Backpropagation code for deep neural net

backpropagation.py

def backprop_this_layer(self, da_curr, z_curr, W_curr, b_curr, A_prev, activation_function):
  if activation_function is 'sigmoid':
    activation_back = self.sigmoid_backward 
  elif activation_function is 'relu':
    activation_back = self.relu_backward
  else:
    return
  m = A_prev.shape[1]
  
  dz_curr = activation_back(da_curr, z_curr)
  dw_curr = np.dot(dz_curr, A_prev.T)/m
  db_curr = np.sum(dz_curr, axis=1, keepdims=True)/m
  da_prev = np.dot(W_curr.T, dz_curr)

  return da_prev, dw_curr, db_curr

def backward(self, ytrue, ypred, cache):
  grads = {}
  m = ytrue.shape[1]
  da_prev = np.divide(1-ytrue, 1-ypred) - np.divide(ytrue, ypred)
  for prev_layer_id, layer in reversed(list(enumerate(self.architecture))):
    layer_id = prev_layer_id + 1
    activation = layer['activation']

    da_curr = da_prev
    A_prev = cache['A'+str(prev_layer_id)]
    Z_curr = cache['Z'+str(layer_id)]

    W_curr = self.params['W'+str(layer_id)]
    b_curr = self.params['b'+str(layer_id)]

    da_prev, dw_curr, db_curr = self.backprop_this_layer(
        da_curr, Z_curr, W_curr, b_curr, A_prev, activation)
    grads["dw"+str(layer_id)] = dw_curr
    grads['db'+str(layer_id)] = db_curr
  return grads