def compute_dz2_dW2(a1, c):
'''
Compute local gradient of the logits function z2 w.r.t. the weights W2.
Input:
a1: the activations of sigmoid function, a numpy float vector of shape h by 1.
Output:
dz2_dW2: the partial gradient of logits z2 w.r.t. the weight matrix W2, a numpy float matrix of shape (c by h).
The (i,j)-th element represents the partial gradient of the i-th logit (z2[i]) w.r.t. the weight W2[i,j]: d_z2[i] / d_W2[i,j]
'''
dz2_dW2 = sr.compute_dz_dW(a1, c)
return dz2_dW2
def compute_dz1_dW1(x, h):
'''
Compute local gradient of the logits function z1 w.r.t. the weights W1 in the 1st layer.
Input:
x: the feature vector of a data instance, a float numpy vector of shape p by 1. Here p is the number of features/dimensions.
h: the number of output activations in the first layer, an integer.
Output:
dz1_dW1: the partial gradient of logits z1 w.r.t. the weight matrix W1, a numpy float matrix of shape (h by p).
The (i,j)-th element represents the partial gradient of the i-th logit (z1[i]) w.r.t. the weight W1[i,j]: d_z1[i] / d_W1[i,j]
'''
dz1_dW1 = sr.compute_dz_dW(x, h)
return dz1_dW1