Neural nets not learning mnist dataset

Question

I tried training a 2 hidden layer network using the mnist dataset, but I am not getting any results. I have tried tuning the learning rate(tried 0.1 and 0.0001) and the number of epochs(tried 10 and 50). I even changed the size of hidden layer from 10 to 250. First i had initialized the weights between 0 and 1 and was getting the same classification for all test samples but added (-) sign to 50% of them(chose the figure of 50% by myself) and that problem was solved. Now I cant figure out why it is not working.

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import math
from sklearn.preprocessing import StandardScaler

scaler = StandardScaler() 

def to_array(img):
    img = np.asarray(img)
    return img

'''def standardize(gray):
    st_gray = (gray-np.amin(gray))/(np.amax(gray)-np.amin(gray))
    return st_gray
'''
def activ_func(x):
    for i in range(x.shape[0]):
        '''x[i][0]=(1-np.e**(-2*x[i][0]))/(1+np.e**(-2*x[i][0]))'''
        x[i][0] = 1/(1+np.e**(-x[i][0]))
    return x

def deriv_activ_func(x):
    for i in range(x.shape[0]):
        '''x[i][0] = 1-math.pow(x[i][0],2)'''
        x[i][0] = (x[i][0])*(1-x[i][0])
    return x

def cost(out_layer, label, ind):
    cost = (out_layer[ind]-label)**2
    return cost

def update(x, grad, r):
    for i in range(x.shape[0]):
        x[i][0] = x[i][0]+r*grad[i][0]
    return x

path = "mnist/mnist_train.csv"
gray = pd.read_csv(path)
labels = gray['label']
gray = gray.drop(['label'], axis=1)
gray = to_array(gray)
labels = to_array(labels)
st_gray = np.empty(shape=(gray.shape[1],1))

def rand_sign(w):
    n = np.random.randint(2,size=w.shape[0]*w.shape[1]).reshape(w.shape[0],w.shape[1])
    for i in range(w.shape[0]):
        for j in range(w.shape[1]):
            if(n[i][j]==1):
                w[i][j]=(-1)*w[i][j]
    return w

def initialize():
    in_layer = np.empty(shape=(st_gray.shape[0],1))
    out_layer = np.unique(labels).reshape(-1,1)
    w1 = rand_sign(np.random.rand(250,in_layer.shape[0]))
    b1 = rand_sign(np.random.rand(250,1))
    l1 = np.empty(shape=(250,1))
    w2 = rand_sign(np.random.rand(250,l1.shape[0]))
    b2 = rand_sign(np.random.rand(250,1))
    l2 = np.empty(shape=(250,1))
    w3 = rand_sign(np.random.rand(out_layer.shape[0],l2.shape[0]))
    b3 = rand_sign(np.random.rand(out_layer.shape[0],1))
    l3 = np.empty_like(out_layer)
    return l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer,out_layer

def feed_forward(l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer,i):
    st_gray = scaler.fit_transform(gray[i][:].reshape(-1,1))
    in_layer = st_gray
    l1 = np.dot(w1,in_layer)+b1
    l1 = activ_func(l1)
    l2 = np.dot(w2,l1)+b2
    l2 = activ_func(l2)
    l3 = np.dot(w3,l2)+b3
    l3 = activ_func(l3)
    return l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer

def one_hot(out_layer,label):
    for j in range(out_layer.shape[0]):
        if(out_layer[j][0]==label):
            out_layer[j][0] = 1
        else:
            out_layer[j][0] = 0
    return out_layer 

def back_prop(l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer,out_layer,lr):
    error = out_layer-l3
    grad = np.dot(error*deriv_activ_func(l3),l2.T)
    w3 = update(w3, grad, lr)
    grad = error*deriv_activ_func(l3)
    b3 = update(b3, grad, lr)
    grad = np.dot(w3.T,error*deriv_activ_func(l3))
    error = grad
    grad = error*deriv_activ_func(l2)
    w2 = update(w2, grad, lr)
    grad = error*deriv_activ_func(l2)
    b2 = update(b2, grad, lr)
    grad = np.dot(w2.T,error*deriv_activ_func(l2))
    error = grad
    grad = error*deriv_activ_func(l1)
    w1 = update(w1, grad, lr)
    grad = error*deriv_activ_func(l1)
    b1 = update(b1, grad, lr)
    return l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer

def predict(l3):
    out = np.amax(l3)
    count = 0
    for j in range(l3.shape[0]):
        count=count+1
        if(l3[j]==out):
            break
    return count

def trainer():
    l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer,out_layer = initialize()
    for epochs in range(50):
        for i in range(gray.shape[0]):
            out_layer = np.unique(labels).reshape(-1,1)
            l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer = feed_forward(l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer,i)
            out_layer = one_hot(out_layer,labels[i])
            l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer = back_prop(l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer,out_layer,0.0001)
        print("End of epoch :",epochs+1) 
    return l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer,out_layer

l1,l2,l3,w1,w2,w3,b1,b2,b3,in_layer,out_layer = trainer()

path = "mnist/mnist_train.csv"
gray = pd.read_csv(path)
labels = gray['label']
gray = gray.drop(['label'], axis=1)
gray = to_array(gray)
labels = to_array(labels)
st_gray = np.empty(shape=(gray.shape[1],1))

for i in range(10):
    st_gray = scaler.fit_transform(gray[i][:].reshape(-1,1))
    in_layer = st_gray
    l1 = np.dot(w1,in_layer)+b1
    l1 = activ_func(l1)
    l2 = np.dot(w2,l1)+b2
    l2 = activ_func(l2)
    l3 = np.dot(w3,l2)+b3
    l3 = activ_func(l3)
    count = predict(l3)
    print("Expected: ",labels[i]," Predicted: ",count)