def monitor(self):
if not self._monitor:
return
val_mean_abs = np.array(ca.mean(ca.fabs(self._array)))
grad_mean_abs = np.array(ca.mean(ca.fabs(self._tmp_grad_array)))
step_mean_abs = np.array(ca.mean(ca.fabs(self._tmp_last_step)))
logger.info("%s:\t%.1e [%.1e, %.1e]" % (self.name, val_mean_abs, grad_mean_abs, step_mean_abs))
def monitor(self):
if not self._monitor:
return
val_mean_abs = np.array(ca.mean(ca.fabs(self._array)))
grad_mean_abs = np.array(ca.mean(ca.fabs(self._tmp_grad_array)))
step_mean_abs = np.array(ca.mean(ca.fabs(self._tmp_step)))
log.info('%s:\t%.1e [%.1e, %.1e]', self.name, val_mean_abs,
grad_mean_abs, step_mean_abs)
def monitor(self):
for param, step in zip(self.params, self.steps):
if param.monitor:
val_mean_abs = np.array(ca.mean(ca.fabs(param.values)))
grad_mean_abs = np.array(ca.mean(ca.fabs(param.grad())))
step_mean_abs = np.array(ca.mean(ca.fabs(step)))
logger.info('%s:\t%.1e [%.1e, %.1e]'
% (param.name, val_mean_abs, grad_mean_abs,
step_mean_abs))
def monitor(self):
for param, step in zip(self.params, self.steps):
if param.monitor:
val_mean_abs = np.array(ca.mean(ca.fabs(param.values)))
grad_mean_abs = np.array(ca.mean(ca.fabs(param.grad())))
step_mean_abs = np.array(ca.mean(ca.fabs(step)))
logger.info(
'%s:\t%.1e [%.1e, %.1e]' %
(param.name, val_mean_abs, grad_mean_abs, step_mean_abs))
def _update(self):
# Forward propagation
next_x = self.x.array
x_feats = [None]*len(self.layers)
x_grams = [None]*len(self.layers)
for l, layer in enumerate(self.layers):
next_x = layer.fprop(next_x)
if self.subject_weights[l] > 0:
x_feats[l] = next_x
if self.style_weights[l] > 0:
x_feats[l] = next_x
x_grams[l] = gram_matrix(next_x)
# Backward propagation
grad = ca.zeros_like(next_x)
loss = ca.zeros(1)
for l, layer in reversed(list(enumerate(self.layers))):
if self.subject_weights[l] > 0:
diff = x_feats[l] - self.subject_feats[l]
norm = ca.sum(ca.fabs(diff)) + 1e-8
weight = float(self.subject_weights[l]) / norm
grad += diff * weight
loss += 0.5*weight*ca.sum(diff**2)
if self.style_weights[l] > 0:
diff = x_grams[l] - self.style_grams[l]
n_channels = diff.shape[0]
x_feat = ca.reshape(x_feats[l], (n_channels, -1))
style_grad = ca.reshape(ca.dot(diff, x_feat), x_feats[l].shape)
norm = ca.sum(ca.fabs(style_grad))
weight = float(self.style_weights[l]) / norm
style_grad *= weight
grad += style_grad
loss += 0.25*weight*ca.sum(diff**2)
grad = layer.bprop(grad)
if self.tv_weight > 0:
x = ca.reshape(self.x.array, (3, 1) + grad.shape[2:])
tv = self.tv_conv.fprop(x, self.tv_kernel)
tv *= self.tv_weight
grad -= ca.reshape(tv, grad.shape)
ca.copyto(self.x.grad_array, grad)
return loss
def _update(self):
# Forward propagation
next_x = self.x.array
x_feats = [None] * len(self.layers)
for l, layer in enumerate(self.layers):
next_x = layer.fprop(next_x)
if self.subject_weights[l] > 0 or self.style_weights[l] > 0:
x_feats[l] = next_x
# Backward propagation
grad = ca.zeros_like(next_x)
loss = ca.zeros(1)
for l, layer in reversed(list(enumerate(self.layers))):
if self.subject_weights[l] > 0:
diff = x_feats[l] - self.subject_feats[l]
norm = ca.sum(ca.fabs(diff)) + 1e-8
weight = float(self.subject_weights[l]) / norm
grad += diff * weight
loss += 0.5 * weight * ca.sum(diff**2)
if self.style_weights[l] > 0:
diff = gram_matrix(x_feats[l]) - self.style_grams[l]
n_channels = diff.shape[0]
x_feat = ca.reshape(x_feats[l], (n_channels, -1))
style_grad = ca.reshape(ca.dot(diff, x_feat), x_feats[l].shape)
norm = ca.sum(ca.fabs(style_grad))
weight = float(self.style_weights[l]) / norm
style_grad *= weight
grad += style_grad
loss += 0.25 * weight * ca.sum(diff**2)
grad = layer.bprop(grad)
if self.tv_weight > 0:
x = ca.reshape(self.x.array, (3, 1) + grad.shape[2:])
tv = self.tv_conv.fprop(x, self.tv_kernel)
tv *= self.tv_weight
grad -= ca.reshape(tv, grad.shape)
ca.copyto(self.x.grad_array, grad)
return loss
def fprop(self):
ca.fabs(self.x.out, out=self.out)
def fprop(self):
ca.fabs(self.x.array, out=self.array)
def fprop(self):
ca.fabs(self.x.array, out=self.array)
def fprop(self):
ca.fabs(self.x.out, out=self.out)
