def compute_da2_dz2(a2):
'''
Compute local gradient of the softmax activations a2 w.r.t. the logits z2 in the 2nd layer.
Input:
a2: the activation values of softmax function, a numpy float vector of shape c by 1. Here c is the number of classes.
Output:
da2_dz2: the local gradient of the activations a2 w.r.t. the logits z2, a float numpy matrix of shape (c by c).
The (i,j)-th element represents the partial gradient ( d_a2[i] / d_z2[j] )
'''
da2_dz2 = sr.compute_da_dz(a2)
return da2_dz2
def compute_da1_dz1(a1):
'''
Compute local gradient of the sigmoid activations a1 w.r.t. the logits z1 in the first layer.
Input:
a1: the activations of sigmoid function, a numpy float vector of shape h by 1.
a1: the non-linear activations in the 1st layer.
Output:
da1_dz1: the local gradient of the activations a1 w.r.t. the logits z1, a float numpy vector of shape h by 1.
The i-th element of da1_dz1 represents the partial gradient ( d_a1[i] / d_z1[i] )
Hint: you could solve this problem using 1 line of code.
'''
#########################################
## INSERT YOUR CODE HERE
da1_dz1 = np.asmatrix(np.diagonal(sr.compute_da_dz(a1))).T
#########################################
return da1_dz1
