728x90
2021.6.18
경남 테크노파크 스마트팩토리 인공지능 과정에서 CNN을 학습한 후 예제들을 활용하여 각자 원하는 데이터를 가지고 실습을 진행하였다.
내가 진행한 간단한 프로젝트는 마블 캐릭터 3명을 골라서 분류하는 것이다.
1. 마블 캐릭터 이미지를 kaggle(https://www.kaggle.com/hchen13/marvel-heroes)에서 다운로드 받았다.
2. 데이터가 정리되지 않은 부분이 많아서 데이터들을 걸러내었다.
3. 데이터들을 폴더에 정리하였다.
4. 코드 작성
Marvel Heroes (Using CNN)
In [ ]:
import os
import zipfile
import urllib.request
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import tensorflow as tf
from tensorflow import keras
from keras.models import Sequential
from keras.layers import Flatten, Dense, Conv2D, MaxPooling2D, Dropout
from keras.preprocessing.image import ImageDataGenerator
from keras import optimizers
(1) 데이터 로드
구글 drive mount.
In [ ]:
from google.colab import drive
drive.mount('/content/drive')
Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True).
In [ ]:
pwd
Out[ ]:
'/content/drive'데이터가 있는 디렉토리로 이동.
marvel 디렉토리에 저장된 학습 이미지는 black_widow, spiderman, captain america 각각 200~300개 정도, 전체 745개
In [ ]:
black_widow_dir = os.path.join('/marvel/train/black_widow')
spider_man_dir = os.path.join('/marvel/train/spider-man')
captain_america_dir = os.path.join('/marvel/train/captain_america')
print('total training black widow images:', len(os.listdir(black_widow_dir)))
print('total training spiderman images:', len(os.listdir(spider_man_dir)))
print('total training captain america images:', len(os.listdir(captain_america_dir)))
total training black widow images: 206
total training spiderman images: 300
total training captain america images: 239
In [ ]:
black_widow_files = os.listdir(blackwidow_dir)
print(black_widow_files[:5])
spider_man_files = os.listdir(spiderman_dir)
print(spider_man_files[:5])
captain_america_files = os.listdir(captainamerica_dir)
print(captain_america_files[:5])
['pic_034.jpg', 'pic_028.jpg', 'pic_027.jpg', 'pic_025.jpg', 'pic_048.jpg']
['pic_070.jpg', 'pic_069.jpg', 'pic_066.jpg', 'pic_065.jpg', 'pic_064.jpg']
['pic_005.jpg', 'pic_004.jpg', 'pic_003.jpg', 'pic_002.jpg', 'pic_017.jpg']
디렉토리에 저장된 테스트 이미지는 black widow, spiderman, captain america 전체 109개|
In [ ]:
black_widow_test_dir = os.path.join('/marvel/valid/black_widow')
spider_man_test_dir = os.path.join('/marvel/valid/spider-man')
captain_america_test_dir = os.path.join('/marvel/valid/captain_america')
print('total testing rock images:', len(os.listdir(black_widow_test_dir)))
print('total testing scissors images:', len(os.listdir(spider_man_test_dir)))
print('total testing paper images:', len(os.listdir(captain_america_test_dir)))
total testing rock images: 24
total testing scissors images: 52
total testing paper images: 33
In [ ]:
rps = 3
pic_index = 2
next_black_widow = [os.path.join(black_widow_dir, fname) for fname in black_widow_files[:pic_index]]
next_spider_man = [os.path.join(spider_man_dir, fname) for fname in spider_man_files[:pic_index]]
next_captain_america = [os.path.join(captain_america_dir, fname) for fname in captain_america_files[:pic_index]]
fig = plt.gcf()
fig.set_size_inches(pic_index*4, rps*4)
for i, img_path in enumerate(next_black_widow + next_spider_man + next_captain_america):
sp = plt.subplot(rps, pic_index, i + 1)
sp.axis('Off')
img = mpimg.imread(img_path)
plt.imshow(img)
plt.show()
(2) Data Augmentation
In [ ]:
TRAINING_DIR = "/marvel/train"
training_datagen = ImageDataGenerator(
rescale = 1./255,
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
train_generator = training_datagen.flow_from_directory(
TRAINING_DIR,
target_size=(150,150),
class_mode='categorical',
batch_size=74
)
Found 745 images belonging to 3 classes.
테스트 데이터는 스케일만 변경
In [ ]:
VALIDATION_DIR = "/marvel/valid"
validation_datagen = ImageDataGenerator(rescale = 1./255)
validation_generator = validation_datagen.flow_from_directory(
VALIDATION_DIR,
target_size=(150,150),
class_mode='categorical',
batch_size=36
)
Found 109 images belonging to 3 classes.
(3) model define
In [ ]:
model = Sequential([
Conv2D(64, (3,3), activation='relu', input_shape=(150, 150, 3)),
MaxPooling2D(2, 2),
Conv2D(64, (3,3), activation='relu'),
MaxPooling2D(2,2),
Conv2D(128, (3,3), activation='relu'),
MaxPooling2D(2,2),
Conv2D(128, (3,3), activation='relu'),
MaxPooling2D(2,2),
Conv2D(128, (3,3), activation='relu'),
MaxPooling2D(2,2),
Flatten(),
Dropout(0.5),
Dense(512, activation='relu'),
Dense(3, activation='softmax')
])
In [ ]:
model.summary()
Model: "sequential_5"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_23 (Conv2D) (None, 148, 148, 64) 1792
_________________________________________________________________
max_pooling2d_23 (MaxPooling (None, 74, 74, 64) 0
_________________________________________________________________
conv2d_24 (Conv2D) (None, 72, 72, 64) 36928
_________________________________________________________________
max_pooling2d_24 (MaxPooling (None, 36, 36, 64) 0
_________________________________________________________________
conv2d_25 (Conv2D) (None, 34, 34, 128) 73856
_________________________________________________________________
max_pooling2d_25 (MaxPooling (None, 17, 17, 128) 0
_________________________________________________________________
conv2d_26 (Conv2D) (None, 15, 15, 128) 147584
_________________________________________________________________
max_pooling2d_26 (MaxPooling (None, 7, 7, 128) 0
_________________________________________________________________
conv2d_27 (Conv2D) (None, 5, 5, 128) 147584
_________________________________________________________________
max_pooling2d_27 (MaxPooling (None, 2, 2, 128) 0
_________________________________________________________________
flatten_5 (Flatten) (None, 512) 0
_________________________________________________________________
dropout_5 (Dropout) (None, 512) 0
_________________________________________________________________
dense_10 (Dense) (None, 512) 262656
_________________________________________________________________
dense_11 (Dense) (None, 3) 1539
=================================================================
Total params: 671,939
Trainable params: 671,939
Non-trainable params: 0
_________________________________________________________________
(4) model compile
In [ ]:
model.compile(loss = 'categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
batch * steps_per_epoch = image 개수 train, valid 따로 맞춰줘야함.
(5) 모델 학습
In [ ]:
history = model.fit(
train_generator,
epochs=50,
steps_per_epoch=10,
validation_data = validation_generator,
verbose = 1,
validation_steps=3)
Epoch 1/50
10/10 [==============================] - 17s 1s/step - loss: 1.1146 - accuracy: 0.4042 - val_loss: 1.0595 - val_accuracy: 0.4722
Epoch 2/50
10/10 [==============================] - 15s 1s/step - loss: 1.0506 - accuracy: 0.4192 - val_loss: 0.9280 - val_accuracy: 0.4815
Epoch 3/50
10/10 [==============================] - 15s 1s/step - loss: 0.9742 - accuracy: 0.4872 - val_loss: 1.0504 - val_accuracy: 0.4537
Epoch 4/50
10/10 [==============================] - 14s 1s/step - loss: 1.0113 - accuracy: 0.5226 - val_loss: 0.8303 - val_accuracy: 0.6019
Epoch 5/50
10/10 [==============================] - 15s 2s/step - loss: 1.0152 - accuracy: 0.5777 - val_loss: 0.8096 - val_accuracy: 0.5648
Epoch 6/50
10/10 [==============================] - 15s 2s/step - loss: 0.9105 - accuracy: 0.5596 - val_loss: 0.8162 - val_accuracy: 0.6389
Epoch 7/50
10/10 [==============================] - 15s 1s/step - loss: 0.8699 - accuracy: 0.5997 - val_loss: 0.8693 - val_accuracy: 0.6111
Epoch 8/50
10/10 [==============================] - 15s 1s/step - loss: 0.8810 - accuracy: 0.5631 - val_loss: 0.7487 - val_accuracy: 0.6296
Epoch 9/50
10/10 [==============================] - 15s 2s/step - loss: 0.9104 - accuracy: 0.5920 - val_loss: 0.8055 - val_accuracy: 0.6481
Epoch 10/50
10/10 [==============================] - 15s 1s/step - loss: 0.8244 - accuracy: 0.6545 - val_loss: 0.6916 - val_accuracy: 0.6389
Epoch 11/50
10/10 [==============================] - 14s 1s/step - loss: 0.8204 - accuracy: 0.5736 - val_loss: 0.7244 - val_accuracy: 0.6759
Epoch 12/50
10/10 [==============================] - 15s 1s/step - loss: 0.8002 - accuracy: 0.6042 - val_loss: 0.6994 - val_accuracy: 0.6759
Epoch 13/50
10/10 [==============================] - 15s 1s/step - loss: 0.7963 - accuracy: 0.6109 - val_loss: 0.6925 - val_accuracy: 0.6944
Epoch 14/50
10/10 [==============================] - 15s 2s/step - loss: 0.7776 - accuracy: 0.6256 - val_loss: 0.6906 - val_accuracy: 0.6667
Epoch 15/50
10/10 [==============================] - 15s 1s/step - loss: 0.7601 - accuracy: 0.6347 - val_loss: 0.7725 - val_accuracy: 0.7222
Epoch 16/50
10/10 [==============================] - 15s 1s/step - loss: 0.8110 - accuracy: 0.5967 - val_loss: 0.6884 - val_accuracy: 0.6574
Epoch 17/50
10/10 [==============================] - 15s 1s/step - loss: 0.8265 - accuracy: 0.6259 - val_loss: 0.7566 - val_accuracy: 0.6111
Epoch 18/50
10/10 [==============================] - 15s 2s/step - loss: 0.7835 - accuracy: 0.6104 - val_loss: 0.6915 - val_accuracy: 0.6944
Epoch 19/50
10/10 [==============================] - 15s 1s/step - loss: 0.7489 - accuracy: 0.6186 - val_loss: 0.6400 - val_accuracy: 0.7315
Epoch 20/50
10/10 [==============================] - 15s 1s/step - loss: 0.7460 - accuracy: 0.6138 - val_loss: 0.6744 - val_accuracy: 0.7315
Epoch 21/50
10/10 [==============================] - 14s 1s/step - loss: 0.7281 - accuracy: 0.6571 - val_loss: 0.7921 - val_accuracy: 0.6852
Epoch 22/50
10/10 [==============================] - 15s 1s/step - loss: 0.7703 - accuracy: 0.6223 - val_loss: 0.7639 - val_accuracy: 0.6667
Epoch 23/50
10/10 [==============================] - 15s 2s/step - loss: 0.7861 - accuracy: 0.6272 - val_loss: 0.6806 - val_accuracy: 0.7037
Epoch 24/50
10/10 [==============================] - 15s 1s/step - loss: 0.7457 - accuracy: 0.6237 - val_loss: 0.6695 - val_accuracy: 0.7130
Epoch 25/50
10/10 [==============================] - 15s 1s/step - loss: 0.7641 - accuracy: 0.6205 - val_loss: 0.6924 - val_accuracy: 0.7315
Epoch 26/50
10/10 [==============================] - 15s 2s/step - loss: 0.7406 - accuracy: 0.6191 - val_loss: 0.6900 - val_accuracy: 0.7037
Epoch 27/50
10/10 [==============================] - 14s 1s/step - loss: 0.7118 - accuracy: 0.6426 - val_loss: 0.6520 - val_accuracy: 0.7315
Epoch 28/50
10/10 [==============================] - 15s 1s/step - loss: 0.7618 - accuracy: 0.6468 - val_loss: 0.6821 - val_accuracy: 0.7593
Epoch 29/50
10/10 [==============================] - 14s 1s/step - loss: 0.6980 - accuracy: 0.6587 - val_loss: 0.6199 - val_accuracy: 0.7778
Epoch 30/50
10/10 [==============================] - 15s 1s/step - loss: 0.7198 - accuracy: 0.6700 - val_loss: 0.7007 - val_accuracy: 0.7685
Epoch 31/50
10/10 [==============================] - 15s 1s/step - loss: 0.7100 - accuracy: 0.6959 - val_loss: 0.6786 - val_accuracy: 0.7685
Epoch 32/50
10/10 [==============================] - 15s 1s/step - loss: 0.7414 - accuracy: 0.6561 - val_loss: 0.8073 - val_accuracy: 0.6296
Epoch 33/50
10/10 [==============================] - 15s 2s/step - loss: 0.7585 - accuracy: 0.6529 - val_loss: 0.6342 - val_accuracy: 0.7593
Epoch 34/50
10/10 [==============================] - 14s 1s/step - loss: 0.7112 - accuracy: 0.6687 - val_loss: 0.6843 - val_accuracy: 0.7315
Epoch 35/50
10/10 [==============================] - 15s 1s/step - loss: 0.7205 - accuracy: 0.6610 - val_loss: 0.6242 - val_accuracy: 0.7685
Epoch 36/50
10/10 [==============================] - 15s 1s/step - loss: 0.6494 - accuracy: 0.7021 - val_loss: 0.6796 - val_accuracy: 0.7778
Epoch 37/50
10/10 [==============================] - 15s 1s/step - loss: 0.7280 - accuracy: 0.7147 - val_loss: 0.6326 - val_accuracy: 0.7130
Epoch 38/50
10/10 [==============================] - 15s 2s/step - loss: 0.6793 - accuracy: 0.6589 - val_loss: 0.6249 - val_accuracy: 0.8148
Epoch 39/50
10/10 [==============================] - 15s 2s/step - loss: 0.6699 - accuracy: 0.6635 - val_loss: 0.6244 - val_accuracy: 0.7130
Epoch 40/50
10/10 [==============================] - 15s 1s/step - loss: 0.7451 - accuracy: 0.6519 - val_loss: 0.6154 - val_accuracy: 0.7593
Epoch 41/50
10/10 [==============================] - 15s 2s/step - loss: 0.7205 - accuracy: 0.6430 - val_loss: 0.6957 - val_accuracy: 0.6574
Epoch 42/50
10/10 [==============================] - 15s 1s/step - loss: 0.6871 - accuracy: 0.6466 - val_loss: 0.6462 - val_accuracy: 0.7222
Epoch 43/50
10/10 [==============================] - 15s 2s/step - loss: 0.7044 - accuracy: 0.6156 - val_loss: 0.6402 - val_accuracy: 0.7407
Epoch 44/50
10/10 [==============================] - 15s 1s/step - loss: 0.6943 - accuracy: 0.6821 - val_loss: 0.6227 - val_accuracy: 0.7500
Epoch 45/50
10/10 [==============================] - 15s 2s/step - loss: 0.6670 - accuracy: 0.6832 - val_loss: 0.6030 - val_accuracy: 0.7870
Epoch 46/50
10/10 [==============================] - 15s 1s/step - loss: 0.6636 - accuracy: 0.7129 - val_loss: 0.5931 - val_accuracy: 0.7593
Epoch 47/50
10/10 [==============================] - 15s 2s/step - loss: 0.6279 - accuracy: 0.7117 - val_loss: 0.6102 - val_accuracy: 0.7593
Epoch 48/50
10/10 [==============================] - 15s 2s/step - loss: 0.6526 - accuracy: 0.7133 - val_loss: 0.6281 - val_accuracy: 0.7037
Epoch 49/50
10/10 [==============================] - 15s 1s/step - loss: 0.6887 - accuracy: 0.6929 - val_loss: 0.5941 - val_accuracy: 0.7500
Epoch 50/50
10/10 [==============================] - 15s 1s/step - loss: 0.6171 - accuracy: 0.7128 - val_loss: 0.5912 - val_accuracy: 0.7778
(5) model evaluate (Loss & Accuracy)
In [ ]:
acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(len(acc))
plt.plot(epochs, loss, 'r')
plt.plot(epochs, val_loss, 'b')
plt.title ('Training and validation loss')
Out[ ]:
Text(0.5, 1.0, 'Training and validation loss')
In [ ]:
plt.plot(epochs, acc, 'r', label='Training accuracy')
plt.plot(epochs, val_acc, 'b', label='Validation accuracy')
plt.title('Training and validation accuracy')
plt.legend(loc=0)
plt.figure()
plt.show()
<Figure size 432x288 with 0 Axes>모델 학습 후기
- 데이터의 양이 많지 않았고 데이터가 여러 인물이 섞여있는 경우도 있고 사진마다 캐릭터의 상태가 많이 바뀌는 경우가 많아서 성능이 좋지 않다.
- 더 많은 데이터와 깔끔한 데이터를 사용해야한다.
- 다른 캐릭터들도 추후에 진행해보겠다.
- github 링크
https://github.com/Hm-source/Marvel_heroes_Classification
728x90
반응형