def relu(x): return (np.maximum(0, x))
for i in range(20): # Do not change this, we will compare performance at 1000 epochs
    # Forward pass
    L1 = relu(W1.dot(X))
    L2 = relu(W2.dot(L1))
    L3 = sigmoid(W3.dot(L2))
    # Backward pass
    dW3 = (L3 - T) * L3*(1 - L3)
    dW2 = W3.T.dot(dW3)
    dW2[(W2.dot(L1) <= 0)] = 0
    dW1 = W2.T.dot(dW2)
    dW1[(W1.dot(X) <=0)] = 0
    # Update
    W3 -= lr*np.dot(dW3, L2.T)
    W2 -= lr*np.dot(dW2, L1.T)
    W1 -= lr*np.dot(dW1, X.T)

    loss = np.sum((L3 - T)**2)/len(T.T)