Open
Description
class NesterovAcceleratedGradient():
def init(self, learning_rate=0.001, momentum=0.4):
self.learning_rate = learning_rate
self.momentum = momentum
self.w_updt = None#np.array([])
def update(self, w, grad_func):
# Calculate the gradient of the loss a bit further down the slope from w
if self.w_updt is None:
self.w_updt = np.zeros(np.shape(w))
# print("shape og w",w.shape)
# print("shape og w",self.w_updt.shape)
approx_future_grad = np.clip((grad_func(w - self.momentum * self.w_updt)), -1, 1)
#print(approx_future_grad)
# Initialize on first update
if not self.w_updt.any():
self.w_updt = np.zeros(np.shape(w))
self.w_updt = self.momentum * self.w_updt + self.learning_rate * approx_future_grad
# Move against the gradient to minimize loss
return w - self.w_updt
Here grad_func is not implemented!
Metadata
Metadata
Assignees
Labels
No labels