Nestrov Accelerated Gradient Descent

[DataSenseiAryan]

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!

[MikeMavrok]

can i contribute?