def fprop(self):
if self.phase == 'train':
# Calculate batch mean
tmp = ca.mean(self.x.out, axis=0, keepdims=True)
# Center input
ca.subtract(self.x.out, tmp, self._tmp_batch_centered)
# Update running mean
tmp *= 1 - self.momentum
self.running_mean *= self.momentum
self.running_mean += tmp
# Calculate batch variance
ca.power(self._tmp_batch_centered, 2, self.out)
ca.mean(self.out, axis=0, keepdims=True,
out=self._tmp_batch_inv_std)
# Calculate 1 / E([x - E(x)]^2)
self._tmp_batch_inv_std += self.eps
ca.sqrt(self._tmp_batch_inv_std, self._tmp_batch_inv_std)
ca.power(self._tmp_batch_inv_std, -1, self._tmp_batch_inv_std)
# Normalize input
ca.multiply(self._tmp_batch_centered, self._tmp_batch_inv_std,
self.out)
# Update running std
self.running_std *= self.momentum
ca.multiply(self._tmp_batch_inv_std, 1-self.momentum, tmp)
self.running_std += tmp
elif self.phase == 'test':
ca.subtract(self.x.out, self.running_mean, self.out)
self.out *= self.running_std
else:
raise ValueError('Invalid phase: %s' % self.phase)
if self.affine:
self.out *= self.gamma.array
self.out += self.beta.array
def fprop(self):
if self.phase == 'train':
# Calculate batch mean
tmp = ca.mean(self.x.out, axis=0, keepdims=True)
# Center input
ca.subtract(self.x.out, tmp, self._tmp_batch_centered)
# Update running mean
tmp *= 1 - self.momentum
self.running_mean *= self.momentum
self.running_mean += tmp
# Calculate batch variance
ca.power(self._tmp_batch_centered, 2, self.out)
ca.mean(self.out,
axis=0,
keepdims=True,
out=self._tmp_batch_inv_std)
# Calculate 1 / E([x - E(x)]^2)
self._tmp_batch_inv_std += self.eps
ca.sqrt(self._tmp_batch_inv_std, self._tmp_batch_inv_std)
ca.power(self._tmp_batch_inv_std, -1, self._tmp_batch_inv_std)
# Normalize input
ca.multiply(self._tmp_batch_centered, self._tmp_batch_inv_std,
self.out)
# Update running std
self.running_std *= self.momentum
ca.multiply(self._tmp_batch_inv_std, 1 - self.momentum, tmp)
self.running_std += tmp
elif self.phase == 'test':
ca.subtract(self.x.out, self.running_mean, self.out)
self.out *= self.running_std
else:
raise ValueError('Invalid phase: %s' % self.phase)
if self.affine:
self.out *= self.gamma.array
self.out += self.beta.array
def fprop(self):
# e_i = exp(x_i - max(x))
# y = e_i / sum(e)
tmp1 = ca.amax(self.x.array, axis=1, keepdims=True)
ca.subtract(self.x.array, tmp1, self.array)
ca.exp(self.array, self.array)
ca.sum(self.array, axis=1, keepdims=True, out=tmp1)
self.array /= tmp1
def bprop(self):
# -(target/pred - (1 - target)/(1 - pred))
tmp1 = 1 - self.target.out
tmp2 = 1 - self.pred.out
tmp2 += self.eps
ca.divide(tmp1, tmp2, tmp1)
ca.add(self.pred.out, self.eps, tmp2)
ca.divide(self.target.out, tmp2, out=tmp2)
ca.subtract(tmp1, tmp2, self.pred.out_grad)
self.pred.out_grad *= self.out_grad
def bprop(self):
# -(target/pred - (1 - target)/(1 - pred))
tmp1 = 1 - self.target.array
tmp2 = 1 - self.pred.array
tmp2 += self.eps
ca.divide(tmp1, tmp2, tmp1)
ca.add(self.pred.array, self.eps, tmp2)
ca.divide(self.target.array, tmp2, out=tmp2)
ca.subtract(tmp1, tmp2, self.pred.grad_array)
self.pred.grad_array *= ca.reshape(self.grad_array, self.bcast_shape)
def bprop(self):
# -(target/pred - (1 - target)/(1 - pred))
tmp1 = 1 - self.target.array
tmp2 = 1 - self.pred.array
tmp2 += self.eps
ca.divide(tmp1, tmp2, tmp1)
ca.add(self.pred.array, self.eps, tmp2)
ca.divide(self.target.array, tmp2, out=tmp2)
ca.subtract(tmp1, tmp2, self.pred.grad_array)
self.pred.grad_array *= ca.reshape(self.grad_array, self.bcast_shape)
def fprop(self):
if self.phase == 'train':
# Calculate batch mean
tmp = ca.mean(ca.mean(self.x.array, axis=0, keepdims=True),
axis=(2, 3), keepdims=True)
# Center input
ca.subtract(self.x.array, tmp, self._tmp_batch_centered)
# Update running mean
tmp *= 1 - self.momentum
self.running_mean *= self.momentum
self.running_mean += tmp
# Calculate batch variance
ca.power(self._tmp_batch_centered, 2, self.array)
ca.mean(ca.mean(self.array, axis=0, keepdims=True), axis=(2, 3),
keepdims=True, out=self._tmp_batch_inv_std)
# Calculate 1 / E([x - E(x)]^2)
self._tmp_batch_inv_std += self.eps
ca.sqrt(self._tmp_batch_inv_std, self._tmp_batch_inv_std)
ca.power(self._tmp_batch_inv_std, -1, self._tmp_batch_inv_std)
# Normalize input
ca.multiply(self._tmp_batch_centered, self._tmp_batch_inv_std,
self.array)
# Update running std
self.running_std *= self.momentum
ca.multiply(self._tmp_batch_inv_std, 1-self.momentum, tmp)
self.running_std += tmp
if self.noise_std > 0.0:
noise = ca.random.normal(scale=self.noise_std,
size=self.shape)
ca.add(self.array, noise, self.array)
elif self.phase == 'test':
ca.subtract(self.x.array, self.running_mean, self.array)
self.array *= self.running_std
else:
raise ValueError('Invalid phase: %s' % self.phase)
if self.affine:
self.array *= self.gamma.array
self.array += self.beta.array
def matrix_factorization(R, P, Q, mask, steps=200000000, alpha=0.00005, beta=0.02):
Q = ca.transpose(Q)
for step in xrange(steps):
E = ca.subtract(R, ca.multiply(ca.dot(P,Q), mask))
rmse = ca.sqrt(ca.sum(ca.power(E,2)) / ca.sum(mask))
rmse = np.array(rmse)[0]
print 'step: %i RMSE: %f' % (step, rmse)
if rmse < 0.65:
break
P = ca.add(ca.multiply(P,(1-alpha*beta)),ca.multiply(ca.dot(E,ca.transpose(Q)), 2*alpha))
Q = ca.add(ca.multiply(Q,(1-alpha*beta)),ca.multiply(ca.dot(ca.transpose(P),E),2*alpha))
return P, Q
def test_binary():
a_np = np.random.normal(size=(5, 5))
b_np = np.random.normal(size=(5, 5))
a_ca = ca.array(a_np)
b_ca = ca.array(b_np)
c_np = np.add(a_np, b_np)
c_ca = ca.add(a_ca, b_ca)
print(np.allclose(c_np, np.array(c_ca)))
np.add(a_np, b_np, a_np)
ca.add(a_ca, b_ca, a_ca)
print(np.allclose(a_np, np.array(a_ca)))
np.multiply(a_np, b_np, a_np)
ca.multiply(a_ca, b_ca, a_ca)
print(np.allclose(a_np, np.array(a_ca)))
a_np = np.random.normal(size=(5, 5))
b_np = np.random.normal(size=(5, 5)) > 0
a_ca = ca.array(a_np)
b_ca = ca.array(b_np)
c_np = np.multiply(a_np, b_np)
c_ca = ca.multiply(a_ca, b_ca)
print(np.allclose(c_np, np.array(c_ca)))
a_np = np.random.normal()
b_np = np.random.normal(size=(5, 5))
a_ca = ca.array(a_np)
b_ca = ca.array(b_np)
c_np = np.multiply(a_np, b_np)
c_ca = ca.multiply(a_ca, b_ca)
print(np.allclose(c_np, np.array(c_ca)))
a_ca = ca.array(a_np)
b_ca = ca.array(b_np)
c_np = np.divide(a_np, b_np)
c_ca = ca.divide(a_ca, b_ca)
print(np.allclose(c_np, np.array(c_ca)))
a_ca = ca.array(a_np)
b_ca = ca.array(b_np)
c_np = np.subtract(a_np, b_np)
c_ca = ca.subtract(a_ca, b_ca)
print(np.allclose(c_np, np.array(c_ca)))
def fprop(self, x):
n_channels = x.shape[1]
# Calculate local mean
tmp = self.conv_op.fprop(x, self.ca_kernel)
if n_channels > 1:
ca.divide(tmp, n_channels, tmp)
# Center input with local mean
centered = ca.subtract(x, tmp)
# Calculate local standard deviation
tmp = ca.power(centered, 2)
tmp = self.conv_op.fprop(tmp, self.ca_kernel)
if n_channels > 1:
ca.divide(tmp, n_channels, tmp)
ca.sqrt(tmp, tmp)
# Scale centered input with standard deviation
return centered / (tmp + self.eps)
def fprop(self, x, phase):
n_channels = x.shape[1]
# Calculate local mean
tmp = self.conv_op.fprop(x, self.ca_kernel)
if n_channels > 1:
ca.divide(tmp, n_channels, tmp)
# Center input with local mean
centered = ca.subtract(x, tmp)
# Calculate local standard deviation
tmp = ca.power(centered, 2)
tmp = self.conv_op.fprop(tmp, self.ca_kernel)
if n_channels > 1:
ca.divide(tmp, n_channels, tmp)
ca.sqrt(tmp, tmp)
# Scale centered input with standard deviation
return centered / (tmp + self.eps)
def matrix_factorization(R,
P,
Q,
mask,
steps=200000000,
alpha=0.00005,
beta=0.02):
Q = ca.transpose(Q)
for step in xrange(steps):
E = ca.subtract(R, ca.multiply(ca.dot(P, Q), mask))
rmse = ca.sqrt(ca.sum(ca.power(E, 2)) / ca.sum(mask))
rmse = np.array(rmse)[0]
print 'step: %i RMSE: %f' % (step, rmse)
if rmse < 0.65:
break
P = ca.add(ca.multiply(P, (1 - alpha * beta)),
ca.multiply(ca.dot(E, ca.transpose(Q)), 2 * alpha))
Q = ca.add(ca.multiply(Q, (1 - alpha * beta)),
ca.multiply(ca.dot(ca.transpose(P), E), 2 * alpha))
return P, Q
def bprop(self):
ca.subtract(self.pred.array, self.target.array, self.pred.grad_array)
self.pred.grad_array *= 2
self.pred.grad_array *= ca.reshape(self.grad_array, self.bcast_shape)
def bprop(self):
ca.subtract(self.pred.array, self.target.array, self.pred.grad_array)
self.pred.grad_array *= ca.reshape(self.grad_array, self.bcast_shape)
def fprop(self):
ca.subtract(self.lhs.out, self.rhs.out, out=self.out)
def bprop(self):
ca.subtract(self.pred.array, self.target.array, self.pred.grad_array)
if self.sigma != 1.0:
self.pred.grad_array *= 2 * self.multiplier
self.pred.grad_array *= ca.reshape(self.grad_array, self.bcast_shape)
def fprop(self):
ca.subtract(self.lhs.out, self.rhs.out, out=self.out)
def fprop(self):
ca.subtract(self.lhs.array, self.rhs.array, out=self.array)
def fprop(self):
ca.subtract(self.lhs.array, self.rhs.array, out=self.array)
def bprop(self):
ca.subtract(self.pred.out, self.target.out, self.pred.out_grad)
self.pred.out_grad *= 2*self.multiplier
self.pred.out_grad *= self.out_grad
def bprop(self):
ca.subtract(self.pred.array, self.target.array, self.pred.grad_array)
if self.sigma != 1.0:
self.pred.grad_array *= 2*self.multiplier
self.pred.grad_array *= ca.reshape(self.grad_array, self.bcast_shape)
def bprop(self):
# y_i * (y_grad_i - sum(y_grad * y))
ca.multiply(self.array, self.grad_array, self.x.grad_array)
tmp1 = ca.sum(self.x.grad_array, axis=1, keepdims=True)
ca.subtract(self.grad_array, tmp1, self.x.grad_array)
self.x.grad_array *= self.array
def bprop(self):
ca.subtract(self.pred.out, self.target.out, self.pred.out_grad)
self.pred.out_grad *= 2
self.pred.out_grad *= self.out_grad