Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly
Sign up
Appearance settings

Commit 6bbecd8

Browse files
Add files via upload
1 parent 5afb5e5 commit 6bbecd8

File tree

2 files changed

+544
-0
lines changed

2 files changed

+544
-0
lines changed

‎multiclass_classification_cpu.py

Lines changed: 267 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,267 @@
1+
"""
2+
Created on Sat Mar 24 11:25:41 2018
3+
4+
@author: Muhammed Buyukkinaci
5+
"""
6+
#Importing libraries
7+
import tensorflow as tf # ML library for graphs
8+
import cv2 # image processing
9+
import numpy as np # mathematical operations
10+
import os # working with directories
11+
from random import shuffle # mixing up or currently ordered data that might lead our network astray in training.
12+
from tqdm import tqdm # a nice progress bar for loops
13+
14+
IMG_SIZE = 80# our images are 80x80x3
15+
16+
print(os.getcwd())
17+
18+
"""
19+
#Converting the output into one-hot format
20+
def label_img(img):
21+
word_label = img.split('_')[0]
22+
if word_label == 'chair': return [1,0,0,0]
23+
elif word_label == 'kitchen': return [0,1,0,0]
24+
elif word_label == 'knife': return [0,0,1,0]
25+
elif word_label == 'saucepan': return [0,0,0,1]
26+
"""
27+
"""
28+
#A function to read images from training folder
29+
def create_train_data():
30+
training_data = []
31+
for img in tqdm(os.listdir(TRAIN_DIR)):
32+
label = label_img(img)
33+
path = TRAIN_DIR + '\\' + img
34+
img = cv2.imread(path,1)
35+
img = cv2.resize(img, (IMG_SIZE,IMG_SIZE))
36+
training_data.append([np.array(img),np.array(label)])
37+
shuffle(training_data)
38+
np.save('train_data_mc.npy', training_data)
39+
return training_data
40+
"""
41+
#If you are reading images, make the line below uncomment
42+
#train_data = create_train_data()
43+
#If you are reading .npy file, use the line below.
44+
train_data = np.load('train_data_mc.npy')
45+
46+
"""
47+
#A function to read images from testing folder
48+
def process_test_data():
49+
testing_data = []
50+
for img in tqdm(os.listdir(TEST_DIR)):
51+
label = label_img(img)
52+
path = TEST_DIR + '\\' + img
53+
img = cv2.imread(path,1)
54+
img = cv2.resize(img, (IMG_SIZE,IMG_SIZE))
55+
testing_data.append([np.array(img),label])
56+
shuffle(testing_data)
57+
np.save('test_data_mc.npy', testing_data)
58+
return testing_data
59+
"""
60+
#If you are reading images, make the line below uncomment.
61+
62+
#test_data = process_test_data()
63+
64+
#If you are reading .npy file, use the line below.
65+
test_data = np.load('test_data_mc.npy')
66+
67+
#For visualizations
68+
import matplotlib.pyplot as plt
69+
70+
#A CHAIR
71+
plt.imshow(np.array(train_data[180][0]))
72+
plt.show()
73+
74+
#KITCHEN
75+
plt.imshow(np.array(train_data[100][0]))
76+
plt.show()
77+
78+
#KNIFE
79+
plt.imshow(np.array(train_data[111][0]))
80+
plt.show()
81+
82+
#SAUCEPAN
83+
plt.imshow(np.array(train_data[154][0]))
84+
plt.show()
85+
86+
87+
#Splitting train and CV data
88+
train = train_data[:4800]
89+
cv = train_data[4800:]
90+
91+
X = np.array([i[0] for i in train]).reshape(-1,IMG_SIZE,IMG_SIZE,3)
92+
Y = np.array([i[1] for i in train])
93+
94+
cv_x = np.array([i[0] for i in cv]).reshape(-1,IMG_SIZE,IMG_SIZE,3)
95+
cv_y = np.array([i[1] for i in cv])
96+
97+
test_x = np.array([i[0] for i in test_data]).reshape(-1,IMG_SIZE,IMG_SIZE,3)
98+
test_y = np.array([i[1] for i in test_data])
99+
100+
print(X.shape)
101+
print(Y[2])
102+
print(cv_x.shape)
103+
print(test_x.shape)
104+
105+
#HYPERPARAMETERS
106+
steps = 4800
107+
epochs = 60
108+
step_size = 16
109+
total_batch = int(steps/step_size)
110+
LR = 0.00001
111+
112+
#FUNCTIONS TO CODE LAYERS
113+
def init_weights(shape):
114+
init_random_dist = tf.truncated_normal(shape, stddev=0.1)
115+
return tf.Variable(init_random_dist)
116+
117+
def init_bias(shape):
118+
init_bias_vals = tf.constant(0.1, shape=shape)
119+
return tf.Variable(init_bias_vals)
120+
121+
def conv2d(x, W):
122+
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
123+
124+
def max_pool_2by2(x):
125+
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
126+
strides=[1, 2, 2, 1], padding='SAME')
127+
128+
def convolutional_layer(input_x, shape):
129+
W = init_weights(shape)
130+
b = init_bias([shape[3]])
131+
return tf.nn.relu(conv2d(input_x, W) + b)
132+
133+
def normal_full_layer(input_layer, size):
134+
input_size = int(input_layer.get_shape()[1])
135+
W = init_weights([input_size, size])
136+
b = init_bias([size])
137+
return tf.matmul(input_layer, W) + b
138+
139+
print(X[5].shape)
140+
141+
#GRAPH
142+
with tf.Session() as sess:
143+
tf.reset_default_graph()
144+
145+
#Defining Placeholdera
146+
x = tf.placeholder(tf.float32,shape=[None,80,80,3])
147+
y_true = tf.placeholder(tf.float32,shape=[None,4])
148+
149+
#Defining GRAPH
150+
with tf.name_scope('Model'):
151+
convo_1 = convolutional_layer(x,shape=[6,6,3,32])
152+
convo_1_pooling = max_pool_2by2(convo_1)
153+
convo_2 = convolutional_layer(convo_1_pooling,shape=[6,6,32,64])
154+
convo_2_pooling = max_pool_2by2(convo_2)
155+
convo_3 = convolutional_layer(convo_2_pooling,shape=[6,6,64,64])
156+
convo_3_pooling = max_pool_2by2(convo_3)
157+
convo_4 = convolutional_layer(convo_3_pooling,shape=[6,6,64,128])
158+
convo_4_pooling = max_pool_2by2(convo_4)
159+
convo_2_flat = tf.reshape(convo_4_pooling,[-1,5*5*128])
160+
161+
full_layer_one = tf.nn.relu(normal_full_layer(convo_2_flat,4096))
162+
hold_prob1 = tf.placeholder(tf.float32)
163+
full_one_dropout = tf.nn.dropout(full_layer_one,keep_prob=hold_prob1)
164+
165+
full_layer_two = tf.nn.relu(normal_full_layer(full_one_dropout,2048))
166+
hold_prob2 = tf.placeholder(tf.float32)
167+
full_two_dropout = tf.nn.dropout(full_layer_two,keep_prob=hold_prob2)
168+
169+
full_layer_three = tf.nn.relu(normal_full_layer(full_two_dropout,1024))
170+
hold_prob3 = tf.placeholder(tf.float32)
171+
full_three_dropout = tf.nn.dropout(full_layer_three,keep_prob=hold_prob3)
172+
173+
full_layer_four = tf.nn.relu(normal_full_layer(full_three_dropout,512))
174+
hold_prob4 = tf.placeholder(tf.float32)
175+
full_four_dropout = tf.nn.dropout(full_layer_four,keep_prob=hold_prob4)
176+
177+
y_pred = normal_full_layer(full_four_dropout,4)
178+
179+
#Writing Loss and Accuracy
180+
with tf.name_scope('Loss'):
181+
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_true,logits=y_pred))
182+
183+
with tf.name_scope('ADAM'):
184+
train = tf.train.AdamOptimizer(learning_rate=LR).minimize(cross_entropy)
185+
186+
with tf.name_scope('Accuracy'):
187+
matches = tf.equal(tf.argmax(y_pred,1),tf.argmax(y_true,1))
188+
acc = tf.reduce_mean(tf.cast(matches,tf.float32))
189+
190+
init = tf.global_variables_initializer()
191+
192+
tf.summary.scalar("loss", cross_entropy)
193+
tf.summary.scalar("accuracy", acc)
194+
merged_summary_op = tf.summary.merge_all()
195+
196+
acc_list = []
197+
cross_entropy_list = []
198+
acc_train = []
199+
200+
saver = tf.train.Saver()
201+
202+
#If you are using CPU, just use with tf.Session() as sess:
203+
#Starting Session
204+
205+
with tf.Session() as sess:
206+
sess.run(init)
207+
summary_writer = tf.summary.FileWriter(TRAIN_DIR, graph=tf.get_default_graph())
208+
for i in range(epochs):
209+
for j in range(0,steps,step_size):
210+
_ , c , summary,d = sess.run([train,cross_entropy,merged_summary_op,acc],feed_dict={x:X[j:j+step_size] , y_true:Y[j:j+step_size] ,hold_prob1:0.5,hold_prob2:0.5,hold_prob3:0.5,hold_prob4:0.5})
211+
summary_writer.add_summary(summary, i * total_batch + j)
212+
acc_train.append(d)
213+
mean_of_cross_entropy = sess.run(cross_entropy,feed_dict={x:cv_x,y_true:cv_y ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})
214+
mean_of_acc = sess.run(acc,feed_dict={x:cv_x ,y_true:cv_y ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})
215+
cross_entropy_list.append(mean_of_cross_entropy)
216+
acc_list.append(mean_of_acc)
217+
print(i,mean_of_cross_entropy,mean_of_acc)
218+
saver.save(sess, os.getcwd()+"\\CNN_MC.ckpt")
219+
print("test accuracy = ",np.mean([sess.run(acc,feed_dict={x:test_x[:400],y_true:test_y[:400] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(acc,feed_dict={x:test_x[400:800],y_true:test_y[400:800] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(acc,feed_dict={x:test_x[800:],y_true:test_y[800:] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})]))
220+
print("cross_entropy loss = ",np.mean([sess.run(cross_entropy,feed_dict={x:test_x[:400],y_true:test_y[:400] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(cross_entropy,feed_dict={x:test_x[400:800],y_true:test_y[400:800] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0}),sess.run(cross_entropy,feed_dict={x:test_x[800:],y_true:test_y[800:] ,hold_prob1:1.0,hold_prob2:1.0,hold_prob3:1.0,hold_prob4:1.0})]))
221+
222+
import pandas as pd
223+
224+
pd.Series(acc_list).plot(kind='line',title='Accuracy on CV data')
225+
226+
pd.Series(cross_entropy_list).plot(kind='line',title='CV ERROR')
227+
228+
#RESTORING and MAKING PREDICTIONS FOR FIRST 64 IMAGES
229+
with tf.Session() as session:
230+
saver.restore(session, "CNN_MC.ckpt")
231+
print("Model restored.")
232+
print('Initialized')
233+
k = session.run([y_pred], feed_dict={x:test_x[0:64] , hold_prob1:1,hold_prob2:1,hold_prob3:1,hold_prob4:1})
234+
235+
print(np.array(k).shape)
236+
237+
#Reshaping k
238+
k = np.array(k).reshape(64,4)
239+
240+
pred_labels = []
241+
242+
for i in range(64):
243+
r = np.round(np.array(np.exp(k[i])/sum(np.exp(k[i]))),3).argmax()
244+
if r ==0 : pred_labels.append("chair")
245+
elif r ==1: pred_labels.append("kitchen")
246+
elif r ==2: pred_labels.append("knife")
247+
elif r ==3: pred_labels.append("saucepan")
248+
249+
#Multiple images parameters
250+
w=80
251+
h=80
252+
columns = 8
253+
rows = 8
254+
255+
images = test_x[:64]
256+
257+
print(images.shape)
258+
259+
fig = plt.figure(figsize=(20, 20))
260+
for m in range(1, columns*rows +1):
261+
img = images[m-1].reshape([80, 80, 3])
262+
fig.add_subplot(rows, columns, m)
263+
plt.imshow(img)
264+
plt.title("Pred: " + pred_labels[m-1])
265+
plt.axis('off')
266+
plt.show()
267+

0 commit comments

Comments
(0)

AltStyle によって変換されたページ (->オリジナル) /