[TensorFlow][Keras][ResNet50] ResNet50, Transfer Learning, Tutorial

1. Transfer Learning

2. ResNet50 (Keras)

3. Dataset

4. Tutorial Code

 

---

1. Transfer Learning

 전이 학습은 기존에 핟습된 모델을 다른 작업에 재사용하는 기법이며 기존 모델이 학습한 특징을 활용하여 새로운 작업에 대한 학습을 빠르고 효율적으로 수행할 수 있음

 

장점

• 학습 시간 단축: 기존 모델의 특징을 활용하여 학습을 시작할 수 있으므로, 새로운 작업에 대한 학습 시간을 단축할 수 있다
• 데이터 효율성 향상: 기존 모델이 학습한 데이터를 활용하여 학습을 수행할 수 있으므로, 새로운 작업에 대한 데이터 수집 및 전처리 비용을 절감할 수 있다.
• 성능 향상: 기존 모델의 특징을 활용하여 새로운 작업에 대한 성능을 향상시킬 수 있다.

단점

• 데이터의 유사성: 기존 모델이 학습한 데이터와 새로운 작업에 사용되는 데이터가 유사할수록 전이 학습의 효과가 높다.
• 모델의 종류: 기존 모델의 종류에 따라 전이 학습의 효과가 달라질 수 있다.

ref: https://www.superannotate.com/blog/speed-up-labeling-process-using-transfer-learning

 

---

2. ResNet50 (Keras)

ResNet

• Resnet은 깊은 신경망의 학습 문제를 해결하기 위해 잔차 연결을 사용하는 신경망 구조임.
• 잔차 연결은 이전 층의 출력을 다음 층의 입력으로 직접 연결하여 깊은 신경망의 학습 속도와 성능을 향상시킴.

ResNet50 (Keras)

• Keras ResNet50은 Keras에서 제공하는 Resnet50 모델임.
• Keras ResNet50은 50개의 잔차 블록으로 구성되어 있으며, 이미지 분류, 객체 감지, 객체 인식 등 다양한 분야에서 활용될 수 있음.

ref: https://medium.com/@arashserej/resnet-50-83b3ff33be7d

---

3. Dataset (kaggle)

- Flower Classification

https://www.kaggle.com/datasets/sauravagarwal/flower-classification?resource=download

Flower Classification

- Flower types - daisy, dandelion, roses, sunflowers, tulips

---

4. Tutorial Code

1) library

# python librairies installation
!pip install split-folders matplotlib opencv-python spicy

# display, transform, read, split ...
import numpy as np
import cv2 as cv
import os
import splitfolders
import matplotlib.pyplot as plt

# tensorflow
import tensorflow.keras as keras
import tensorflow as tf

# image processing
from tensorflow.keras.preprocessing import image
from tensorflow.keras.preprocessing.image import ImageDataGenerator, load_img

# model / neural network
from tensorflow.keras import layers
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.applications import ResNet50
from tensorflow.keras.applications.resnet50 import preprocess_input

---

2) Data Preprocessing

- check the images each type

# daisy
img_daisy = image.load_img("C:/Users/USER/Desktop/resnet50/data_full/daisy/100080576_f52e8ee070_n.jpg")
img_daisy

 

# dandelion
img_dandelion = image.load_img("C:/Users/USER/Desktop/resnet50/data_full/dandelion/10043234166_e6dd915111_n.jpg")
img_dandelion

# roses
img_roses = image.load_img("C:/Users/USER/Desktop/resnet50/data_full/roses/10090824183_d02c613f10_m.jpg")
img_roses

# sunflowers
img_sunflowers = image.load_img("C:/Users/USER/Desktop/resnet50/data_full/sunflowers/1008566138_6927679c8a.jpg")
img_sunflowers

 

# tulips
img_tulips = image.load_img("C:/Users/USER/Desktop/resnet50/data_full/tulips/100930342_92e8746431_n.jpg")
img_tulips

 

---

- spilt dataset

# split data in a new folder named data-split
splitfolders.ratio("C:/Users/USER/Desktop/resnet50/data_full", 
					output="C:/Users/USER/Desktop/resnet50/data_split", 
					seed=1337, ratio=(0.8, 0.1, 0.1), group_prefix=None, move=False)

 

---

- Create Keras data generators

# ImageDataGenerator()

datagen = ImageDataGenerator()

 

# Classes

# define classes name
class_names = ['daisy','dandelion','roses','sunflowers','tulips']

 

# Train set

# training data
train_generator = datagen.flow_from_directory( 
    directory="C:/Users/USER/Desktop/resnet50/data_split/train/", 
    classes = class_names,
    target_size=(224, 224),  
    batch_size=32, 
    class_mode="binary", 
)

 

# Validation set

# validation data
valid_generator = datagen.flow_from_directory( 
    directory="C:/Users/USER/Desktop/resnet50/data_split/val/", 
    classes = class_names,
    target_size=(224, 224), 
    batch_size=32, 
    class_mode="binary", 
)

 

# Test set

# test data
test_generator = datagen.flow_from_directory( 
    directory="C:/Users/USER/Desktop/resnet50/data_split/test/", 
    classes = class_names,
    target_size=(224, 224), 
    batch_size=32, 
    class_mode="binary", 
)

---

3) Build the morel, ResNet50

# ResNet50 model
resnet_50 = ResNet50(include_top=False, weights='imagenet', input_shape=(224,224,3))
for layer in resnet_50.layers:
    layer.trainable = False

# build the entire model
x = resnet_50.output
x = layers.GlobalAveragePooling2D()(x)
x = layers.Dense(512, activation='relu')(x) 
x = layers.Dropout(0.5)(x)
x = layers.Dense(256, activation='relu')(x) 
x = layers.Dropout(0.5)(x)
x = layers.Dense(128, activation='relu')(x) 
x = layers.Dropout(0.5)(x)
x = layers.Dense(64, activation='relu')(x) 
x = layers.Dropout(0.5)(x)
predictions = layers.Dense(5, activation='softmax')(x)
model = Model(inputs = resnet_50.input, outputs = predictions)

---

4) Training

- optimizer='adam', epochs=10

# define training function
def trainModel(model, epochs, optimizer):
    batch_size = 32
    model.compile(optimizer=optimizer, loss="sparse_categorical_crossentropy", metrics=["accuracy"])
    return model.fit(train_generator, validation_data=valid_generator, epochs=epochs, batch_size=batch_size)

# launch the training
model_history = trainModel(model = model, epochs = 10, optimizer = "Adam")

---

# Loss curve

loss_train_curve = model_history.history["loss"]
loss_val_curve = model_history.history["val_loss"]
plt.plot(loss_train_curve, label = "Train")
plt.plot(loss_val_curve, label = "Validation")
plt.legend(loc = 'upper right')
plt.title("Loss")
plt.show()

 

---

# Accuracy curve

acc_train_curve = model_history.history["accuracy"]
acc_val_curve = model_history.history["val_accuracy"]
plt.plot(acc_train_curve, label = "Train")
plt.plot(acc_val_curve, label = "Validation")
plt.legend(loc = 'lower right')
plt.title("Accuracy")
plt.show()

---

5) Evaluation the performace

test_loss, test_acc = model.evaluate(test_generator)
print("The test loss is: ", test_loss)
print("The best accuracy is: ", test_acc*100)

---

6) Predict New data

img = tf.keras.preprocessing.image.load_img('C:/Users/USER/Desktop/resnet50/data_full/roses/123128873_546b8b7355_n.jpg', 
                                            target_size=(224, 224))
img_array = tf.keras.preprocessing.image.img_to_array(img)
img_array = np.array([img_array]) 
img

---

# generate predictions for samples
predictions = model.predict(img_array)
print(predictions)

# generate argmax for predictions
class_id = np.argmax(predictions, axis = 1)
print(class_id)

# transform classes number into classes name
print(class_names[class_id.item()])

---

7) Save model

# directory
model.save('C:Users/USER/Desktop/resnet50/saved_model/my_model')

# or h5 file
model.save('C:/Users/USER/Desktop/resnet50/saved_model/my_model.h5')

model = tf.keras.models.load_model('C:/Users/USER/Desktop/resnet50/saved_model/my_model')
model.summary()

- 전체 24809605 parameters 중에서 1221893 개 > 약 5%만 이용하여 모델을 설계함.

- Tresnfer learning > 굉장히 효율적인 접근방법

 

 

ref: https://github.com/ovh/ai-training-examples/blob/main/notebooks/computer-vision/image-classification/tensorflow/resnet50/notebook-resnet-transfer-learning-image-classification.ipynb