Trjtdtk

I'm trying to implement Nokland's Direct Feedback Alignment in Python following his paper.
Here's my implementation so far:

import numpy as np
import matplotlib.pyplot as plt
from sklearn.utils.extmath import softmax
from scipy.special import expit

def f_sigmoid(x): return expit(x)
def df_sigmoid(x): return f_sigmoid(x) * (1-f_sigmoid(x))

f_activation = f_sigmoid
df_activation = df_sigmoid

class NeuralNet(object):

 def __init__(self, num_input, num_hidden, num_output):

 #Using Lillicrap's initialization between -0.5 and 0.5
 self.W1 = np.random.uniform(-0.5, 0.5)
 self.W2 = np.random.uniform(-0.5, 0.5)
 self.W3 = np.random.uniform(-0.5, 0.5)
 self.B1 = np.random.uniform(-0.5, 0.5, self.W1.size)
 self.B2 = np.random.uniform(-0.5, 0.5, self.W2.size)

 def forward(self, X): #HOW DO I INITIALIZE b1,b2,b3 ?
 a1 = np.matmul( X, self.W1) + self.b1
 h1 = f_activation(a1)

 a2 = np.matmul(h1, self.W2) + self.b2
 h2 = f_activation(a2)

 a_y = np.matmul(h2, self.W3) + self.b3
 y_hat = softmax(a_y)

 return y_hat, h2, h1, a1, a2

 def loss(self, predicted, target): pass

def backpropagation(self, X, y, lr):
 h3, h2, h1, a1, a2 = self.forward(X)
 y_hat = h3
 e = y_hat - y

 # Backpropagation
 d2 = df_activation(a2) * np.matmul(e, self.W3.T)
 d1 = df_activation(a1) * np.matmul(d2, self.W2.T)

 # Feedback Alignment
 """no W anymore!"""
 d2 = df_activation(a2) * np.matmul(e, self.B2)
 d1 = df_activation(a1) * np.matmul(d2, self.B1)

 # Direct Feedback Alignment 
 d2 = df_activation(a2) * np.matmul(e, self.B2)
 d1 = df_activation(a1) * np.matmul(e, self.B1)

 # Indirect Feedback Alignment
 d1 = df_activation(a1) * np.matmul(e, self.B1)
 d2 = df_activation(a2) * np.matmul(d1, self.W2)

 # Weights update, same for all methods
 self.W3 = self.W3 - lr * np.matmul(h2.T,e)
 self.W2 = self.W2 - lr * np.matmul(h1.T, d2)
 self.W1 = self.W1 - lr * np.matmul(X.reshape((-1,1)), d1)

 # Biases update, same for all methods
 self.b3 = self.b3 - lr *e
 self.b2 = self.b2 - lr * d2
 self.b1 = self.b1 - lr * d1

I don't know how to initialize b1,b2,b3 in forward function. I'm also not sure about W and B in __init__

Do you have any suggestions?