FCN

import tensorflow as tf
import numpy as np
import os
import struct
import cv2
from matplotlib import pyplot as plt

def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)

def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)

def conv2d(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

#Variable initialization
x = tf.placeholder(tf.float32, shape=[None, 384, 512, 3])
y_ = tf.placeholder(tf.float32, shape=[None, 48, 64])
x_image = tf.reshape(x, [-1, 384, 512, 3])

#Network forward-pass definition

#First convolutional layer(2 convolutions)
W_conv1_1 = weight_variable([5, 5, 3, 64])
b_conv1_1 = bias_variable([64])
h_conv1_1 = tf.nn.relu(conv2d(x_image, W_conv1_1) + b_conv1_1)
W_conv1_2 = weight_variable([5, 5, 64, 64])
b_conv1_2 = bias_variable([64])
h_conv1_2 = tf.nn.relu(conv2d(h_conv1_1, W_conv1_2) + b_conv1_2)
h_pool1 = max_pool_2x2(h_conv1_2)
#Second convolutional layer(2 convolutions)
W_conv2_1 = weight_variable([5, 5, 64, 64])
b_conv2_1 = bias_variable([64])
h_conv2_1 = tf.nn.relu(conv2d(h_pool1, W_conv2_1) + b_conv2_1)
W_conv2_2 = weight_variable([5, 5, 64, 64])
b_conv2_2 = bias_variable([64])
h_conv2_2 = tf.nn.relu(conv2d(h_conv2_1, W_conv2_2) + b_conv2_2)
h_pool2 = max_pool_2x2(h_conv2_2)
#Third convolutional layer(3 convolutions)
W_conv3_1 = weight_variable([5, 5, 64, 64])
b_conv3_1 = bias_variable([64])
h_conv3_1 = tf.nn.relu(conv2d(h_pool2, W_conv3_1) + b_conv3_1)
W_conv3_2 = weight_variable([5, 5, 64, 64])
b_conv3_2 = bias_variable([64])
h_conv3_2 = tf.nn.relu(conv2d(h_conv3_1, W_conv3_2) + b_conv3_2)
W_conv3_3 = weight_variable([5, 5, 64, 64])
b_conv3_3 = bias_variable([64])
h_conv3_3 = tf.nn.relu(conv2d(h_conv3_2, W_conv3_3) + b_conv3_3)
h_pool3 = max_pool_2x2(h_conv3_3)
#Fourth convolutional layer(3 convolutions)
W_conv4_1 = weight_variable([5, 5, 64, 64])
b_conv4_1 = bias_variable([64])
h_conv4_1 = tf.nn.relu(conv2d(h_pool3, W_conv4_1) + b_conv4_1)
W_conv4_2 = weight_variable([5, 5, 64, 64])
b_conv4_2 = bias_variable([64])
h_conv4_2 = tf.nn.relu(conv2d(h_conv4_1, W_conv4_2) + b_conv4_2)
W_conv4_3 = weight_variable([5, 5, 64, 64])
b_conv4_3 = bias_variable([64])
h_conv4_3 = tf.nn.relu(conv2d(h_conv4_2, W_conv4_3) + b_conv4_3)
h_pool4 = max_pool_2x2(h_conv4_3)

#Train step definition
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=h_pool4))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
#Initializing the saver to save our model after training
saver = tf.train.Saver()
#Training
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    images = np.load("dataset/images.npy")
    labels = np.load("dataset/labels.npy")
    for i in range(2, 100):
        train_step.run(feed_dict={x: images[i % 29].reshape(1, 384, 512, 3), y_: labels[i % 29].reshape(1, 48, 64)})
        print("Step: %f; Cross-entropy: %d"%(i, cross_entropy.eval(feed_dict={x: images[i % 29].reshape(1, 384, 512, 3), y_: labels[i % 29].reshape(1, 48, 64)})))
    save_path = saver.save(sess, "Kernels/model")