[DL/CV/PRJ] Image Classification DL 모델 만들기 2
Kaggle 프로젝트의 실전 데이터로 프로젝트 진행해보기
Kaggle: Dogs vs. Cats Redux: Kernels Edition
| 개 고양이 구분 DL 프로젝트 |
Code
import os
import shutil
import tensorflow as tf
# Data Download
os.environ['KAGGLE_CONFIG_DIR'] = '/Users/jinsolkim/.kaggle/'
!kaggle competitions download -c dogs-vs-cats-redux-kernels-edition
!unzip -q dogs-vs-cats-redux-kernels-edition.zip -d .
!unzip -q train.zip -d .
os.mkdir('./dataset')
os.mkdir('./dataset/cat')
os.mkdir('./dataset/dog')
for i in os.listdir('./train/'):
if 'cat' in i:
shutil.copyfile('./train/' + i, './dataset/cat/' + i)
if 'dog' in i:
shutil.copyfile('./train/' + i, './dataset/dog/' + i)
# Pre-process
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
'./dataset/',
image_size=(64,64),
batch_size=64,
subset='training',
validation_split=0.2,
seed=100
)
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
'./dataset/',
image_size=(64,64),
batch_size=64,
subset='validation',
validation_split=0.2,
seed=100
)
def pre_process(i, ans):
i = tf.cast(i/255.0, tf.float32)
return i, ans
train_ds = train_ds.map(pre_process)
val_ds = val_ds.map(pre_process)
# Model
model = tf.keras.Sequential([
tf.keras.layers.Conv2D(32, (3, 3), padding='same', activation='relu', input_shape=(64, 64, 3)),
tf.keras.layers.MaxPooling2D((2, 2)),
tf.keras.layers.Conv2D(64, (3, 3), padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D((2, 2)),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Conv2D(128, (3, 3), padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D((2, 2)),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Dense(1, activation='sigmoid')
])
model.summary()
# Train
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit(train_ds, validation_data=val_ds, epochs=5)
1. Kaggle Data Download
- Kaggle 홈페이지에서
My Account > API - Create New API Token클릭 - 다운로드 된 kaggle.json 파일을
~/.kaggle/에 위치시기키 (Mac 기준)
import os
os.environ['KAGGLE_CONFIG_DIR'] = '/Users/jinsolkim/.kaggle/'
!kaggle competitions download -c dogs-vs-cats-redux-kernels-edition
다운로드 된 zip파일 압축 풀기
!unzip -q dogs-vs-cats-redux-kernels-edition.zip -d .
!unzip -q train.zip -d .
이미지 파일 갯수 확인
print(len(os.listdir('./train/')))
2. Data Preprocessing
💡 Image classification할 때는 dataset 폴더 안에 각 카테고리 별로 폴더 만들어서 해당하는 데이터셋 옮겨놓기!! (항상!⭐️)
1) Dataset
dataset 폴더 만들기
os.mkdir('./dataset')
os.mkdir('./dataset/cat')
os.mkdir('./dataset/dog')
개 vs 고양이 파일 파일 분류해서 dataset 폴더로 복사
import shutil
for i in os.listdir('./train/'): # 파일명 (ex. cat.5077.jpg)
if 'cat' in i:
shutil.copyfile('./train/' + i, './dataset/cat/' + i)
if 'dog' in i:
shutil.copyfile('./train/' + i, './dataset/dog/' + i)
2) Image to Tensor
이미지 숫자로 변환하기
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
'./dataset/',
image_size=(64,64),
batch_size=64,
subset='training',
validation_split=0.2,
seed=100
)
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
'./dataset/',
image_size=(64,64),
batch_size=64,
subset='validation',
validation_split=0.2,
seed=100
)
print(train_ds)
print(val_ds)
------------------------------------------------------------
Found 25000 files belonging to 2 classes.
Using 20000 files for training.
Found 25000 files belonging to 2 classes.
Using 5000 files for validation.
<BatchDataset element_spec=(TensorSpec(shape=(None, 64, 64, 3), dtype=tf.float32, name=None), TensorSpec(shape=(None,), dtype=tf.int32, name=None))>
<BatchDataset element_spec=(TensorSpec(shape=(None, 64, 64, 3), dtype=tf.float32, name=None), TensorSpec(shape=(None,), dtype=tf.int32, name=None))>
tf.keras.preprocessing.image_dataset_from_directory(): 폴더 내 이미지들을 바로 숫자화된 dataset으로 만들어 준다.image_size=(64,64): 모든 이미지를 64x64 pixel로 압축batch_size=64: 이미지 2만장을 모델에 한번에 넣지 않고 batch 숫자만큼 넣도록subset='training': train dataset 이름 붙이기. 아래validation_split에 따라 이 데이터는 전체 데이터의 80%)subset='validation': validation dataset 이름 붙이기. 아래validation_split에 따라 이 데이터는 전체 데이터의 80%)validation_split=0.2- dataset에서 validation dataset split
train_ds블록 밑에val_ds도 똑같은 형태로 한번 더 써줘야 한다.train_ds블록에서는subset='training',val_ds블록에서는subset='validation'써줘서 각각 무슨 데이터를 의미하는지 나타내주기!
train_ds는( (xxx), (yyy) )형태. 즉,( (이미지 2만개), (정답 2만개) ).
3) Image Pre-processing
데이터 0~1 사이의 값으로 scaling
def pre_process(i, ans):
i = tf.cast(i/255.0, tf.float32)
return i, ans
train_ds = train_ds.map(pre_process)
val_ds = val_ds.map(pre_process)
데이터셋 확인
import matplotlib.pyplot as plt
for i, ans in train_ds.take(1):
print(i.shape)
print(i)
print(ans.shape)
print(ans)
plt.imshow(i[0].numpy().astype('uint8'))
plt.show()
------------------------------------------------------------
(64, 64, 64, 3)
tf.Tensor(
[[[[0.23100491 0.21568628 0.20318627]
[0.1858226 0.1740579 0.13876379]
[0.18834636 0.17266008 0.16089538]
...
[0.16947381 0.1069738 0.07952283]
[0.13084023 0.11123239 0.08770297]
[0.13333334 0.11372549 0.09019608]]
[[0.23925781 0.22393918 0.21143919]
[0.1872664 0.17550169 0.14020757]
[0.18532859 0.16964231 0.15787761]
...
[0.16032858 0.10784314 0.07745098]
[0.1420343 0.12242647 0.09889706]
[0.13725491 0.11764706 0.09411765]]
[[0.25 0.22683823 0.23284313]
[0.21454887 0.184892 0.16246553]
[0.23345588 0.18639706 0.1942402 ]
...
[0.1764706 0.14901961 0.11764706]
[0.17058824 0.12745099 0.11960784]
[0.17634805 0.13321078 0.12536764]]
...
(64,)
tf.Tensor(
[0 1 1 0 0 1 0 1 1 0 0 1 1 0 1 1 0 1 0 0 0 1 0 1 1 0 0 1 0 0 1 0 0 0 1 1 1
1 0 0 1 1 1 0 0 1 0 0 0 1 0 0 1 1 0 1 0 1 0 0 1 1 1 0], shape=(64,), dtype=int32)
i = tf.cast( i/255.0, tf.float32 )- (바로
i = i/255.0로 하면 좋겠지만) tensor이므로tf.cast사용 tf.float32: 만약을 위해서 자료형 강제
- (바로
3. Model
model = tf.keras.Sequential([
tf.keras.layers.Conv2D(32, (3, 3), padding='same', activation='relu', input_shape=(64, 64, 3)),
tf.keras.layers.MaxPooling2D((2, 2)),
tf.keras.layers.Conv2D(64, (3, 3), padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D((2, 2)),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Conv2D(128, (3, 3), padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D((2, 2)),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dropout(0.2), # 128개의 노드 중 20% 제거
tf.keras.layers.Dense(1, activation='sigmoid')
])
model.summary()
------------------------------------------------------------
Model: "sequential_2"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_4 (Conv2D) (None, 64, 64, 32) 896
max_pooling2d_4 (MaxPooling (None, 32, 32, 32) 0
2D)
conv2d_5 (Conv2D) (None, 32, 32, 64) 18496
max_pooling2d_5 (MaxPooling (None, 16, 16, 64) 0
2D)
dropout (Dropout) (None, 16, 16, 64) 0
conv2d_6 (Conv2D) (None, 16, 16, 128) 73856
max_pooling2d_6 (MaxPooling (None, 8, 8, 128) 0
2D)
flatten_3 (Flatten) (None, 8192) 0
dense_5 (Dense) (None, 128) 1048704
dropout_1 (Dropout) (None, 128) 0
dense_6 (Dense) (None, 1) 129
=================================================================
Total params: 1,142,081
Trainable params: 1,142,081
Non-trainable params: 0
_________________________________________________________________
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit(train_ds, validation_data=val_ds, epochs=5)
- Dropout 레이어
- overfitting 완화 가능
- 윗 레이어의 노드를 일부 제거해준다.
- ex)
tf.keras.layers.Dropout(0.2): 윗 레이어의 노드 중 20% 제거해주세요.
4. Results
- 데이터 전처리 0~1 scaling 했을 때 vs 안 했을 때 모델 성능 차이
- 데이터 전처리하지 않았을 때
Epoch 1/5 313/313 [==============================] - 10s 26ms/step - loss: 2.5032 - accuracy: 0.5839 - val_loss: 0.6012 - val_accuracy: 0.6806 Epoch 2/5 313/313 [==============================] - 8s 25ms/step - loss: 0.5912 - accuracy: 0.6789 - val_loss: 0.6017 - val_accuracy: 0.6486 Epoch 3/5 313/313 [==============================] - 8s 25ms/step - loss: 0.5473 - accuracy: 0.7165 - val_loss: 0.5435 - val_accuracy: 0.7136 Epoch 4/5 313/313 [==============================] - 8s 25ms/step - loss: 0.5230 - accuracy: 0.7341 - val_loss: 0.5169 - val_accuracy: 0.7546 Epoch 5/5 313/313 [==============================] - 8s 25ms/step - loss: 0.4988 - accuracy: 0.7554 - val_loss: 0.5080 - val_accuracy: 0.7536 - 데이터 전처리했을 때 (0~1 scaling)
Epoch 1/5 313/313 [==============================] - 8s 23ms/step - loss: 0.5454 - accuracy: 0.7229 - val_loss: 0.5101 - val_accuracy: 0.7456 Epoch 2/5 313/313 [==============================] - 8s 24ms/step - loss: 0.4639 - accuracy: 0.7816 - val_loss: 0.4541 - val_accuracy: 0.7910 Epoch 3/5 313/313 [==============================] - 7s 23ms/step - loss: 0.4262 - accuracy: 0.8047 - val_loss: 0.4284 - val_accuracy: 0.7974 Epoch 4/5 313/313 [==============================] - 7s 23ms/step - loss: 0.3850 - accuracy: 0.8260 - val_loss: 0.3896 - val_accuracy: 0.8176 Epoch 5/5 313/313 [==============================] - 7s 23ms/step - loss: 0.3539 - accuracy: 0.8416 - val_loss: 0.3769 - val_accuracy: 0.8208
- 데이터 전처리하지 않았을 때
- Input data의 숫자를 0~1 사이로 압축하면 ⇒ 연산 속도 빨라지고(최적의 w값 빨리 찾는다), accuracy 증가.
- val_acc 약 85%가 이 모델에서의 최대 accuracy.
- 문제는 데이터의 퀄리티.
- 결국, 데이터의 전처리(양, 질)가 모델의 성능을 높이는 데에 큰 효과를 가져온다!
Leave a comment