def create(self):
# current gradient
gradient_computer = gc.ScopeGradientComputer(self.opt, self.loss,
self.model.weights)
grads_and_vars = gradient_computer.compute()
g_cur, v_cur = zip(*grads_and_vars)
g_pre = self.load_tensors(self.learning_spec.model_dir, 'fisher')
v_pre = self.load_tensors(self.learning_spec.model_dir, 'main')
meta_batch = self.combine_meta_features(g_cur, g_pre, v_pre, v_cur)
meta_output = self.meta_model(meta_batch)
tf.summary.scalar(name='losses/meta_output',
tensor=tf.reshape(meta_output, shape=[]))
self.total_loss = self.loss + self.alpha * meta_output * self.add_meta_loss(
g_pre, self.model.weights, v_pre)
tf.summary.scalar(name='losses/total_loss',
tensor=tf.reshape(self.total_loss, shape=[]))
total_gradient_computer = gc.ScopeGradientComputer(
self.opt, self.total_loss, self.model.weights)
total_grads_and_vars = total_gradient_computer.compute()
train_op = self.global_step_increase(grads_and_vars)
gradient_hook = self.compute_curvature(total_grads_and_vars)
return tf.estimator.EstimatorSpec(self.mode,
loss=self.total_loss,
train_op=train_op,
training_hooks=gradient_hook)
def create(self):
# current gradient
tf.summary.scalar(name='losses/cur_loss', tensor=self.loss)
gradient_computer = gc.ScopeGradientComputer(self.opt, self.loss,
self.model.weights)
grads_and_vars = gradient_computer.compute()
g_cur, v_cur = zip(*grads_and_vars)
joint_gradient_computer = gc.ScopeGradientComputer(
self.opt, self.joint_loss, self.model.weights)
joint_grads_and_vars = joint_gradient_computer.compute()
g_joint, _ = zip(*joint_grads_and_vars)
g_pre = self.load_tensors(self.meta_learning_spec.model_dir, 'fisher')
v_pre = self.load_tensors(self.meta_learning_spec.model_dir, 'main')
meta_batch = self.combine_meta_features(g_cur, g_pre, v_cur, v_pre)
meta_label = self.make_meta_labels(g_cur, g_joint, v_cur, v_pre, g_pre)
tf.summary.scalar(name='losses/meta_label',
tensor=tf.reshape(meta_label, []))
meta_output = self.meta_model(meta_batch)
tf.summary.scalar(name='losses/meta_output',
tensor=tf.reshape(meta_output, []))
meta_loss = tf.losses.absolute_difference(meta_output, meta_label)
tf.summary.scalar(name='losses/meta_loss', tensor=meta_loss)
meta_gradient_computer = gc.ScopeGradientComputer(
self.meta_opt, meta_loss, self.meta_model.weights)
meta_grads_and_vars = meta_gradient_computer.compute()
ops = self.global_step_increase_meta(meta_grads_and_vars)
gradient_hook = self.compute_curvature(grads_and_vars)
return tf.estimator.EstimatorSpec(self.mode,
loss=self.loss,
train_op=tf.group(ops),
training_hooks=gradient_hook)
def create(self):
gradient_computer = gc.ScopeGradientComputer(self.opt, self.loss,
self.model.weights)
grads_and_vars = gradient_computer.compute()
if self.mode == tf.estimator.ModeKeys.EVAL:
return self.evaluate(self.loss)
train_op = self.global_step_increase(grads_and_vars)
return tf.estimator.EstimatorSpec(self.mode,
loss=self.loss,
train_op=train_op)
def create(self):
self.loss = self.loss + self.alpha * self.add_ewc_loss(
self.model.weights)
gradient_computer = gc.ScopeGradientComputer(self.opt, self.loss,
self.model.weights)
grads_and_vars = gradient_computer.compute()
if self.mode == tf.estimator.ModeKeys.EVAL:
return self.evaluate(self.loss)
train_op = self.global_step_increase(grads_and_vars)
gradient_hook = self.compute_curvature(grads_and_vars)
return tf.estimator.EstimatorSpec(self.mode,
loss=self.loss,
train_op=train_op,
training_hooks=gradient_hook)
def create(self):
g_pre = self.load_tensors(self.learning_spec.model_dir, 'fisher')
v_pre = self.load_tensors(self.learning_spec.model_dir, 'center')
self.loss = self.loss + self.alpha * self.add_ewc_loss(
self.model.weights, v_pre, g_pre)
gradient_computer = gc.ScopeGradientComputer(self.opt, self.loss,
self.model.weights)
grads_and_vars = gradient_computer.compute()
if self.mode == tf.estimator.ModeKeys.EVAL:
return self.evaluate(self.loss)
train_op = self.global_step_increase(grads_and_vars)
gradient_hook = self.compute_curvature(grads_and_vars)
return tf.estimator.EstimatorSpec(self.mode,
loss=self.loss,
train_op=train_op,
training_hooks=gradient_hook)
