def update_gradient_with_optimizer(self, x: Variable,
optimizer: Optimizer):
# print(type(x))
# Gradient Clipping
mask = (x.gradient < GRADIENT_CLIPPING_THRESHOLD).astype(int)
mask = np.multiply(
mask, (x.gradient > -GRADIENT_CLIPPING_THRESHOLD).astype(int))
contra_mask = 1 - mask
x.gradient = np.multiply(
mask, x.gradient) + contra_mask * GRADIENT_CLIPPING_THRESHOLD
if x.back_prop is not None:
# which means x is an input node
x.back_prop()
if x.trainable:
optimizer.update_once(x)
if x.lchild is not None:
self.update_gradient_with_optimizer(x.lchild, optimizer)
if x.rchild is not None:
self.update_gradient_with_optimizer(x.rchild, optimizer)
def set_and_update_gradient(self, x: Variable, gradient):
assert x.gradient.shape == gradient.shape
x.gradient = gradient
self.update_gradient(x)
