def main():
# Generate dataset and initial weight
x_t, y_t = generate_dataset(1, 1, -5, point_num=100)
# add extra dim to build homogenous coordinates
x_t = np.concatenate((x_t, np.ones((x_t.shape[0], 1))), axis=1)
W_val = np.random.rand(3, 1)
# draw initial decision superplane
draw(W_val, x_t, y_t)
# Create the model
x = ad.Variable(name='x')
W = ad.Variable(name='W')
y = ad.sigmoid_op(ad.matmul_op(x, W))
# Define loss
y_ = ad.Variable(name='y_')
cross_entropy = ad.reduce_mean_op(-ad.reduce_sum_op(
y_ * ad.log_op(y) +
(1 - y_) * ad.log_op(1 - y), reduction_indices=[1]))
# Update rule
learning_rate = 0.5
W_grad, = ad.gradients(cross_entropy, [W])
W_train_step = W - learning_rate * W_grad
# Training
executor = ad.Executor([cross_entropy, y, W_train_step])
steps = 200
plt.ion()
for i in range(steps):
plt.cla()
loss_val, y_val, W_val = executor.run(feed_dict={
x: x_t,
y_: y_t,
W: W_val,
})
print("Step {}: loss: {}".format(i + 1, loss_val))
# draw trained decision superplane
draw(W_val, x_t, y_t)
plt.pause(0.1)
plt.ioff()
plt.show()
def test_exp():
x1 = ad.Variable(name = "x1")
y = 1 + 2 * ad.exp_op(ad.log_op(x1))
x1_val = np.ones((2, 1))
grad_y, = ad.gradients(y, [x1])
executor = ad.Executor([y, grad_y])
y_val, grad_y_val, = executor.run(feed_dict = {x1: x1_val})
assert np.array_equal(y_val, 3 * np.ones_like(y_val))
assert np.array_equal(grad_y_val, 2 * np.ones_like(grad_y_val))
def softmax_ce_loss(preds, truth):
"""
calculate the softmax and xent loss in a more efficient way
:param preds: the pred is the output of the model
:param truth: the true label, a one-hot vector
:return: the loss
"""
pred_max = ad.max_op(preds)
preds_shift = ad.add_byscalar_op(ad.neg_op(pred_max), x)
exps = ad.exp_op(preds_shift)
return minus_op(ad.log_op(ad.sum_op(exps)), ad.sum_op(ad.mul_op(preds_shift, truth)))
def test_log():
x1 = ad.Variable(name = "x1")
x2 = ad.Variable(name = "x2")
y = ad.log_op(x1) / x2
grad_y, = ad.gradients(y, [x1])
x1_val = 2 * np.ones((2, 1))
x2_val = np.ones((1,1))
executor = ad.Executor([y, grad_y])
y_val, grad_y_val, = executor.run(feed_dict = {x1: x1_val, x2:x2_val})
assert np.array_equal(y_val, np.log(x1_val))
assert np.array_equal(grad_y_val, 0.5 * np.ones_like(grad_y_val))
def test_log(): # P
x2 = ad.Variable(name="x2")
y = ad.log_op(x2)
grad_x2, = ad.gradients(y, [x2])
executor = ad.Executor([y, grad_x2])
x2_val = np.linspace(0.1, 2, 5)
y_val, grad_x2_val = executor.run(feed_dict={x2: x2_val})
assert isinstance(y, ad.Node)
assert np.array_equal(y_val, np.log(x2_val))
assert np.array_equal(grad_x2_val, 1 / x2_val)
def fit(self, X, Y):
x = ad.Variable(name='x')
w = ad.Variable(name='w')
y = ad.Variable(name='y')
p = 1 / (1 + ad.exp_op(0 - ad.matmul_op(w, x)))
# cross entropy
loss = 0 - y * ad.log_op(p) - (1 - y) * ad.log_op(1 - p)
grad_w, = ad.gradients(loss, [w])
# SGD
length = np.shape(X)[0]
self.num_feature = np.shape(X)[1]
executor = ad.Executor([loss, grad_w])
self.coef_ = np.random.rand(1, self.num_feature) / 1000.0
for i in range(self.maxiter):
grad = np.zeros((1, self.num_feature))
loss = 0
for j in range(self.batch):
t = random.choice(range(length))
x_val = X[t].reshape((self.num_feature, 1))
if Y[t] == self.labels[0]:
y_val = 0
else:
y_val = 1
loss_val, grad_w_val = executor.run(feed_dict={
x: x_val,
w: self.coef_,
y: y_val
})
grad = grad + grad_w_val
loss = loss + loss_val
self.coef_ = self.coef_ - self.learning_rate * grad / self.batch
if i % 100 == 0:
print(loss)
def get_model_params(x_train, y_train, class_1, class_2):
'''returns the weights after performing gradient descdent'''
learning_rate = 0.01
batch_size = 8
x = ad.Variable(name='x')
w = ad.Variable(name='w')
y = ad.Variable(name='y')
logistic_regression = 1 / (1 + ad.exp_op(0 - ad.matmul_op(w, x)))
cross_entropy = -1 * y * ad.log_op(logistic_regression) - (1 - y) * ad.log_op(1 - logistic_regression)
gradients = ad.gradients(cross_entropy, [w])[0]
executor = ad.Executor([cross_entropy, gradients])
weights = np.random.rand(1, np.shape(x_train)[1]) / 1000.0
#batch = 0
#previous_loss = 0
for i in range(5000):
grad = np.zeros((1, np.shape(x_train)[1]))
loss = 0
#go ramdomly over examples in each batch
for _ in range(batch_size):
t = random.choice(range(np.shape(x_train)[0]))
x_flat = x_train[t].reshape((np.shape(x_train)[1], 1))
y_label = 0 if y_train[t] == class_1 else 1
loss_delta, grad_delta = executor.run(feed_dict={x : x_flat, w : weights, y : y_label})
grad += grad_delta
loss += loss_delta
weights = weights - (learning_rate * grad / batch_size)
if i % 1000 == 0:
print("loss = {:.3f} loss_delta = {:.3f}".format(loss[0][0], loss_delta[0][0]))
return weights
def test_log():
x2 = ad.Variable(name="x2")
y = ad.log_op(x2)
grad_x2, = ad.gradients(y, [x2])
executor = ad.Executor([y, grad_x2])
x2_val = 2 * np.ones(3)
y_val, grad_x2_val = executor.run(feed_dict={x2: x2_val})
epsilon = 1e-6
zero_arr = np.zeros(3) + epsilon
assert isinstance(y, ad.Node)
assert np.all(np.less_equal(np.abs(y_val - np.log(x2_val)), zero_arr))
assert np.all(np.less_equal(np.abs(1 / x2_val - grad_x2_val), zero_arr))
def test():
x1 = ad.Variable(name="x1")
x2 = ad.Variable(name="x2")
x3 = ad.Variable(name="x3")
y = (ad.sin_op(x1 + 1) + ad.cos_op(2 * x2)) * ad.tan_op(ad.log_op(x3)) + (
ad.sin_op(x2 + 1)) + ad.cos_op(2 * x1) * ad.exp_op(1 + ad.sin_op(x3))
grad_x1, grad_x2, grad_x3 = ad.gradients(y, [x1, x2, x3])
executor = ad.Executor([y, grad_x1, grad_x2, grad_x3])
x1_val = 1 * np.ones(1)
x2_val = 2 * np.ones(1)
x3_val = 3 * np.ones(1)
y_val, grad_x1_val, grad_x2_val, grad_x3_val = executor.run(feed_dict={
x1: x1_val,
x2: x2_val,
x3: x3_val
})
print('x1=', x1_val[0])
print('x2=', x2_val[0])
print('x3=', x3_val[0])
print('---------------------------------------------------------------')
print('y0_val=', y_val[0])
print('grad_x1_val= ', grad_x1_val[0])
print('grad_x2_val= ', grad_x2_val[0])
print('grad_x3_val= ', grad_x3_val[0])
print('---------------------------------------------------------------')
y_numerical, grad_numerical = numerical_diff(f, [x1_val, x2_val, x3_val],
1e-10)
print('y0_numerical= ', y_numerical)
grad_numerical_x1, grad_numerical_x2, grad_numerical_x3 = grad_numerical[
0], grad_numerical[1], grad_numerical[2]
print('grad_numerical_x1 =', grad_numerical_x1)
print('grad_numerical_x2 =', grad_numerical_x2)
print('grad_numerical_x3 =', grad_numerical_x3)
print('---------------------------------------------------------------')
print('gradients Offset:')
print('x1:', abs(grad_x1_val - grad_numerical_x1))
assert abs(grad_x1_val - grad_numerical_x1) < 1e-5
print('x2:', abs(grad_x2_val - grad_numerical_x2))
assert abs(grad_x2_val - grad_numerical_x2) < 1e-5
print('x3:', abs(grad_x3_val - grad_numerical_x3))
assert abs(grad_x3_val - grad_numerical_x3) < 1e-5
def cross_entropy(output, labels):
loss = -1.0 * ad.reduce_sum_op(labels * ad.log_op(output) +
(1.0 - labels) * ad.log_op(1.0 - output),
axis=1)
return loss
import numpy as np import autodiff as ad x = ad.Variable(name = "x") w = ad.Variable(name = "w") b = ad.Variable(name = "b") labels = ad.Variable(name = "lables") # Define Computation graph p = 1.0 / (1.0 + ad.exp_op((-1.0 * ad.matmul_op(w, x)))) loss = -1.0 * ad.reduce_sum_op(labels * ad.log_op(p) + (1.0 - labels) * ad.log_op(1.0 - p), axis = 1) grad_y_w, = ad.gradients(loss, [w]) num_features = 2 num_points = 500 num_iterations = 1000 learning_rate = 0.001 # The dummy dataset consists of two classes. # The classes are modelled as a random normal variables with different means. class_1 = np.random.normal(2, 0.1, (num_points / 2, num_features)) class_2 = np.random.normal(4, 0.1, (num_points / 2, num_features))
import numpy as np
import autodiff as ad
x = ad.Variable(name="x")
w = ad.Variable(name="w")
b = ad.Variable(name="b")
labels = ad.Variable(name="lables")
# Define Computation graph
p = 1.0 / (1.0 + ad.exp_op((-1.0 * ad.matmul_op(w, x))))
loss = -1.0 * ad.reduce_sum_op(
labels * ad.log_op(p) + (1.0 - labels) * ad.log_op(1.0 - p), axis=1)
grad_y_w, = ad.gradients(loss, [w])
num_features = 2
num_points = 200
num_iterations = 1000
learning_rate = 0.01
# The dummy dataset consists of two classes.
# The classes are modelled as a random normal variables with different means.
class_1 = np.random.normal(2, 0.1, (num_points / 2, num_features))
class_2 = np.random.normal(4, 0.1, (num_points / 2, num_features))
x_val = np.concatenate((class_1, class_2), axis=0).T
x_val = np.concatenate((x_val, np.ones((1, num_points))), axis=0)
w_val = np.random.normal(size=(1, num_features + 1))
import autodiff as ad
import numpy as np
w = ad.Variable(name="w")
x = ad.Variable(name="x")
b = ad.Variable(name="b")
labels = ad.Variable(name="lables")
out = 1.0 / (1.0 + ad.exp_op((-1.0 * (w * x + b))))
ce_loss = -1.0 * ((labels * ad.log_op(out)) +
((1.0 - labels) * ad.log_op(1.0 - out)))
grad_w, grad_b = ad.gradients(ce_loss, [w, b])
# weights our model initially starts at
w_val = 10
b_val = 1
# weights our model should reach to
w_required = 5
b_required = 20
# we are simulating the training dataset for logistic regression
# taking x as a continuous array from -10 to 10 with step size of 0.01
x_val = np.arange(-10, 6, 0.01)
# finding the labels by doing the exact calculation of logistic regression using numpy
labels_val = 1 / (1 + np.exp(-(w_required * x_val + b_required)))
import autodiff as ad
import numpy as np
# construct the computation graph
x = ad.Variable(name="x")
w = ad.Variable(name="w")
y_ = ad.Variable(name="lables")
prob = 1.0 / (1.0 + ad.exp_op((-1.0 * ad.matmul_op(w, x))))
loss = -1.0 * ad.reduce_sum_op(
y_ * ad.log_op(prob) + (1.0 - y_) * ad.log_op(1.0 - prob), axis=1)
# pay attention that there is a ','
grad_w, = ad.gradients(loss, [w])
# Data
data1 = np.random.normal(1, 0.1, size=(100, 10))
data2 = np.random.normal(5, 0.4, size=(200, 10))
data = np.concatenate((data1, data2), axis=0).T
x_val = np.concatenate((data, np.ones((1, 300))), axis=0)
y_val = np.concatenate((np.zeros(
(data1.shape[0], 1)), np.ones((data2.shape[0], 1))),
axis=0).T
# Variables
w_val = np.random.normal(size=(1, 11))
# Params
learning_rate = 0.0001
# Execute
executor = ad.Executor([loss, grad_w])