train.py

import os
import cv2
import matplotlib.pyplot as plt
import numpy as np
from tqdm import tqdm
from random import shuffle
import tensorflow as tf
import pandas as pd
from sklearn.metrics import roc_auc_score

#HyperParameters
epochs = 50
step_size = 8
IMG_SIZE_ALEXNET = 227 # image size
validating_size = 20 # while cross validating, we are evaluating batch by batch
nodes_fc1 = 4096 # no of nodes on fc layer 1
nodes_fc2 = 4096 # no of nodes on fc layer 2
output_classes = 3 # three classes: eggplant, 
output_locations = 4 # minx, miny, maxx, maxy

TRAIN_DIR = os.getcwd()

#Separating train and validation sets
data = np.load('object_localization.npy')
train = data[:int(len(data)*0.8)]
cv = data[int(len(data)*0.8):]

# X for train input
X = np.array([i[0] for i in train]).reshape(-1,IMG_SIZE_ALEXNET,IMG_SIZE_ALEXNET,3)
# Y1 for classification head
Y1 = np.array([i[1] for i in train])
# Y2 for regression head
Y2 = np.array([i[2] for i in train])

# cv_x for train input
cv_x = np.array([i[0] for i in cv]).reshape(-1,IMG_SIZE_ALEXNET,IMG_SIZE_ALEXNET,3)
# cv_y1 for classification head
cv_y1 = np.array([i[1] for i in cv])
# cv_y2 for regression head
cv_y2 = np.array([i[2] for i in cv])

print(cv_y1[:10])
print(X.shape)

print(Y1.shape)

print(cv_x.shape)

print(cv_y1.shape)

#How many trainin images are kept as 'steps'
steps = len(train)
print(steps)
remaining = steps % step_size

#Resetting graph
tf.reset_default_graph()

#Defining Placeholders
x = tf.placeholder(tf.float32,shape=[None,IMG_SIZE_ALEXNET,IMG_SIZE_ALEXNET,3])
y_true_1 = tf.placeholder(tf.float32,shape=[None,output_classes])
y_true_2 = tf.placeholder(tf.float32,shape=[None,output_locations])

##CONVOLUTION LAYER 1
#Weights for layer 1
w_1 = tf.Variable(tf.truncated_normal([11,11,3,96], stddev=0.01))
#Bias for layer 1
b_1 = tf.Variable(tf.constant(0.0, shape=[[11,11,3,96][3]]))
#Applying convolution
c_1 = tf.nn.conv2d(x, w_1,strides=[1, 4, 4, 1], padding='VALID')
#Adding bias
c_1 = c_1 + b_1
#Applying RELU
c_1 = tf.nn.relu(c_1)
								
print(c_1)
##POOLING LAYER1
p_1 = tf.nn.max_pool(c_1, ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1], padding='VALID')
print(p_1)

##CONVOLUTION LAYER 2
#Weights for layer 2
w_2 = tf.Variable(tf.truncated_normal([5,5,96,256], stddev=0.01))
#Bias for layer 2
b_2 = tf.Variable(tf.constant(1.0, shape=[[5,5,96,256][3]]))
#Applying convolution
c_2 = tf.nn.conv2d(p_1, w_2,strides=[1, 1, 1, 1], padding='SAME')
#Adding bias
c_2 = c_2 + b_2
#Applying RELU
c_2 = tf.nn.relu(c_2)

print(c_2)

##POOLING LAYER2
p_2 = tf.nn.max_pool(c_2, ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1], padding='VALID')
print(p_2)

##CONVOLUTION LAYER 3
#Weights for layer 3
w_3 = tf.Variable(tf.truncated_normal([3, 3, 256, 384], stddev=0.01))
#Bias for layer 3
b_3 = tf.Variable(tf.constant(0.0, shape=[[3, 3, 256, 384][3]]))
#Applying convolution
c_3 = tf.nn.conv2d(p_2, w_3,strides=[1, 1, 1, 1], padding='SAME')
#Adding bias
c_3 = c_3 + b_3
#Applying RELU
c_3 = tf.nn.relu(c_3)

print(c_3)

##CONVOLUTION LAYER 4
#Weights for layer 4
w_4 = tf.Variable(tf.truncated_normal([3, 3, 384, 384], stddev=0.01))
#Bias for layer 4
b_4 = tf.Variable(tf.constant(0.0, shape=[[3, 3, 384, 384][3]]))
#Applying convolution
c_4 = tf.nn.conv2d(c_3, w_4,strides=[1, 1, 1, 1], padding='SAME')
#Adding bias
c_4 = c_4 + b_4
#Applying RELU
c_4 = tf.nn.relu(c_4)

print(c_4)

##CONVOLUTION LAYER 5
#Weights for layer 5
w_5 = tf.Variable(tf.truncated_normal([3, 3, 384, 256], stddev=0.01))
#Bias for layer 5
b_5 = tf.Variable(tf.constant(0.0, shape=[[3, 3, 384, 256][3]]))
#Applying convolution
c_5 = tf.nn.conv2d(c_4, w_5,strides=[1, 1, 1, 1], padding='SAME')
#Adding bias
c_5 = c_5 + b_5
#Applying RELU
c_5 = tf.nn.relu(c_5)

print(c_5)

##POOLING LAYER3
p_3 = tf.nn.max_pool(c_5, ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1], padding='VALID')
print(p_3)

#Flattening
flattened = tf.reshape(p_3,[-1,6*6*256])
print(flattened)

##Fully Connected Layer 1
#Getting input nodes in FC layer 1
input_size = int( flattened.get_shape()[1] )
#Weights for FC Layer 1
w1_fc = tf.Variable(tf.truncated_normal([input_size, nodes_fc1], stddev=0.01))
#Bias for FC Layer 1
b1_fc = tf.Variable( tf.constant(1.0, shape=[nodes_fc1] ) )
#Summing Matrix calculations and bias
s_fc1 = tf.matmul(flattened, w1_fc) + b1_fc
#Applying RELU
s_fc1 = tf.nn.relu(s_fc1)

#Dropout Layer 1
hold_prob1 = tf.placeholder(tf.float32)
s_fc1 = tf.nn.dropout(s_fc1,keep_prob=hold_prob1)

print(s_fc1)

##Fully Connected Layer 2
#Weights for FC Layer 2
w2_fc = tf.Variable(tf.truncated_normal([nodes_fc1, nodes_fc2], stddev=0.01))
#Bias for FC Layer 2
b2_fc = tf.Variable( tf.constant(1.0, shape=[nodes_fc2] ) )
#Summing Matrix calculations and bias
s_fc2 = tf.matmul(s_fc1, w2_fc) + b2_fc
#Applying RELU
s_fc2 = tf.nn.relu(s_fc2)
print(s_fc2)

#Dropout Layer 2
hold_prob2 = tf.placeholder(tf.float32)
s_fc2 = tf.nn.dropout(s_fc2,keep_prob=hold_prob1)

##Fully Connected Layer 3 -- CLASSIFICATION HEAD
#Weights for FC Layer 3
w3_fc_1 = tf.Variable(tf.truncated_normal([nodes_fc2,output_classes], stddev=0.01))
#Bias for FC Layer 3b3_fc = tf.Variable( tf.constant(1.0, shape=[output_classes] ) )
b3_fc_1 = tf.Variable( tf.constant(1.0, shape=[output_classes] ) )
#Summing Matrix calculations and bias
y_pred_1 = tf.matmul(s_fc2, w3_fc_1) + b3_fc_1
#Applying RELU
print(y_pred_1)

##Fully Connected Layer 3 -- REGRESSION HEAD
#Weights for FC Layer 3
w3_fc_2 = tf.Variable(tf.truncated_normal([nodes_fc2,output_locations], stddev=0.01))
#Bias for FC Layer 3b3_fc = tf.Variable( tf.constant(1.0, shape=[output_classes] ) )
b3_fc_2 = tf.Variable( tf.constant(1.0, shape=[output_locations] ) )
#Summing Matrix calculations and bias
y_pred_2 = tf.matmul(s_fc2, w3_fc_2) + b3_fc_2
#Applying RELU
print(y_pred_2)

#Defining Classification function
cross_entropy = tf.multiply(tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_true_1,logits=y_pred_1)),10)

#Defining Regression Loss
regression_loss = tf.multiply(tf.reduce_mean(tf.square(y_pred_2 - y_true_2)),1.0)

#Defining total loss
final_loss = cross_entropy + regression_loss

#Defining objective
train = tf.train.AdamOptimizer(learning_rate=0.00001).minimize(final_loss)

#Defining Accuracy
matches = tf.equal(tf.argmax(y_pred_1,1),tf.argmax(y_true_1,1))
acc = tf.reduce_mean(tf.cast(matches,tf.float32))

#Initializing weights
init = tf.global_variables_initializer()

#Starting Empty lists to keep results
acc_list = []
auc_list = []
loss_list = []
regression_list = []
#In order to save, creating a tf.train.Saver() object.
saver = tf.train.Saver()

#GPU settings
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
config.gpu_options.allocator_type = 'BFC'

tf.add_to_collection("classification_head", y_pred_1)
tf.add_to_collection("regression_head", y_pred_2)

def main():
    with tf.Session(config=config) as sess:
        sess.run(init)
        for i in range(epochs):
            for j in range(0,steps-remaining,step_size):
				#Feeding step_size-amount data with 0.5 keeping probabilities on DROPOUT LAYERS
            	_,c = sess.run([train,final_loss],
				feed_dict = {x:X[j:j+step_size], y_true_1:Y1[j:j+step_size],
				y_true_2:Y2[j:j+step_size],hold_prob1:0.5,hold_prob2:0.5})
			
		    #Writing for loop to calculate test statistics. GTX 1050 isn't able to calculate all cv data.
            cv_auc_list = []
            cv_acc_list = []
            cv_loss_list = []
            cv_regression_list = []
            for v in range(0,len(cv_x)-int(len(cv_x) % validating_size),validating_size):
                acc_on_cv,loss_on_cv,preds,coordinates = sess.run([acc,cross_entropy,tf.nn.softmax(y_pred_1),y_pred_2],
                feed_dict={x:cv_x[v:v+validating_size], y_true_1:cv_y1[v:v+validating_size], y_true_2:cv_y2[v:v+validating_size],
				  hold_prob1:1.0,hold_prob2:1.0})
				
                auc_on_cv = roc_auc_score(cv_y1[v:v+validating_size],preds)
                regression_loss = np.mean(pow(cv_y2[v:v+validating_size] - coordinates , 2 ) )
                cv_acc_list.append(acc_on_cv)
                cv_auc_list.append(auc_on_cv)
                cv_loss_list.append(loss_on_cv)
                cv_regression_list.append(regression_loss)
            acc_cv_ = round(np.mean(cv_acc_list),5)
            auc_cv_ = round(np.mean(cv_auc_list),5)
            loss_cv_ = round(np.mean(cv_loss_list),5)
            regression_loss_cv_ = round(np.mean(cv_loss_list),5)
            acc_list.append(acc_cv_)
            auc_list.append(auc_cv_)
            loss_list.append(loss_cv_)
            regression_list.append(regression_loss_cv_)
            print("Epoch:",i,"Accuracy:",acc_cv_,"Loss:",loss_cv_ ,"AUC:",auc_cv_)
        print("Training has finished and model is saved")
        saver.save(sess, os.path.join(os.getcwd(),"CNN_OL.ckpt"))

if __name__ == "__main__":
    main()