 'searchsorted', 'shape', 'size', 'sometrue', 'sort', 'squeeze',
# 'std', 'sum', 'swapaxes', 'take', 'trace', 'transpose', 'var',
a = [4, 3, 5, 7, 6, 8]
indices = [0, 1, 4]
np.take(a, indices)
a = np.array(a)
# a[indices]
np.take(a, [[0, 1], [2, 3]])
a = np.zeros((10, 2))
b = a.T
a = np.arange(6).reshape((3, 2))
np.reshape(a, (2, 3)) # C-like index ordering
np.reshape(np.ravel(a), (2, 3)) # equivalent to C ravel then C reshape
np.reshape(a, (2, 3), order='F') # Fortran-like index ordering
np.reshape(np.ravel(a, order='F'), (2, 3), order='F')
a = np.array([[1, 2, 3], [4, 5, 6]])
np.reshape(a, 6)
np.reshape(a, 6, order='F')
np.reshape(a, (3, -1)) # the unspecified value is inferred to be 2
choices = [[0, 1, 2, 3], [10, 11, 12, 13], [20, 21, 22, 23],
[30, 31, 32, 33]]
np.choose([2, 3, 1, 0], choices)
np.choose([2, 4, 1, 0], choices, mode='clip') # 4 goes to 3 (4-1)
np.choose([2, 4, 1, 0], choices, mode='wrap') # 4 goes to (4 mod 4)
a = [[1, 0, 1], [0, 1, 0], [1, 0, 1]]
choices = [-10, 10]
np.choose(a, choices)
a = np.array([0, 1]).reshape((2, 1, 1))
c1 = np.array([1, 2, 3]).reshape((1, 3, 1))
c2 = np.array([-1, -2, -3, -4, -5]).reshape((1, 1, 5))
np.choose(a, (c1, c2)) # result is 2x3x5, res[0,:,:]=c1, res[1,:,:]=c2
np.repeat(3, 4)
x = np.array([[1, 2], [3, 4]])
np.repeat(x, 2)
np.repeat(x, 3, axis=1)
np.repeat(x, [1, 2], axis=0)
a = np.arange(5)
np.put(a, [0, 2], [-44, -55])
a = np.arange(5)
np.put(a, 22, -5, mode='clip')
x = np.array([[1, 2, 3]])
np.swapaxes(x, 0, 1)
x = np.array([[[0, 1], [2, 3]], [[4, 5], [6, 7]]])
np.swapaxes(x, 0, 2)
x = np.arange(4).reshape((2, 2))
np.transpose(x)
x = np.ones((1, 2, 3))
np.transpose(x, (1, 0, 2)).shape
a = np.array([3, 4, 2, 1])
np.partition(a, 3)
np.partition(a, (1, 3))
x = np.array([3, 4, 2, 1])
x[np.argpartition(x, 3)]
x[np.argpartition(x, (1, 3))]
x = [3, 4, 2, 1]
np.array(x)[np.argpartition(x, 3)]
a = np.array([[1, 4], [3, 1]])
np.sort(a) # sort along the last axis
np.sort(a, axis=None) # sort the flattened array
np.sort(a, axis=0) # sort along the first axis
dtype = [('name', 'S10'), ('height', float), ('age', int)]
values = [('Arthur', 1.8, 41), ('Lancelot', 1.9, 38), ('Galahad', 1.7, 38)]
a = np.array(values, dtype=dtype) # create a structured array
np.sort(a, order='height') # doctest: +SKIP
np.sort(a, order=['age', 'height']) # doctest: +SKIP
x = np.array([3, 1, 2])
np.argsort(x)
x = np.array([[0, 3], [2, 2]])
np.argsort(x, axis=0)
np.argsort(x, axis=1)
x = np.array([(1, 0), (0, 1)], dtype=[('x', '<i4'), ('y', '<i4')])
np.argsort(x, order=('x', 'y'))
np.argsort(x, order=('y', 'x'))
a = np.arange(6).reshape(2, 3)
np.argmax(a)
np.argmax(a, axis=0)
np.argmax(a, axis=1)
b = np.arange(6)
b[1] = 5
np.argmax(b) # Only the first occurrence is returned.
a = np.arange(6).reshape(2, 3)
np.argmin(a)
np.argmin(a, axis=0)
np.argmin(a, axis=1)
b = np.arange(6)
b[4] = 0
np.argmin(b) # Only the first occurrence is returned.
np.searchsorted([1, 2, 3, 4, 5], 3)
np.searchsorted([1, 2, 3, 4, 5], 3, side='right')
np.searchsorted([1, 2, 3, 4, 5], [-10, 10, 2, 3])
a = np.array([[0, 1], [2, 3]])
np.resize(a, (2, 3))
np.resize(a, (1, 4))
np.resize(a, (2, 4))
x = np.array([[[0], [1], [2]]])
x.shape
np.squeeze(x).shape
np.squeeze(x, axis=(2, )).shape
a = np.arange(4).reshape(2, 2)
a = np.arange(8).reshape(2, 2, 2)
a
a[:, :, 0] # main diagonal is [0 6]
a[:, :, 1] # main diagonal is [1 7]
np.trace(np.eye(3))
a = np.arange(8).reshape((2, 2, 2))
np.trace(a)
a = np.arange(24).reshape((2, 2, 2, 3))
np.trace(a).shape
x = np.array([[1, 2, 3], [4, 5, 6]])
np.ravel(x)
x.reshape(-1)
np.ravel(x, order='F')
np.ravel(x.T)
np.ravel(x.T, order='A')
a = np.arange(3)[::-1]
a
# a = np.arange(12).reshape(2,3,2).swapaxes(1,2); a
x = np.eye(3)
np.nonzero(x)
x[np.nonzero(x)]
np.transpose(np.nonzero(x))
a = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
a > 3
np.nonzero(a > 3)
np.shape(np.eye(3))
np.shape([[1, 2]])
np.shape([0])
np.shape(0)
a = np.array([(1, 2), (3, 4)], dtype=[('x', 'i4'), ('y', 'i4')])
np.shape(a)
a.shape
a = np.array([[1, 2], [3, 4], [5, 6]])
np.compress([0, 1], a, axis=0)
np.compress([False, True, True], a, axis=0)
np.compress([False, True], a, axis=1)
np.compress([False, True], a)
a = np.arange(10)
np.clip(a, 1, 8)
np.clip(a, 3, 6, out=a)
a = np.arange(10)
np.clip(a, [3, 4, 1, 1, 1, 4, 4, 4, 4, 4], 8)
np.sum([])
np.sum([0.5, 1.5])
np.sum([0.5, 0.7, 0.2, 1.5], dtype=np.int32)
np.sum([[0, 1], [0, 5]])
np.sum([[0, 1], [0, 5]], axis=0)
np.sum([[0, 1], [0, 5]], axis=1)
# np.ones(128, dtype=np.int8).sum(dtype=np.int8)
# np.any([[True, False], [True, True]])
# np.any([[True, False], [False, False]], axis=0)
# np.any([-1, 0, 5])
# np.any(np.nan)
# np.all([[True,False],[True,True]])
# np.all([[True,False],[True,True]], axis=0)
# np.all([-1, 4, 5])
# np.all([1.0, np.nan])
a = np.array([[1, 2, 3], [4, 5, 6]])
np.cumsum(a)
np.cumsum(a, dtype=float) # specifies type of output value(s)
np.cumsum(a, axis=0) # sum over rows for each of the 3 columns
np.cumsum(a, axis=1) # sum over columns for each of the 2 rows
x = np.arange(4).reshape((2, 2))
np.ptp(x, axis=0)
np.ptp(x, axis=1)
a = np.arange(4).reshape((2, 2))
np.amax(a) # Maximum of the flattened array
np.amax(a, axis=0) # Maxima along the first axis
np.amax(a, axis=1) # Maxima along the second axis
b = np.arange(5, dtype=np.float)
# b[2] = np.NaN
np.amax(b)
np.nanmax(b)
a = np.arange(4).reshape((2, 2))
np.amin(a) # Minimum of the flattened array
np.amin(a, axis=0) # Minima along the first axis
np.amin(a, axis=1) # Minima along the second axis
b = np.arange(5, dtype=np.float)
# b[2] = np.NaN
np.amin(b)
np.nanmin(b)
a = np.zeros((7, 4, 5))
a.shape[0]
np.alen(a)
x = np.array([536870910, 536870910, 536870910, 536870910])
np.prod(x) #random
np.prod([])
np.prod([1., 2.])
np.prod([[1., 2.], [3., 4.]])
np.prod([[1., 2.], [3., 4.]], axis=1)
x = np.array([1, 2, 3], dtype=np.uint8)
# np.prod(x).dtype == np.uint
x = np.array([1, 2, 3], dtype=np.int8)
# np.prod(x).dtype == np.int
a = np.array([1, 2, 3])
np.cumprod(a) # intermediate results 1, 1*2
a = np.array([[1, 2, 3], [4, 5, 6]])
np.cumprod(a, dtype=float) # specify type of output
np.cumprod(a, axis=0)
np.cumprod(a, axis=1)
np.ndim([[1, 2, 3], [4, 5, 6]])
np.ndim(np.array([[1, 2, 3], [4, 5, 6]]))
np.ndim(1)
a = np.array([[1, 2, 3], [4, 5, 6]])
np.size(a)
np.size(a, 1)
np.size(a, 0)
np.around([0.37, 1.64])
np.around([0.37, 1.64], decimals=1)
np.around([.5, 1.5, 2.5, 3.5, 4.5]) # rounds to nearest even value
np.around([1, 2, 3, 11], decimals=1) # ndarray of ints is returned
np.around([1, 2, 3, 11], decimals=-1)
a = np.array([[1, 2], [3, 4]])
np.mean(a)
np.mean(a, axis=0)
np.mean(a, axis=1)
a = np.zeros((2, 512 * 512), dtype=np.float32)
a[0, :] = 1.0
a[1, :] = 0.1
np.mean(a)
np.mean(a, dtype=np.float64)
a = np.array([[1, 2], [3, 4]])
np.std(a)
np.std(a, axis=0)
np.std(a, axis=1)
a = np.zeros((2, 512 * 512), dtype=np.float32)
a[0, :] = 1.0
a[1, :] = 0.1
np.std(a)
np.std(a, dtype=np.float64)
a = np.array([[1, 2], [3, 4]])
np.var(a)
np.var(a, axis=0)
np.var(a, axis=1)
a = np.zeros((2, 512 * 512), dtype=np.float32)
a[0, :] = 1.0
a[1, :] = 0.1
np.var(a)
np.var(a, dtype=np.float64)
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))
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
def test_context():
set_context(gpu(1)) # set the global context as gpu(1)
def sigmoid(x):
return 0.5 * (np.tanh(x / 2) + 1)
def predict(weights, inputs):
return sigmoid(np.dot(inputs, weights))
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 training_accuracy(weights, inputs):
preds = predict(weights, inputs)
error = np.count_nonzero(
np.argmax(preds, axis=1) - np.argmax(targets, axis=1))
return (256 - error) * 100 / 256.0
with gpu(0):
xshape = (256, 500)
wshape = (500, 250)
tshape = (256, 250)
inputs = random.rand(*xshape) - 0.5
targets = np.zeros(tshape)
truth = random.randint(0, 250, 256)
targets[np.arange(256), truth] = 1
weights = random.rand(*wshape) - 0.5
training_gradient_fun = grad(training_loss)
for i in range(20):
print('Trained loss accuracy #{}: {}%'.format(
i, training_accuracy(weights, inputs)))
gr = training_gradient_fun(weights, inputs)
weights -= gr * 0.01
print("\nff and bp on {0}".format(weights.context))
print("\nexecute on cpu")
with cpu():
x_cpu = random.rand(32, 64) - 0.5
y_cpu = random.rand(64, 32) - 0.5
z_cpu = np.dot(x_cpu, y_cpu)
print('z_cpu.context = {0}'.format(z_cpu.context))
print("\nexecute on gpu(0)")
with gpu(0):
x_gpu0 = random.rand(32, 64) - 0.5
y_gpu0 = random.rand(64, 32) - 0.5
z_gpu0 = np.dot(x_gpu0, y_gpu0)
z_gpu0.asnumpy()
print('z_gpu0.context = {0}'.format(z_gpu0.context))
print("\n[use global context] execute on gpu(1)")
x_gpu1 = random.rand(32, 64) - 0.5
y_gpu1 = random.rand(64, 32) - 0.5
z_gpu1 = np.dot(x_gpu1, y_gpu1)
z_gpu1.asnumpy()
print('z_gpu1.context = {0}'.format(z_gpu1.context))