과제 목표
CNN 모델 만들어 5가지 꽃 품종을 분류하고자 한다.
과제 내용
CNN 이론을 이해하고 Tensorflow로 구현하고 CNN의 동작과정을 이해하고자 한다.
주어진 데이터셋을 가지고 5가지 꽃의 품종을 분류하는 CNN모델을 학습시키고자 한다.
사용 라이브러리
tensorflow1.15, tqdm, numpy, matplotlib, tensorflow-datasets 1.2.0
데이터셋 구성
학습시간을 위해 64X64 크기로 사전에 전처리 하였다.
학습데이터: 꽃 이미지 2220장
검증데이터: 꽃 이미지 730장
테스트데이터: 꽃 이미지 720장
기본 코드
데이터 전처리 및 배치사이즈 함수 정의
from tqdm.notebook import tqdm
import tensorflow as tf
import tensorflow_datasets as tfds
import numpy as np
import matplotlib.pyplot as plt
import cv2
import os
import time
tf.reset_default_graph()
tf.compat.v1.disable_eager_execution()
def data():
train_dataset, validation_dataset, test_dataset=tfds.load(name='tf_flowers', split=[tfds.Split.TRAIN.subsplit(tfds.percent[:60]),
tfds.Split.TRAIN.subsplit(tfds.percent[60:80]),
tfds.Split.TRAIN.subsplit(tfds.percent[80:])
],shuffle_files=False)
return train_dataset, validation_dataset, test_datasetdef preprocess():
train_dataset, validation_dataset, _ =data()
#-----------------------Train Data 전처리
train_image_array=[]
train_label_array=[]
train_size=0
for train_data in tfds.as_numpy(train_dataset):
train_image_data=train_data['image']
train_image_array.append(train_image_data)
train_label_data=train_data['label']
train_label_array.append(train_label_data)
train_size+=1
for i in range(train_size):
train_image_array[i]=cv2.resize(train_image_array[i], (64, 64))/255
validation_image_array=[]
validation_label_array=[]
validation_size=0
for validation_data in tfds.as_numpy(validation_dataset):
validation_image_data=validation_data['image']
validation_image_array.append(validation_image_data)
validation_label_data=validation_data['label']
validation_label_array.append(validation_label_data)
validation_size+=1
for i in range(validation_size):
validation_image_array[i]=cv2.resize(validation_image_array[i], (64, 64))/255
return train_image_array, train_label_array, validation_image_array, validation_label_array
def next_batch(num, data, labels):
idx=np.arange(0, len(data))
np.random.shuffle(idx)
idx=idx[:num]
data_shuffle=[data[i] for i in idx]
labels_shuffle=[labels[i] for i in idx]
return np.asarray(data_shuffle), np.asarray(labels_shuffle)시각화 함수 정의
def visualization():
print("\n=================================")
print(" Student Info")
print("=================================")
print("이름 : ", student_name)
print("\n=================================")
print(" Dataset Info")
print("=================================")
print("- 총 3670개의 꽃 이미지 학습 데이터로 구성")
fig=plt.figure(figsize=(10, 4))
for i in range(10):
ax=fig.add_subplot(2, 5, i+1)
image_show=batch[0]
label_show=batch[1]
plt.imshow(image_show[i])
ax.set_xlabel(label_show[i])
ax.axes.xaxis.set_ticks([])
ax.axes.yaxis.set_ticks([])
if i == 2:
ax.set_title("Dataset example", fontsize=18)
plt.show()
print("0:Dandelion 1:Daisy 2:Tulips 3:Sunflowers 4:Roses")
print("\n=================================")
print(" Train Info")
print("=================================")
print("학습시간 : {}초".format(round(train_time, 3)))
print("학습 epoch : ", train_epoch)
print("dropout rate : ", dropout_rate)
print("\n=================================")
print(" Train Loss")
print("=================================")
for i in range(len(plt_index)-1):
print("epoch : {}\tloss : {}".format(plt_index[i+1], round(plt_train_loss[i+1], 5)))
plt.title("Train loss", fontsize=14)
plt.xlabel("epoch")
plt.ylabel("loss")
if train_epoch >= 10:
plt.xticks([i for i in range(0, plt_index[-1]+1, int(plt_index[-1]/5))])
plt.plot(plt_index, plt_train_loss, '-')
plt.show()
print("\n=================================")
print(" Train Accuracy")
print("=================================")
for i in range(len(plt_index)-1):
print("epoch : {}\taccuracy : {}%".format(plt_index[i+1], round(plt_train_accuracy[i+1]*100, 5)))
plt.title("Train Accuracy", fontsize=14)
plt.xlabel("epoch")
plt.ylabel("accuracy")
plt.plot(plt_index, plt_train_accuracy, "g")
if train_epoch >= 10:
plt.xticks([i for i in range(0, plt_index[-1]+1, int(plt_index[-1]/5))])
plt.show()
print("\n=================================")
print(" Validation")
print("=================================")
print("Validation Loss : {}".format(round(validation_loss, 5)))
print("Validation Accuracy : {}%".format(round(validation_accuracy*100, 5)))
plt.suptitle('Validation Result',fontsize=14)
plt.subplot(121)
plt.bar([0, 1, 2], [0, validation_loss, 0], color="orange")
plt.xticks([1],["Validation Loss"])
plt.subplot(122)
plt.bar([0, 1, 2], [0, validation_accuracy, 0], color="blue")
plt.xticks([1],["Validation Accuracy"])
plt.show()
CNN 모델링
student_name="이름" #자신의 학번과 이름으로 수정
dropout_rate = 0.5 #드롭아웃 비율 설정
train_epoch=1000 #에폭 설정
def build_CNN_classifier(x):
# 입력 이미지
x_image=x
# 첫번째 convolutional layer
W_conv1=tf.Variable(tf.truncated_normal(shape=[5, 5, 3, 64], stddev=5e-2))
b_conv1=tf.Variable(tf.constant(0.1, shape=[64]))
h_conv1=tf.nn.relu(tf.nn.conv2d(x_image, W_conv1, strides=[1, 1, 1, 1], padding='SAME')+b_conv1)
# 첫번째 Pooling layer
h_pool1=tf.nn.max_pool(h_conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')
# 두번째 convolutional layer
W_conv2=tf.Variable(tf.truncated_normal(shape=[5, 5, 64, 64], stddev=5e-2))
b_conv2=tf.Variable(tf.constant(0.1, shape=[64]))
h_conv2=tf.nn.relu(tf.nn.conv2d(h_pool1, W_conv2, strides=[1, 1, 1, 1], padding='SAME')+b_conv2)
# 두번째 pooling layer
h_pool2=tf.nn.max_pool(h_conv2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')
# 세번째 convolutional layer
W_conv3=tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 128], stddev=5e-2))
b_conv3=tf.Variable(tf.constant(0.1, shape=[128]))
h_conv3=tf.nn.relu(tf.nn.conv2d(h_pool2, W_conv3, strides=[1, 1, 1, 1], padding='SAME')+b_conv3)
# 네번째 convolutional layer
W_conv4=tf.Variable(tf.truncated_normal(shape=[3, 3, 128, 128], stddev=5e-2))
b_conv4=tf.Variable(tf.constant(0.1, shape=[128]))
h_conv4=tf.nn.relu(tf.nn.conv2d(h_conv3, W_conv4, strides=[1, 1, 1, 1], padding='SAME')+b_conv4)
# 다섯번째 convolutional layer
W_conv5=tf.Variable(tf.truncated_normal(shape=[3, 3, 128, 128], stddev=5e-2))
b_conv5=tf.Variable(tf.constant(0.1, shape=[128]))
h_conv5=tf.nn.relu(tf.nn.conv2d(h_conv4, W_conv5, strides=[1, 1, 1, 1], padding='SAME')+b_conv5)
# 첫번째 Fully Connected Layer
W_fc1=tf.Variable(tf.truncated_normal(shape=[16*16*128, 512], stddev=5e-2))
b_fc1=tf.Variable(tf.constant(0.1, shape=[512]))
h_conv5_flat=tf.reshape(h_conv5, [-1, 16*16*128])
h_fc1=tf.nn.relu(tf.matmul(h_conv5_flat, W_fc1)+b_fc1)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# 두번째 Fully Connected Layer
W_fc2=tf.Variable(tf.truncated_normal(shape=[512, 5], stddev=5e-2))
b_fc2=tf.Variable(tf.constant(0.1, shape=[5]))
logits=tf.matmul(h_fc1_drop, W_fc2)+b_fc2
y_pred=tf.nn.softmax(logits)
return y_pred, logits
x=tf.placeholder(tf.float32, shape=[None, 64, 64, 3])
y=tf.placeholder(tf.int64, shape=[None])
keep_prob=tf.placeholder(tf.float32)
#-----------------------Convolutional Neural Networks(CNN) 그래프 생성
y_pred, logits=build_CNN_classifier(x)
#-----------------------Cross Entropy를 비용함수(loss function)으로 정의하고, RMSPropOptimizer 사용
loss=tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y))
train_step=tf.train.RMSPropOptimizer(0.001).minimize(loss)
#-----------------------정확도 계산
correct_prediction=tf.equal(tf.argmax(y_pred, 1), y)
accuracy=tf.reduce_mean(tf.cast(correct_prediction, tf.float32))세션 실행, 체크포인트
SAVER_DIR="model"
saver=tf.train.Saver()
checkpoint_path=os.path.join(SAVER_DIR, "model")
ckpt=tf.train.get_checkpoint_state(SAVER_DIR)
#-----------------------세션을 열고 학습 진행
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train_image_array, train_label_array, validation_image_array, validation_label_array=preprocess()
plt_train_accuracy=[0]
plt_train_loss=[0]
plt_index=[0]
start=time.time()
print("\n=================================")
print(" 학습 진행 상황")
print("=================================")
for i in tqdm(range(train_epoch)):
batch=next_batch(32, train_image_array, train_label_array)
if (i+1)%100 == 0:
saver.save(sess, checkpoint_path, global_step=i+1)
train_accuracy=accuracy.eval(feed_dict={x: batch[0], y: batch[1], keep_prob: 1.0})
train_loss=loss.eval(feed_dict={x: batch[0], y: batch[1], keep_prob: 1.0})
plt_train_accuracy.append(train_accuracy)
plt_train_loss.append(train_loss)
plt_index.append(i+1)
sess.run(train_step, feed_dict={x: batch[0], y: batch[1], keep_prob: dropout_rate})
train_time=time.time()-start
validation_batch=next_batch(730, validation_image_array, validation_label_array)
validation_loss=sess.run(loss, feed_dict={x: validation_batch[0], y: validation_batch[1], keep_prob: 1.0})
validation_accuracy=sess.run(accuracy, feed_dict={x: validation_batch[0], y: validation_batch[1], keep_prob: 1.0})
visualization()
기본코드 결과
-
테스트 정확도(Accuracy)
epoch : 100 accuracy : 31.25%
epoch : 200 accuracy : 25.0%
epoch : 300 accuracy : 37.5%
epoch : 400 accuracy : 50.0%
epoch : 500 accuracy : 75.0%
epoch : 600 accuracy : 65.625%
epoch : 700 accuracy : 71.875%
epoch : 800 accuracy : 87.5%
epoch : 900 accuracy : 90.625%
epoch : 1000 accuracy : 78.125% -
테스트 손실(Loss)
epoch : 100 loss : 1.614650011062622
epoch : 200 loss : 1.5362000465393066
epoch : 300 loss : 1.4424500465393066
epoch : 400 loss : 1.0443700551986694
epoch : 500 loss : 0.7260900139808655
epoch : 600 loss : 0.7748000025749207
epoch : 700 loss : 0.7546300292015076
epoch : 800 loss : 0.35062000155448914
epoch : 900 loss : 0.34352999925613403
epoch : 1000 loss : 0.7358199954032898
검증 정확도(Validation Accuracy)
Validation Accuracy : 58.90411%
검증 손실(Validation Loss)
Validation Loss : 1.772510051727295
AI 개발자를 목표로 하고 있는 꿈 많은 공대생입니다. a deo vocatus rite paratus