nn.py

# Testing simple 3 Hidden Layers Neural Network implementation in python
# Multiple examples are feed in single instance.

import numpy as np
from skimage.io import imread, imshow

def get_img(name):
	im = imread(name)
	im_clr = np.zeros((im.shape[0],im.shape[1]))
	for i in range(im.shape[0]):
		for j in range(im.shape[1]):
			im_clr[i][j] = np.sum(im[i][j])
	im_flat = np.reshape(im_clr, (im.shape[0]*im.shape[1],1));
	return im_flat

image_flat = get_img("test_3.png")
image_flat = np.append(image_flat, get_img("test_3_1.png"), axis=1)
image_flat = np.append(image_flat, get_img("test_3_2.png"), axis=1)
image_flat = np.append(image_flat, get_img("test_3_3.png"), axis=1)
image_flat = np.append(image_flat, get_img("test_3_4.png"), axis=1)
image_flat = np.append(image_flat, get_img("test_3_5.png"), axis=1)
image_flat = np.append(image_flat, get_img("test_3_6.png"), axis=1)
image_flat = np.append(image_flat, get_img("d1.jpg"), axis=1)
image_flat = np.append(image_flat, get_img("d2.png"), axis=1)
image_flat = np.append(image_flat, get_img("d3.jpeg"), axis=1)
image_flat = np.append(image_flat, get_img("d4.jpeg"), axis=1)
image_flat = np.append(image_flat, get_img("d5.jpeg"), axis=1)
image_flat = np.append(image_flat, get_img("d6.jpeg"), axis=1)


n=13
y=np.ones((1,n))
lr=1
epoch = 256;

w1 = np.random.randn(1024,image_flat.shape[0])
b1 = np.random.randn(1024,1)*0
z1 = np.dot(w1, image_flat) + b1
a1 = 1/(1+np.exp(z1 * -1))


w2 = np.random.randn(1024,a1.shape[0])
b2 = np.random.randn(1024,1)*0
z2 = np.dot(w2, a1) + b2
a2 = 1/(1+np.exp(z2 * -1))

w3 = np.random.randn(1024,a2.shape[0])
b3 = np.random.randn(1024,1)*0
z3 = np.dot(w3, a2) + b3
a3 = 1/(1+np.exp(z3 * -1))

w4 = np.random.randn(1,a3.shape[0])
b4 = np.random.randn(1,1)*0
z4 = np.dot(w4, a3) + b4
a4 = 1/(1+np.exp(z4 * -1))

#print("INIT Weights: \n", w2, w3, w4)
print("OUT: ", a4)

cost = np.sum((a4-y)**2/2, axis=1)/n
cost_prime = a4-y
#cost_2 = -(1 * np.log(a3) + (1-1)*np.log(1-a3))
print("Cost: ", cost)

while epoch:
	epoch -=1
	z4_prime = a4 * (1-a4)
	d4 = cost_prime * z4_prime
	pd4 = np.sum(a4 * d4, axis=1, keepdims=True)/n

	z3_prime = a3 * (1-a3)
	d3 = np.dot(np.transpose(w4),d4) * z3_prime
	pd3 = np.sum(a3 * d3, axis=1, keepdims=True)/n

	z2_prime = a2 * (1-a2)
	d2 = np.dot(np.transpose(w3), d3) * z2_prime
	pd2 = np.sum(a2 * d2, axis=1, keepdims=True)/n

	z1_prime = a1 * (1-a1)
	d1 = np.dot(np.transpose(w2), d2) * z1_prime
	pd1 = np.sum(a1 * d1, axis=1, keepdims=True)/n


	w4 -= lr*pd4
	w3 -= lr*pd3
	w2 -= lr*pd2
	w1 -= lr*pd1
	
	b4 = -lr*np.sum(z4_prime, axis=1, keepdims=True)/n
	b3 = -lr*np.sum(z3_prime, axis=1, keepdims=True)/n
	b2 = -lr*np.sum(z2_prime, axis=1, keepdims=True)/n
	b1 = -lr*np.sum(z1_prime, axis=1, keepdims=True)/n


	z1 = np.dot(w1, image_flat) + b1
	a1 = 1/(1+np.exp(z1 * -1))

	z2 = np.dot(w2, a1) + b2
	a2 = 1/(1+np.exp(z2 * -1))

	z3 = np.dot(w3, a2) + b3
	a3 = 1/(1+np.exp(z3 * -1))

	z4 = np.dot(w4, a3) + b4
	a4 = 1/(1+np.exp(z4 * -1))
	#print("OUT: ",a4)

	cost = np.sum((a4-y)**2/2, axis=1)/n
	cost_prime = a4-y
	#cost_2 = -(1 * np.log(a3) + (1-1)*np.log(1-a3))

print("Final OUT: ",a4)
print("Final Cost after n iter: ", cost)


def predict(img_name):
	test = get_img(img_name)
	z1 = np.dot(w1, test)
	a1 = 1/(1+np.exp(z1 * -1))
	z2 = np.dot(w2, a1)
	a2 = 1/(1+np.exp(z2 * -1))
	z3 = np.dot(w3, a2)
	a3 = 1/(1+np.exp(z3 * -1))
	z4 = np.dot(w4, a3)
	a4 = 1/(1+np.exp(z4 * -1))
	print("Predict OUT: ",a4)

#print("Final Weights: \n", w2, w3, w4)
predict("test_3174.png")
predict("test_3.png")
predict("ss.png")
predict("d7.jpeg")
predict("d8.jpeg")
predict("d9.jpeg")