Deep Learning (06) - 신경망 / 다중 레이블 분류
신경망 적용해보기
무슨옷과 무슨색?
In [1]:
!pwd
Out [1]:
/content
In [2]:
!mkdir clothes_dataset
In [3]:
!unzip '/content/drive/MyDrive/Colab Notebooks/deep_learning/clothes_dataset.zip' -d ./clothes_dataset/
Out [3]:
inflating: ./clothes_dataset/brown_shoes/30ef20bcb027c99409c81fd6127957502b0e693e.jpg
inflating: ./clothes_dataset/brown_shoes/312cf581fd4ec3678b8794f9f488aa1dad2f2908.jpg
...
inflating: ./clothes_dataset/white_shorts/fb625925bc55cd5c0a147d8214b63d806ac7ecb1.jpg
inflating: ./clothes_dataset/white_shorts/fd76a8fe7e6c5d2fe5f346c3207f75325a781d82.jpg
데이터 정리하기
In [4]:
import numpy as np
import pandas as pd
import tensorflow as tf
import glob as glob
import cv2
all_data = np.array(glob.glob('/content/clothes_dataset/*/*.jpg', recursive=True))
# 색과 옷의 종류를 구별하기 위해 해당되는 label에 1을 삽입합니다.
def check_cc(color, clothes):
labels = np.zeros(11,)
# color check
if(color == 'black'):
labels[0] = 1
color_index = 0
elif(color == 'blue'):
labels[1] = 1
color_index = 1
elif(color == 'brown'):
labels[2] = 1
color_index = 2
elif(color == 'green'):
labels[3] = 1
color_index = 3
elif(color == 'red'):
labels[4] = 1
color_index = 4
elif(color == 'white'):
labels[5] = 1
color_index = 5
# clothes check
if(clothes == 'dress'):
labels[6] = 1
elif(clothes == 'shirt'):
labels[7] = 1
elif(clothes == 'pants'):
labels[8] = 1
elif(clothes == 'shorts'):
labels[9] = 1
elif(clothes == 'shoes'):
labels[10] = 1
return labels, color_index
# label과 color_label을 담을 배열을 선언합니다.
all_labels = np.empty((all_data.shape[0], 11))
all_color_labels = np.empty((all_data.shape[0], 1))
for i, data in enumerate(all_data):
color_and_clothes = all_data[i].split('/')[-2].split('_')
color = color_and_clothes[0]
clothes = color_and_clothes[1]
labels, color_index = check_cc(color, clothes)
all_labels[i] = labels
all_color_labels[i] = color_index
all_labels = np.concatenate((all_labels, all_color_labels), axis = -1)
In [5]:
from sklearn.model_selection import train_test_split
# 훈련, 검증, 테스트 데이터셋으로 나눕니다.
train_x, test_x, train_y, test_y = train_test_split(all_data, all_labels, shuffle = True, test_size = 0.3,
random_state = 99)
train_x, val_x, train_y, val_y = train_test_split(train_x, train_y, shuffle = True, test_size = 0.3,
random_state = 99)
In [6]:
train_df = pd.DataFrame({'image':train_x, 'black':train_y[:, 0], 'blue':train_y[:, 1],
'brown':train_y[:, 2], 'green':train_y[:, 3], 'red':train_y[:, 4],
'white':train_y[:, 5], 'dress':train_y[:, 6], 'shirt':train_y[:, 7],
'pants':train_y[:, 8], 'shorts':train_y[:, 9], 'shoes':train_y[:, 10],
'color':train_y[:, 11]})
val_df = pd.DataFrame({'image':val_x, 'black':val_y[:, 0], 'blue':val_y[:, 1],
'brown':val_y[:, 2], 'green':val_y[:, 3], 'red':val_y[:, 4],
'white':val_y[:, 5], 'dress':val_y[:, 6], 'shirt':val_y[:, 7],
'pants':val_y[:, 8], 'shorts':val_y[:, 9], 'shoes':val_y[:, 10],
'color':val_y[:, 11]})
test_df = pd.DataFrame({'image':test_x, 'black':test_y[:, 0], 'blue':test_y[:, 1],
'brown':test_y[:, 2], 'green':test_y[:, 3], 'red':test_y[:, 4],
'white':test_y[:, 5], 'dress':test_y[:, 6], 'shirt':test_y[:, 7],
'pants':test_y[:, 8], 'shorts':test_y[:, 9], 'shoes':test_y[:, 10],
'color':test_y[:, 11]})
In [7]:
train_df.head()
Out [7]:
| image | black | blue | brown | green | red | white | dress | shirt | pants | shorts | shoes | color | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | /content/clothes_dataset/green_shorts/e74d11d3... | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 | 3.0 |
| 1 | /content/clothes_dataset/black_dress/f1be32393... | 1.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |
| 2 | /content/clothes_dataset/black_shoes/04f78f68a... | 1.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
| 3 | /content/clothes_dataset/brown_pants/0671d132b... | 0.0 | 0.0 | 1.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 | 0.0 | 2.0 |
| 4 | /content/clothes_dataset/white_shoes/59803fb01... | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 5.0 |
In [8]:
!pwd
Out [8]:
/content
In [9]:
!mkdir csv_data
In [10]:
train_df.to_csv('/content/csv_data/train.csv', index=False)
val_df.to_csv('/content/csv_data/val.csv', index=False)
test_df.to_csv('/content/csv_data/test.csv', index=False)
이미지 제네레이터 정의 및 모델 구성하기
In [11]:
from keras.preprocessing.image import ImageDataGenerator
In [12]:
train_datagen = ImageDataGenerator(rescale=1./255) # 불러온 값을 rescale로 지정한 값과 곱함->범위 축소
val_datagen = ImageDataGenerator(rescale=1./255)
# 스탭 계산하는 함수
def get_steps(num_samples, batch_size):
if (num_samples % batch_size) > 0:
return (num_samples // batch_size) + 1
else:
return num_samples // batch_size
- 모델 구성
In [13]:
from keras.models import Sequential
from keras.layers import Dense, Flatten
In [14]:
model = Sequential()
model.add(Flatten(input_shape=(112, 112, 3))) # RGB라 3차원
model.add(Dense(128, activation='relu'))
model.add(Dense(64, activation='relu'))
model.add(Dense(11, activation='sigmoid')) # softmax는 가능성 높은것 중 하나, sigmoid는 모든 값을 0과 1사이값으로
model.compile(optimizer='adam',
loss='binary_crossentropy', # 손실함수: 바이너리 크로스엔트로피
metrics=['binary_accuracy'])
In [15]:
model.summary()
Out [15]:
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten (Flatten) (None, 37632) 0
dense (Dense) (None, 128) 4817024
dense_1 (Dense) (None, 64) 8256
dense_2 (Dense) (None, 11) 715
=================================================================
Total params: 4,825,995
Trainable params: 4,825,995
Non-trainable params: 0
_________________________________________________________________
데이터 제네레이터 정의하기
In [16]:
train_df.columns
Out [16]:
Index(['image', 'black', 'blue', 'brown', 'green', 'red', 'white', 'dress',
'shirt', 'pants', 'shorts', 'shoes', 'color'],
dtype='object')
In [17]:
batch_size = 32
class_col = ['black', 'blue', 'brown', 'green', 'red', 'white',
'dress', 'shirt', 'pants', 'shorts', 'shoes']
In [18]:
# 제네레이터 만들기
train_generator = train_datagen.flow_from_dataframe(
dataframe=train_df, x_col='image', y_col=class_col,
target_size=(112, 112), color_mode='rgb', class_mode='raw',
batch_size=batch_size, shuffle=True, seed=42)
val_generator = val_datagen.flow_from_dataframe(
dataframe=val_df, x_col='image', y_col=class_col,
target_size=(112, 112), color_mode='rgb', class_mode='raw',
batch_size=batch_size, shuffle=True)
Out [18]:
Found 5578 validated image filenames.
Found 2391 validated image filenames.
제네레이터를 통해 모델 학습시키기
In [19]:
model.fit(train_generator,
steps_per_epoch=get_steps(len(train_df), batch_size),
validation_data=val_generator,
validation_steps=get_steps(len(val_df), batch_size),
epochs=10)
Out [19]:
Epoch 1/10
175/175 [==============================] - 31s 158ms/step - loss: 0.6019 - binary_accuracy: 0.8315 - val_loss: 0.3970 - val_binary_accuracy: 0.8343
Epoch 2/10
175/175 [==============================] - 27s 155ms/step - loss: 0.3113 - binary_accuracy: 0.8790 - val_loss: 0.2709 - val_binary_accuracy: 0.8987
Epoch 3/10
175/175 [==============================] - 27s 155ms/step - loss: 0.2832 - binary_accuracy: 0.8913 - val_loss: 0.2794 - val_binary_accuracy: 0.8905
Epoch 4/10
175/175 [==============================] - 29s 164ms/step - loss: 0.2347 - binary_accuracy: 0.9062 - val_loss: 0.2622 - val_binary_accuracy: 0.9001
Epoch 5/10
175/175 [==============================] - 27s 152ms/step - loss: 0.2218 - binary_accuracy: 0.9120 - val_loss: 0.2417 - val_binary_accuracy: 0.9004
Epoch 6/10
175/175 [==============================] - 27s 154ms/step - loss: 0.2011 - binary_accuracy: 0.9193 - val_loss: 0.2290 - val_binary_accuracy: 0.9156
Epoch 7/10
175/175 [==============================] - 27s 153ms/step - loss: 0.1905 - binary_accuracy: 0.9244 - val_loss: 0.2051 - val_binary_accuracy: 0.9205
Epoch 8/10
175/175 [==============================] - 28s 159ms/step - loss: 0.1802 - binary_accuracy: 0.9290 - val_loss: 0.2301 - val_binary_accuracy: 0.9138
Epoch 9/10
175/175 [==============================] - 32s 181ms/step - loss: 0.1873 - binary_accuracy: 0.9262 - val_loss: 0.1955 - val_binary_accuracy: 0.9252
Epoch 10/10
175/175 [==============================] - 27s 154ms/step - loss: 0.1659 - binary_accuracy: 0.9346 - val_loss: 0.1960 - val_binary_accuracy: 0.9233
<keras.callbacks.History at 0x7f1867fb1790>
테스트 데이터 예측하기
In [20]:
test_datagen = ImageDataGenerator(rescale=1./255)
test_generator = test_datagen.flow_from_dataframe(
dataframe=test_df, x_col='image',
target_size=(112, 112), color_mode='rgb', class_mode=None,
batch_size=batch_size, shuffle=False)
preds = model.predict(test_generator, steps=get_steps(len(test_df), batch_size), verbose=1)
Out [20]:
Found 3416 validated image filenames.
107/107 [==============================] - 12s 110ms/step
In [21]:
np.round(preds[0], 2)
Out [21]:
array([0. , 0. , 0. , 1. , 0. , 0. , 0. , 1. , 0. , 0.17, 0. ],
dtype=float32)
In [22]:
import matplotlib.pyplot as plt
In [23]:
do_preds = preds[:8]
for i, pred in enumerate(do_preds):
plt.subplot(2, 4, i+1)
prob = zip(class_col, list(pred))
prob = sorted(list(prob), key=lambda x: x[1], reverse=True)[:2] # 내림차순 정렬
image = cv2.imread(test_df['image'][i]) # 경로 지정 읽어오기
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # BGR 패턴을 RGB 패턴으로 변경
plt.imshow(image)
plt.title(f'{prob[0][0]}: {round(prob[0][1]*100, 2)}% \n {prob[1][0]}:{round(prob[1][1]*100, 2)}%')
plt.tight_layout()
plt.show()
Out [23]:
실제 데이터 가져와 예측하기
In [24]:
data_datagen = ImageDataGenerator(rescale=1./255)
data_glob = sorted(glob.glob('/content/drive/MyDrive/Colab Notebooks/deep_learning/06_clothes_img/*/*.png'))
data_generator = data_datagen.flow_from_directory(
directory='/content/drive/MyDrive/Colab Notebooks/deep_learning/06_clothes_img',
target_size=(112, 112), color_mode='rgb', class_mode=None,
batch_size=batch_size, shuffle=False)
Out [24]:
Found 3 images belonging to 1 classes.
In [25]:
result = model.predict(data_generator, steps=get_steps(3, batch_size), verbose=1)
Out [25]:
1/1 [==============================] - 0s 35ms/step
In [26]:
np.round(result, 2)
Out [26]:
array([[0. , 0. , 0. , 0. , 1. , 0. , 0.01, 0. , 0.22, 0. , 0.64],
[0. , 1. , 0. , 0. , 0. , 0. , 0.01, 0. , 0.05, 0.04, 0.65],
[0.02, 0. , 1. , 0.02, 0.14, 0. , 0. , 0. , 0.92, 0.08, 0.45]],
dtype=float32)
In [27]:
for i, pred in enumerate(result):
plt.subplot(1, 3, i+1)
prob = zip(class_col, list(pred))
prob = sorted(list(prob), key=lambda x: x[1], reverse=True)[:2] # 내림차순 정렬
image = cv2.imread(data_glob[i]) # 경로 지정 읽어오기
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # BGR 패턴을 RGB 패턴으로 변경
plt.imshow(image)
plt.title(f'{prob[0][0]}: {round(prob[0][1]*100, 2)}% \n {prob[1][0]}:{round(prob[1][1]*100, 2)}%')
plt.tight_layout()
plt.show()
Out [27]:
Reference
- 이 포스트는 SeSAC 인공지능 자연어처리, 컴퓨터비전 기술을 활용한 응용 SW 개발자 양성 과정 - 심선조 강사님의 강의를 정리한 내용입니다.
댓글남기기