Kaggle Pistachio Image Dataset
참고 | ❤ENetB7 Explained 98% ✅ Fine Tuning ✅ EDA✅
By making a Fine Tuning to a Pre-Trained network
사용 라이브러리
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
| import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import os
import glob
import pathlib
import splitfolders
import PIL
import cv2
import tensorflow as tf
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras import layers
from tensorflow.keras import Sequential
from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint
from tensorflow.keras.utils import plot_model
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, confusion_matrix
from sklearn.metrics import classification_report
|
Introudction
이미 훈련된 네트워크 모델을 이용해 피스타치오 종 분류
Data Load
1
2
3
| data_dir = "Pistachio_Image_Dataset/Pistachio_Image_Dataset/"
data_dir = pathlib.Path(data_dir)
data_dir
|
1
| WindowsPath('Pistachio_Image_Dataset/Pistachio_Image_Dataset')
|
1
2
3
| Total_Img = glob.glob("./Pistachio_Image_Dataset/Pistachio_Image_Dataset/*/*.jpg")
print(f"Total Images: {len(Total_Img)}")
|
2개의 피스타치오 종 (Kirmizi와 Siirt)이 총 2148개 있음
1
2
3
4
5
6
7
| Total_Img = pd.Series(Total_Img)
total_df = pd.DataFrame()
total_df["Filename"] = Total_Img.map(lambda img_name: img_name.split("\\")[-1])
total_df["ClassId"] = Total_Img.map(lambda img_name: img_name.split("\\")[-2])
total_df.head()
|
| Filename | ClassId |
|---|
| 0 | kirmizi (1).jpg | Kirmizi_Pistachio |
|---|
| 1 | kirmizi (10).jpg | Kirmizi_Pistachio |
|---|
| 2 | kirmizi (11).jpg | Kirmizi_Pistachio |
|---|
| 3 | kirmizi (12).jpg | Kirmizi_Pistachio |
|---|
| 4 | kirmizi (13).jpg | Kirmizi_Pistachio |
|---|
1
2
| class_id_distributionTotal = total_df["ClassId"].value_counts()
class_id_distributionTotal
|
1
2
3
| Kirmizi_Pistachio 1232
Siirt_Pistachio 916
Name: ClassId, dtype: int64
|
전체 데이터의 분포 확인
1
| _ = sns.barplot(x=class_id_distributionTotal.values, y=class_id_distributionTotal.index).set_title("Total Distribution of Data")
|
1
2
| plt.figure(facecolor="white")
_ = plt.pie(class_id_distributionTotal.values, labels=class_id_distributionTotal.index, autopct="%.2f%%")
|
데이터 나누기
Train : Test : Val = 80 : 10 : 10
1
| splitfolders.ratio(data_dir, output="output", ratio=(0.8, 0.1, 0.1))
|
1
| Copying files: 2148 files [00:10, 195.75 files/s]
|
splitfolders 라이브러리를 사용하면, 지정한 경로에 분할된 데이터가 폴더로 생성됨
1
2
3
4
5
6
| train_path = "output/train/"
val_path = "output/val/"
test_path = "output/test/"
class_names = os.listdir(train_path)
class_names_val = os.listdir(val_path)
class_names_test = os.listdir(test_path)
|
1
2
3
4
5
6
7
8
9
| train_image = glob.glob('./output/train/*/*.jpg')
test_image = glob.glob('./output/test/*/*.jpg')
Val_image = glob.glob('./output/val/*/*.jpg')
print(f"\
training images: {len(train_image)}\n\
validation images: {len(Val_image)}\n\
test images: {len(test_image)}\
")
|
1
2
3
| training images: 1717
validation images: 214
test images: 217
|
1
2
3
4
5
6
7
| data = [len(train_image), len(Val_image), len(test_image)]
labels = ["Train", "Val", "Test"]
plt.figure(facecolor="white")
_ = plt.pie(data, labels=labels, autopct="%.2f%%")
plt.title("Ratio - [Train : Val : Test]")
plt.show()
|
Train / Validation / Test
1
2
3
4
5
6
7
8
9
10
11
12
13
| # Train
train_image_names = pd.Series(train_image)
train_df = pd.DataFrame({"Filename":[x for x in train_image_names.map(lambda img_name: img_name.split("\\")[-1])],
"ClassId":[x for x in train_image_names.map(lambda img_name: img_name.split("\\")[-2])]})
# Val
val_image_names = pd.Series(Val_image)
val_df = pd.DataFrame({"Filename":[x for x in val_image_names.map(lambda img_name: img_name.split("\\")[-1])],
"ClassId":[x for x in val_image_names.map(lambda img_name: img_name.split("\\")[-2])]})
# Test
test_image_names = pd.Series(test_image)
test_df = pd.DataFrame({"Filename":[x for x in test_image_names.map(lambda img_name: img_name.split("\\")[-1])],
"ClassId":[x for x in test_image_names.map(lambda img_name: img_name.split("\\")[-2])]})
|
1
2
3
4
5
6
| fig, ax = plt.subplots(1, 3, figsize=(15, 5))
for idx, df in enumerate([train_df, val_df, test_df]):
axex = ax[idx]
df = df["ClassId"].value_counts()
axex.pie(df.values, labels=df.index, autopct="%.2f%%")
plt.show()
|
데이터가 균일하게 나눠졌음
데이터 출력하기 (이미지)
| Filename | ClassId |
|---|
| 0 | kirmizi (10).jpg | Kirmizi_Pistachio |
|---|
| 1 | kirmizi (11).jpg | Kirmizi_Pistachio |
|---|
| 2 | kirmizi (12).jpg | Kirmizi_Pistachio |
|---|
| 3 | kirmizi (13).jpg | Kirmizi_Pistachio |
|---|
| 4 | kirmizi (15).jpg | Kirmizi_Pistachio |
|---|
| ... | ... | ... |
|---|
| 1712 | siirt 95.jpg | Siirt_Pistachio |
|---|
| 1713 | siirt 96.jpg | Siirt_Pistachio |
|---|
| 1714 | siirt 97.jpg | Siirt_Pistachio |
|---|
| 1715 | siirt 98.jpg | Siirt_Pistachio |
|---|
| 1716 | siirt 99.jpg | Siirt_Pistachio |
|---|
1717 rows × 2 columns
1
2
3
4
5
6
7
8
9
10
11
| plot_df = train_df.sample(12).reset_index()
plt.figure(figsize=(15, 15))
for _ in range(12):
img_name = plot_df.loc[_, "Filename"]
label_str = plot_df.loc[_, "ClassId"]
plt.subplot(4, 4, _+1)
plt.imshow(plt.imread(os.path.join(train_path, label_str, img_name)))
plt.title(label_str)
plt.xticks([])
plt.yticks([])
|
Image Data Generator
1
2
3
4
5
6
7
| train_datagen = ImageDataGenerator(zoom_range=0.15, width_shift_range=0.2, shear_range=0.15)
test_datagen = ImageDataGenerator()
val_datagen = ImageDataGenerator()
train_generator = train_datagen.flow_from_directory(train_path, target_size=(224, 224), batch_size=32, shuffle=True, class_mode='binary')
test_generator = test_datagen.flow_from_directory(test_path, target_size=(224,224), batch_size=32, shuffle=False, class_mode='binary')
val_generator = val_datagen.flow_from_directory(val_path, target_size=(224,224), batch_size=32, shuffle=False, class_mode='binary')
|
1
2
3
| Found 1717 images belonging to 2 classes.
Found 217 images belonging to 2 classes.
Found 214 images belonging to 2 classes.
|
Network: Efficent Net
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
스케일링을 체계적으로 연구하고 네트워크 깊이와 폭, 해상도 조절로 성능 향상을 이룬 네트워크 모델
기존에 네트워크를 확장하는 일반적인 방법은 깊이를 확장하는 것이었음
예를들어, ResNet은 18->150으로 층을 확장했지만 차원(Demension)은 여전히 1개뿐임
네트워크가 3차원 모두에서 균일하게 분류되어야 한다는 관점을 적용함
1
2
3
4
5
6
7
| from tensorflow.keras.applications import EfficientNetB0
model = EfficientNetB0(
input_shape=(224, 224, 3),
include_top=False,
weights="imagenet"
)
|
1
2
| Downloading data from https://storage.googleapis.com/keras-applications/efficientnetb0_notop.h5
16711680/16705208 [==============================] - 0s 0us/step
|
Fine Tuning
1
2
3
4
| model.trainable = True
for layer in model.layers[:-15]:
layer.trainable = False
|
1
2
3
4
5
| x = tf.keras.layers.Flatten()(model.output)
x = tf.keras.layers.Dropout(0.5)(x)
x = tf.keras.layers.Dense(1, activation="sigmoid")(x)
model = tf.keras.Model(model.input, x)
|
1
2
3
4
5
| model.compile(
optimizer = "adam",
loss = tf.keras.losses.binary_crossentropy,
metrics = ["accuracy"]
)
|
1
2
| # 모델 볼륨이 커서 출력 안함
# plot_model(model, to_file="convnet.png", show_shapes=True, show_layer_names=True)
|
Call Backs
1
2
| early_stopping = EarlyStopping(monitor="val_accuracy", mode="max", verbose=0, patience=20)
model_check = ModelCheckpoint("model.h5", monitor="val_accuracy", mode="max", save_best_only=True)
|
Train the Model
1
2
| with tf.device("/device:GPU:0"):
history = model.fit(train_generator, validation_data=val_generator, epochs=50, verbose=1, callbacks=[early_stopping, model_check])
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
| Epoch 1/50
54/54 [==============================] - 23s 422ms/step - loss: 0.3424 - accuracy: 0.9330 - val_loss: 0.1209 - val_accuracy: 0.9579
Epoch 2/50
54/54 [==============================] - 17s 309ms/step - loss: 0.2074 - accuracy: 0.9464 - val_loss: 0.7111 - val_accuracy: 0.8692
Epoch 3/50
54/54 [==============================] - 17s 321ms/step - loss: 0.1741 - accuracy: 0.9610 - val_loss: 0.4261 - val_accuracy: 0.8972
Epoch 4/50
54/54 [==============================] - 17s 314ms/step - loss: 0.1761 - accuracy: 0.9616 - val_loss: 0.3144 - val_accuracy: 0.9393
...
Epoch 43/50
54/54 [==============================] - 17s 313ms/step - loss: 0.0324 - accuracy: 0.9936 - val_loss: 0.2229 - val_accuracy: 0.9533
Epoch 44/50
54/54 [==============================] - 17s 311ms/step - loss: 0.0261 - accuracy: 0.9942 - val_loss: 0.3668 - val_accuracy: 0.9626
Epoch 45/50
54/54 [==============================] - 17s 307ms/step - loss: 0.1185 - accuracy: 0.9825 - val_loss: 0.0720 - val_accuracy: 0.9907
|
1
2
3
4
5
6
| hist_df = pd.DataFrame(history.history)
fig, ax = plt.subplots(1, 2, figsize=(14, 5))
hist_df[["accuracy", "val_accuracy"]].plot(ax=ax[0])
hist_df[["loss", "val_loss"]].plot(ax=ax[1])
plt.show()
|
Model Evaluation
1
2
3
| test_loss, test_acc = model.evaluate(test_generator, steps=len(test_generator), verbose=1)
print('Loss: %.3f' % (test_loss * 100.0))
print('Accuracy: %.3f' % (test_acc * 100.0))
|
1
2
3
| 7/7 [==============================] - 2s 268ms/step - loss: 0.3866 - accuracy: 0.9631
Loss: 38.660
Accuracy: 96.313
|
Classification Report
1
2
3
4
5
| y_val = test_generator.classes
y_pred = model.predict(test_generator)
y_pred = np.argmax(y_pred,axis=1)
print(classification_report(y_val, y_pred))
|
1
2
3
4
5
6
7
8
| precision recall f1-score support
0 0.57 1.00 0.73 124
1 0.00 0.00 0.00 93
accuracy 0.57 217
macro avg 0.29 0.50 0.36 217
weighted avg 0.33 0.57 0.42 217
|
Comments powered by Disqus.